01
Remove a person from a shot and keep your soundtrack: H3 Person Remover
October 11, 2026 briefing recipe read the briefing →
(model card, workflow examples/Person-Remover-Window-Reroll-V1.json, H3 Relay)
model card ↗Workflow Index
Recipes, not announcements
Every workflow published in a briefing, kept in one place: what to install, why it is worth the afternoon, and exactly where it falls over. Nothing here ships without the second half.
463 workflows indexed 44 briefings mined Updated with every edition
02
The steps
October 11, 2026 briefing recipe read the briefing →
Update ComfyUI until it has native MiniMax H3 and SAM 3.1 support. Put H3-Person-Remover-V1.safetensors in models/loras/. Conform your clip to 24 fps with width and height divisible by 32, ideally about five seconds and one continuous shot. Paint the person out of frame one in…
03
How it works
October 11, 2026 briefing recipe read the briefing →
The model never sees the person, only a green hole and your clean first frame. It fills the hole window by window, and each window inherits the last 18 frames of picture and sound from the one before, so the fill stays continuous.
04
Why it is good
October 11, 2026 briefing recipe read the briefing →
Rerolling one window keeps every earlier window you approved, so a bad patch costs one window of render time, not the whole clip.
Why it is good
Rerolling one window keeps every earlier window you approved, so a bad patch costs one window of render time, not the whole clip.
05
Where it breaks
October 11, 2026 briefing recipe read the briefing →
The whole frame is regenerated, so check unmasked areas against the original. Shadows, reflections and stray hands the mask misses stay. A sloppy first-frame cleanup spreads through every later window, and hard cuts or big camera moves break continuity. Licensing excludes the…
Where it breaks
The whole frame is regenerated, so check unmasked areas against the original. Shadows, reflections and stray hands the mask misses stay. A sloppy first-frame cleanup spreads through every later window, and hard cuts or big camera moves break continuity.…
06
Board a sequence from one frame: Next-Scene chains
October 11, 2026 briefing recipe read the briefing →
(model card, recipe per ComfyUI Wiki)
model card ↗
07
The steps
October 11, 2026 briefing recipe read the briefing →
Load nextscenestep2500.safetensors on Qwen-Image 2.1's edit workflow with your frame as the input. For a single next shot use strength 0.7 to 0.9. For a chain, drop strength to 0.7 or lower and add a 0.5-pixel blur to each frame before feeding it back. Start every prompt with…
08
How it works
October 11, 2026 briefing recipe read the briefing →
It learned what tends to happen between consecutive frames of two films, so it treats your direction as a cut, not a touch-up.
09
Why it is good
October 11, 2026 briefing recipe read the briefing →
A director or storyboard artist gets five consistent angles on one location in the time it takes to write five sentences.
Why it is good
A director or storyboard artist gets five consistent angles on one location in the time it takes to write five sentences.
10
Where it breaks
October 11, 2026 briefing recipe read the briefing →
At strength 0.9 grain builds up visibly by the third or fourth hop. It trained at 512 px, it avoids portraits, and the base model's research license keeps it out of client deliverables.
Where it breaks
At strength 0.9 grain builds up visibly by the third or fourth hop. It trained at 512 px, it avoids portraits, and the base model's research license keeps it out of client deliverables.
11
Carry-forward: check before you build on it, with SynthID Detector
12
The steps
October 11, 2026 briefing recipe read the briefing →
Upload supplied assets at synthid.com before compositing them into client work, and log the result in the project file.
13
Why it is good
October 11, 2026 briefing recipe read the briefing →
It turns a reputational risk into a two-minute check, and PetaPixel's October 9 report that Nikon disqualified a contest winner shows the stakes.
Why it is good
It turns a reputational risk into a two-minute check, and PetaPixel's October 9 report that Nikon disqualified a contest winner shows the stakes.
14
Where it breaks
October 11, 2026 briefing recipe read the briefing →
It only detects marks from Google and named partners; a "no" proves little about open models.
Where it breaks
It only detects marks from Google and named partners; a "no" proves little about open models.
15
Remove a Person From Video With AI: Why the Frame You Paint by Hand Decides the Whole Shot
October 11, 2026 deep dive read the piece →
Akatz Labs' free H3 Person Remover workflow removes a person from a short clip by never letting the video model see them, so the result is decided by two human jobs, painting a clean first frame and checking the mask, while short rerollable chunks keep a bad patch cheap, and a…
16
Swap the performer, keep the scene and the song: Genjutsu
17
The steps
October 10, 2026 briefing recipe read the briefing →
Run setup.sh, add ENHANCORAPIKEY and REPLICATEAPITOKEN to .env, launch with start.command and open http://127.0.0.1:8770. Drop in a clip and character reference images. The app separates vocals from music with Demucs, pitches the vocals up three semitones, makes a colored depth…
18
How it works
October 10, 2026 briefing recipe read the briefing →
The video model never sees your performer's real face, only a depth silhouette in the real room, so it keeps the motion and camera while inventing the new character from your references.
19
Why it is good
October 10, 2026 briefing recipe read the briefing →
Two approval gates (mask, then draft) stop you paying for HD on a bad mask, and the final audio is your original track, not a generated one.
Why it is good
Two approval gates (mask, then draft) stop you paying for HD on a bad mask, and the final audio is your original track, not a generated one.
20
Where it breaks
October 10, 2026 briefing recipe read the briefing →
It runs on paid hosted services and the repo states no prices; the API notes also admit the provider's full schema "was not independently retrievable." The README does not explain the three-semitone shift. Draft and HD are billed separately. Media passes through public file…
Where it breaks
It runs on paid hosted services and the repo states no prices; the API notes also admit the provider's full schema "was not independently retrievable." The README does not explain the three-semitone shift. Draft and HD are billed separately. Media passes…
21
Draft on a 16 GB Mac with Qwen-Image-2.1-Turbo
22
The steps
October 10, 2026 briefing recipe read the briefing →
Install MLX-Serve 26.10.2 or newer, open the Image tab, pick the Qwen-Image 2.1 Turbo 4-bit pack (10.5 GB). Generate at 1024x1024. Steps are fixed at 8. For edits, attach up to nine reference images.
23
How it works
October 10, 2026 briefing recipe read the briefing →
The 4-bit pack stores the model's numbers at lower precision so it fits in less memory. On a big Mac, the author found "Quantizing saves memory, not time."
24
Why it is good
October 10, 2026 briefing recipe read the briefing →
Free, local, no Python, and fast enough to iterate on composition.
Why it is good
Free, local, no Python, and fast enough to iterate on composition.
25
Where it breaks
October 10, 2026 briefing recipe read the briefing →
The 4-bit pack garbled a "€" in the author's test; use 8-bit for lettering. The timings come from a 256 GB M5 Ultra, not a 16 GB machine. And the research-only license means these drafts are for exploration, not delivery.
Where it breaks
The 4-bit pack garbled a "€" in the author's test; use 8-bit for lettering. The timings come from a 256 GB M5 Ultra, not a 16 GB machine. And the research-only license means these drafts are for exploration, not delivery.
26
Carry-forward: check before you build on it, with SynthID Detector
27
The steps
October 10, 2026 briefing recipe read the briefing →
Before compositing a supplied asset or stock frame into client work, upload it at synthid.com. Note the result in your project file. Run your own deliverable through it too, so you know what a client or contest will see.
28
Why it is good
October 10, 2026 briefing recipe read the briefing →
It turns a reputational risk into a two-minute check.
Why it is good
It turns a reputational risk into a two-minute check.
29
Where it breaks
October 10, 2026 briefing recipe read the briefing →
It only detects SynthID from Google and named partners, and secondary reports put the limit at about ten checks a day.
Where it breaks
It only detects SynthID from Google and named partners, and secondary reports put the limit at about ten checks a day.
30
AI Video Actor Replacement That Keeps Your Shot and Your Soundtrack, and Makes You Approve Twice Before You Pay
October 10, 2026 deep dive read the piece →
The useful part of TEIN's free Genjutsu workflow is less the character swap than its order of operations, which hides the performer from the video model, makes you approve a mask and then a draft before any full-quality render, and puts the original soundtrack back at the end, a…
31
Turn a written explainer into a reframed social cut: Claude Motion to Luma
32
The steps
October 9, 2026 briefing recipe read the briefing →
Build the animated explainer in Claude Motion. Add the Luma connector in Claude. Open the animation in Luma, where it arrives as a video. Ask Luma's agent to restyle it while keeping the motion, then ask for 9:16 and 1:1 versions for social.
33
How it works
October 9, 2026 briefing recipe read the briefing →
Claude writes the animation as code, so timing and layout are exact and editable. Luma treats the render as source footage and runs its video models over it.
34
Why it is good
October 9, 2026 briefing recipe read the briefing →
You get precise motion design (charts that hit the right numbers, type that lands on the beat) and a generated finish, without animating twice.
Why it is good
You get precise motion design (charts that hit the right numbers, type that lands on the beat) and a generated finish, without animating twice.
35
Where it breaks
October 9, 2026 briefing recipe read the briefing →
The beta is Team and Enterprise only. Once the piece is in Luma, it is a video: change a number in the chart and you go back to Claude and re-export. Pricing on the Luma side is not published in the post.
Where it breaks
The beta is Team and Enterprise only. Once the piece is in Luma, it is a video: change a number in the chart and you go back to Claude and re-export. Pricing on the Luma side is not published in the post.
36
Picture-plus-sound video on a 24 GB laptop with Prism-Comfy
37
The steps
October 9, 2026 briefing recipe read the briefing →
Install the Comfyui-Prism-T8 nodes. Download the seven model files (about 38 GiB) into models/diffusionmodels, models/textencoders and models/vae as the card lists. Load 01nativei2va.json: one image in, 848x480, 49 frames at 24 fps (about two seconds), separate prompts for…
38
How it works
October 9, 2026 briefing recipe read the briefing →
Offloading swaps pieces of the model between graphics memory and system memory, trading speed for fit.
39
Why it is good
October 9, 2026 briefing recipe read the briefing →
It is a documented, MIT-licensed path to Prism's synced sound on a laptop.
Why it is good
It is a documented, MIT-licensed path to Prism's synced sound on a laptop.
40
Where it breaks
October 9, 2026 briefing recipe read the briefing →
About 47 minutes for one two-second clip in the author's run. The author reports slight composition drift and soft texture and has not listened to the sample's audio. 720p and long clips are untested.
Where it breaks
About 47 minutes for one two-second clip in the author's run. The author reports slight composition drift and soft texture and has not listened to the sample's audio. 720p and long clips are untested.
41
Depth maps and 4x restores from Iris-3B
42
The steps
October 9, 2026 briefing recipe read the briefing →
Clone the repo and pip install -e .. Download the depth/ or upscaler/ folder. Run python scripts/depth.py photo.jpg --out depthout for a depth map, or python scripts/upscale.py photo.jpg --out upscaled for a 4x restore.
43
How it works
October 9, 2026 briefing recipe read the briefing →
The same backbone that makes images was fine-tuned for each job.
44
Why it is good
October 9, 2026 briefing recipe read the briefing →
A depth pass for fog, relighting or parallax, plus a restore, from Apache-licensed weights.
Why it is good
A depth pass for fog, relighting or parallax, plus a restore, from Apache-licensed weights.
45
Where it breaks
October 9, 2026 briefing recipe read the briefing →
Depth is relative, not measured distance. The authors found the pixel approach no better than a standard model for these jobs, and each folder is another 12 GB.
Where it breaks
Depth is relative, not measured distance. The authors found the pixel approach no better than a standard model for these jobs, and each folder is another 12 GB.
46
AI Motion Graphics Work Better as Code. Write the Animation, Then Hand It to a Video Model.
October 9, 2026 deep dive read the piece →
The reliable way to make motion graphics with AI is to have a language model write the animation as editable code, keep that code as the master file, and only then hand the render to a video model for a generated finish, because a cheap model can now write a working…
47
The steps
October 8, 2026 briefing recipe read the briefing →
Load Qwen-Image-2.1 in edit mode (the card's Diffusers example uses QwenImage21Pipeline with 40 steps and truecfgscale=1.0), add the v2 step-1,500 file at strength 1.0 as the card suggests, and feed your photo. Run one prompt per view, always starting with : front view,…
48
How it works
October 8, 2026 briefing recipe read the briefing →
The add-on learned from 13,328 pairs of rendered 3D objects and rigged characters photographed from known positions, so the angle label works like a camera instruction rather than a style word.
49
Why it is good
October 8, 2026 briefing recipe read the briefing →
A front, side and back set from one reference is what a character designer, a product shooter or a 3D modeler needs before anything else, and it used to mean a reshoot or a sketch session.
Why it is good
A front, side and back set from one reference is what a character designer, a product shooter or a 3D modeler needs before anything else, and it used to mean a reshoot or a sketch session.
50
Where it breaks
October 8, 2026 briefing recipe read the briefing →
If the result starts copying your original view, the card says lower the strength toward 0.8 rather than switching files. Fine detail softens on real-world subjects. The base model's research license rules out client work for now.
Where it breaks
If the result starts copying your original view, the card says lower the strength toward 0.8 rather than switching files. Fine detail softens on real-world subjects. The base model's research license rules out client work for now.
51
The steps
October 8, 2026 briefing recipe read the briefing →
Accept the MiniMax H3 license on Hugging Face and download the base model, skipping the FL2VA, Ref2VA and transformerref folders as the README shows. Put the ComfyUI-format DMAD file in your add-on folder. Open minimaxh3/workflows/dmadh34step15spodcast.json for speed or the…
52
How it works
October 8, 2026 briefing recipe read the briefing →
DMAD trains a student copy of MiniMax H3 to land in four steps what the original needs fifty for, so each take costs a fraction of the passes.
53
Why it is good
October 8, 2026 briefing recipe read the briefing →
It moves a 33-billion-parameter picture-plus-sound model from rented hardware to a card many editors already own.
Why it is good
It moves a 33-billion-parameter picture-plus-sound model from rented hardware to a card many editors already own.
54
Where it breaks
October 8, 2026 briefing recipe read the briefing →
The 24 GB timings are estimates, not measurements. The model's default output is about five seconds (124 frames), so watch 15-second takes for drift late in the shot. The 170 GB base download and the territory clause in the MiniMax license do not go away.
Where it breaks
The 24 GB timings are estimates, not measurements. The model's default output is about five seconds (124 frames), so watch 15-second takes for drift late in the shot. The 170 GB base download and the territory clause in the MiniMax license do not go away.
55
Build a consistent cast scene with [Nano Banana 2.1 in ComfyUI](https://docs.comfy.org/changelog) (v0.39.1 partner node)
October 8, 2026 briefing recipe read the briefing →
Nano Banana 2.1 in ComfyUI ↗
56
The steps
October 8, 2026 briefing recipe read the briefing →
Update ComfyUI to v0.39.1. Load up to four character reference images and your prop references into the Nano Banana 2.1 node. Write the scene in plain sentences naming each character the way you labeled the references. Set thinking to high for posters and diagrams, minimal for…
57
How it works
October 8, 2026 briefing recipe read the briefing →
The model reads every reference alongside your prompt and plans the layout before drawing, which is what the thinking level controls.
58
Why it is good
October 8, 2026 briefing recipe read the briefing →
Storyboards and comic panels with a recurring cast, at half the old per-image price.
Why it is good
Storyboards and comic panels with a recurring cast, at half the old per-image price.
59
Where it breaks
October 8, 2026 briefing recipe read the briefing →
Google itself says character consistency is not always perfect and small text is often blurry at 1K, so judge layout, not lettering, in drafts. Every run is a paid call, and the node goes away if you launch ComfyUI with --disable-partner-nodes.
Where it breaks
Google itself says character consistency is not always perfect and small text is often blurry at 1K, so judge layout, not lettering, in drafts. Every run is a paid call, and the node goes away if you launch ComfyUI with --disable-partner-nodes.
60
One Photo, 72 Camera Positions: Building an AI Turnaround Sheet With the Qwen-Image-2.1 Multiple Angles Add-On
October 8, 2026 deep dive read the piece →
The Multiple Angles add-on works because it treats camera position as a short fixed vocabulary learned from 3D renders, which makes it a reliable way to rough out front, side and back views of an object from one photo, but its research-licensed base model and untested character…
61
Render a talking clip with sound on your own card, with [Kandinsky 6.0](https://github.com/kandinskylab/kandinsky-6) (`just download pro-distill`)
October 7, 2026 briefing recipe read the briefing →
Kandinsky 6.0 ↗
62
The steps
October 7, 2026 briefing recipe read the briefing →
Clone the repo, run just setup (it picks the right PyTorch build for your card), then just download pro-distill and just generate "your prompt". Write the spoken line into the prompt. Outputs land in a timestamped folder alongside the expanded prompt and the settings used. For…
63
How it works
October 7, 2026 briefing recipe read the briefing →
One stream of the model draws the frames and a second stream writes the audio, and the two keep checking each other while they generate, which is why the mouth and the sound land together instead of being dubbed afterwards.
64
Why it is good
October 7, 2026 briefing recipe read the briefing →
Dialogue, room tone and effects in one pass, under a license that lets you deliver it to a paying client.
Why it is good
Dialogue, room tone and effects in one pass, under a license that lets you deliver it to a paying client.
65
Where it breaks
October 7, 2026 briefing recipe read the briefing →
Five seconds is the ceiling. At full quality a Lite clip takes over twenty minutes on a mid-range card; the 10-pass version is faster but unmeasured, so test it on your own card before planning a schedule, and draft ideas in the free Space. Lip-sync quality is the authors'…
Where it breaks
Five seconds is the ceiling. At full quality a Lite clip takes over twenty minutes on a mid-range card; the 10-pass version is faster but unmeasured, so test it on your own card before planning a schedule, and draft ideas in the free Space. Lip-sync quality…
66
The steps
October 7, 2026 briefing recipe read the briefing →
Put the file in models/loras. Load the ltx-2.5-22b-distilled model and the LTX-2.5V2VICLoRASingleStageDistilled.json workflow from ComfyUI-LTXVideo. Connect your clip as the reference video at strength 1.0, set the add-on to 1.0, leave the prompt empty, and run only the first…
ComfyUI-LTXVideo ↗
67
How it works
October 7, 2026 briefing recipe read the briefing →
The add-on teaches the video model to answer "here is a shot, draw what belongs in front" as a black-and-white video, so the matte comes out as a second clip lined up frame for frame with yours.
68
Why it is good
October 7, 2026 briefing recipe read the briefing →
It aims at the material keyers hate: flyaway hair, smoke, glass, water.
Why it is good
It aims at the material keyers hate: flyaway hair, smoke, glass, water.
69
Where it breaks
October 7, 2026 briefing recipe read the briefing →
The model picks the foreground itself, so a shot with two subjects may give you the wrong one. Long clips leak picture into the matte. Full HD needs data-center hardware, so most people will rent a card or work at lower resolution.
Where it breaks
The model picks the foreground itself, so a shot with two subjects may give you the wrong one. Long clips leak picture into the matte. Full HD needs data-center hardware, so most people will rent a card or work at lower resolution.
70
Clone a narrator's voice on a laptop with [KittenTTS 2](https://huggingface.co/KittenML/kitten-tts-2)
October 7, 2026 briefing recipe read the briefing →
KittenTTS 2 ↗
71
The steps
October 7, 2026 briefing recipe read the briefing →
pip install kittenml. Record 5 to 30 seconds of clean speech from one person who has agreed to it. Call m.generate("Your line here.", reference="myvoice.wav") as the card shows. Try the built-in voices first to judge quality before cloning.
72
How it works
October 7, 2026 briefing recipe read the briefing →
The model listens to your sample as part of the request and continues in that voice, so there is no training step and no upload.
73
Why it is good
October 7, 2026 briefing recipe read the briefing →
Scratch narration for an animatic, offline, on a machine without a graphics card.
Why it is good
Scratch narration for an animatic, offline, on a machine without a graphics card.
74
Where it breaks
October 7, 2026 briefing recipe read the briefing →
The license caps commercial use at $1 million in revenue or funding. No published real-time figures on CPU (the card says its llama.cpp-based C++ build is the fastest route), and no hosted demo to hear it before installing.
Where it breaks
The license caps commercial use at $1 million in revenue or funding. No published real-time figures on CPU (the card says its llama.cpp-based C++ build is the fastest route), and no hosted demo to hear it before installing.
75
AI Alpha Mattes Without a Green Screen: Pulling Hair, Smoke and Glass From Ordinary Footage With LTX-2.5 Alpha-Gen
October 7, 2026 deep dive read the piece →
LTX-2.5 Alpha-Gen is the useful kind of AI tool because it hands back a matte you still own and grade rather than a finished frame, but it chooses the subject for you, so it belongs on single-subject shots kept under six seconds.
76
The steps
October 6, 2026 briefing recipe read the briefing →
Lock the camera. Shoot the frame with the subject, then the same frame empty. Install the repo's inference/requirements.txt, then run python inference/qwenextract.py --composite picture.png --background cleanbackground.png --output extracted.png. The script fetches the add-on if…
77
How it works
October 6, 2026 briefing recipe read the briefing →
The model sees both images and learned, over 15,000 training steps, which pixels belong to the scene and which to the thing in front of it. Because it paints the alpha channel directly, it can keep half-transparent edges instead of cutting a hard line.
78
Why it is good
October 6, 2026 briefing recipe read the briefing →
It is the old difference key with judgment added: it can keep hair and soft shadow that a pixel-subtraction key throws away, and it works on drawn frames as well as photos.
Why it is good
It is the old difference key with judgment added: it can keep hair and soft shadow that a pixel-subtraction key throws away, and it works on drawn frames as well as photos.
79
Where it breaks
October 6, 2026 briefing recipe read the briefing →
Research license, so no paid work. Stills only; running it frame by frame on video will likely flicker (our expectation, not the author's claim). Needs a 16 GB NVIDIA card and 64 GB of RAM. And it depends on a true clean plate: if the light or framing shifts between shots,…
Where it breaks
Research license, so no paid work. Stills only; running it frame by frame on video will likely flicker (our expectation, not the author's claim). Needs a 16 GB NVIDIA card and 64 GB of RAM. And it depends on a true clean plate: if the light or framing shifts…
80
Find the shot you know you filmed, with [SCM](https://github.com/allenv0/SCM)
October 6, 2026 briefing recipe read the briefing →
SCM ↗
81
The steps
October 6, 2026 briefing recipe read the briefing →
brew tap allenv0/scm, brew trust allenv0/scm, then brew install --cask allenv0/scm/scm on an Apple-silicon Mac running macOS 12 or later. Add your footage folder as a watched folder and let it index. Use Scenes mode for a described moment ("wide shot, rain on the window, someone…
82
How it works
October 6, 2026 briefing recipe read the briefing →
It cuts each video into scenes and turns each one into a searchable fingerprint of what it looks like, using a local vision model you pick from a short list (the default is about 435 MB, downloaded once). Whisper transcribes speech and Tesseract reads visible text, so three…
83
Why it is good
October 6, 2026 briefing recipe read the briefing →
Nothing leaves the machine, which matters for client footage under NDA, and renamed duplicates are caught by content rather than filename.
Why it is good
Nothing leaves the machine, which matters for client footage under NDA, and renamed duplicates are caught by content rather than filename.
84
Where it breaks
October 6, 2026 briefing recipe read the briefing →
Mac only. Indexing a large drive will take time, and the README gives no speed figures. Search by description is only as good as the vision model's sense of your shots, and the optional chat mode needs a separate local language model.
Where it breaks
Mac only. Indexing a large drive will take time, and the README gives no speed figures. Search by description is only as good as the vision model's sense of your shots, and the optional chat mode needs a separate local language model.
85
The steps
October 6, 2026 briefing recipe read the briefing →
Install FastVideo with uv and the CUDA 13 build of PyTorch, as the V2 card shows. The V2 card's eight-pass example is python examples/inference/basic/basicfasth38step.py --prompt "your prompt" --no-warmup --repeats 1. For the consumer build, the card's only code downloads the…
FastVideo ↗
86
How it works
October 6, 2026 briefing recipe read the briefing →
The team trained a fast student to match H3's results in eight passes, then shrank the model and its prompt reader to four-bit numbers so they fit on one card.
87
Why it is good
October 6, 2026 briefing recipe read the briefing →
Synced dialogue and sound effects on a desk machine, with no per-second fees.
Why it is good
Synced dialogue and sound effects on a desk machine, with no per-second fees.
88
Where it breaks
October 6, 2026 briefing recipe read the briefing →
Blackwell cards only. The precomputed table works for the included eight-pass schedule only and errors on anything else. No published speed on a 5090, weaker fine detail and hard motion than the full model, and the H3 license.
Where it breaks
Blackwell cards only. The precomputed table works for the included eight-pass schedule only and errors on anything else. No published speed on a 5090, weaker fine detail and hard motion than the full model, and the H3 license.
89
Clean-Plate Cutouts With AI: How PlateExtract Turns Two Photos Into a Transparent PNG With Soft Edges
October 6, 2026 deep dive read the piece →
PlateExtract gives the oldest trick in compositing, the clean-plate difference key, the judgment it always lacked on hair, glass and shadow, but the result still depends on shoot discipline, and its research license makes it a tool for tests and personal work rather than paid…
90
Batch-finish a folder of PSDs from the command line with [PhotoCraft](https://github.com/storytold/photocraft) (`photocraft-cli`)
October 5, 2026 briefing recipe read the briefing →
PhotoCraft ↗
91
The steps
October 5, 2026 briefing recipe read the briefing →
Install a v0.1.x build from the releases page. Run the README's own example on one file first: photocraft-cli run wave.psd --cmd filter.sharpen.smartSharpen --params '{"amount":80}' --cmd layer.newAdjustmentLayer.curves --params '{"points":[[0,0],[64,48],[192,212],[255,255]]}'…
releases page ↗
92
How it works
October 5, 2026 briefing recipe read the briefing →
The menus, the command line and the MCP server all call the same 500+ commands, so a recipe you test by hand is the recipe the script or agent runs.
93
Why it is good
October 5, 2026 briefing recipe read the briefing →
A curves layer added this way is still an adjustment layer, so the client file stays editable instead of baked flat.
Why it is good
A curves layer added this way is still an adjustment layer, so the client file stays editable instead of baked flat.
94
Where it breaks
October 5, 2026 briefing recipe read the briefing →
Early alpha. One of 135 test PSDs failed to round-trip, so open a few outputs in Photoshop before trusting a batch. No generative fill. And "menu item live" is not the same as "matches Photoshop's math," so check your sharpening by eye.
Where it breaks
Early alpha. One of 135 test PSDs failed to round-trip, so open a few outputs in Photoshop before trusting a batch. No generative fill. And "menu item live" is not the same as "matches Photoshop's math," so check your sharpening by eye.
95
The steps
October 5, 2026 briefing recipe read the briefing →
Download the repo, run npm install, then node runtime/cli.js --package . --prompt "your idea" -o out.png to see the default. For anything with type, write a structured caption and pass it with --caption-file caption.json, which skips the caption writer (about 90 seconds saved).…
96
How it works
October 5, 2026 briefing recipe read the briefing →
Ideogram 4 reads a structured caption, not a sentence. The bundled writer turns your idea into one, and that is where the invented signs come from. Writing the caption yourself removes the middleman.
97
Why it is good
October 5, 2026 briefing recipe read the briefing →
Poster layouts with exact text, on a card you own, with no per-image fee.
Why it is good
Poster layouts with exact text, on a card you own, with no per-image fee.
98
Where it breaks
October 5, 2026 briefing recipe read the briefing →
Non-commercial license, the grey "memorized mode" panel on short prompts (the Node runtime re-renders automatically), Firefox fails at 1024 × 1024, and the author claims no accuracy figure against the full model.
Where it breaks
Non-commercial license, the grey "memorized mode" panel on short prompts (the Node runtime re-renders automatically), Firefox fails at 1024 × 1024, and the author claims no accuracy figure against the full model.
99
Dress a model in a full look with [Garments2Look](https://github.com/ArtmeScienceLab/Garments2Look)
October 5, 2026 briefing recipe read the briefing →
Garments2Look ↗
100
The steps
October 5, 2026 briefing recipe read the briefing →
Make a collage of every piece (top, layer, bottoms, shoes, bag, belt). For repainting, grey out the clothing on the person photo in flat 128 grey. Number each item in the prompt with styling notes ("partially unbuttoned, tucked-in"), then end with the layering order. Run the…
101
How it works
October 5, 2026 briefing recipe read the briefing →
The add-ons learned from 20,000 outfit pairs where the numbered list matched the garments, so the numbers tell the model which collage piece goes where and which sits on top.
102
Why it is good
October 5, 2026 briefing recipe read the briefing →
A lookbook comp in one pass instead of six separate swaps.
Why it is good
A lookbook comp in one pass instead of six separate swaps.
103
Where it breaks
October 5, 2026 briefing recipe read the briefing →
No stated add-on license, fine accessory detail drifts, no ComfyUI node, and the mask is on you.
Where it breaks
No stated add-on license, fine accessory detail drifts, no ComfyUI node, and the mask is on you.
104
Garments2Look: Dress a Whole Outfit in One AI Edit by Writing the Layering Order
October 5, 2026 deep dive read the piece →
AI try-on fails on full outfits because nothing tells the model what sits on top of what, and Garments2Look fixes that with a numbered item list and a written layering order, which makes a whole-look comp a single edit, at the cost of rented hardware, a hand-made mask and an…
105
The steps
October 4, 2026 briefing recipe read the briefing →
Pick a native canvas in multiples of 8, say 128 × 128. Generate at four times that, 512 × 512, with the card's prompt pattern: "[Shot 1] pxgrid4 nativeres128x128 sourcefps10. Native pixel canvas: 128 by 128 pixels. Each native pixel is a 4 by 4 solid-color square… 2D pixel-art…
106
How it works
October 4, 2026 briefing recipe read the briefing →
The model learned to paint in 4 × 4 blocks, so the cleanup step only has to tidy cells that are already close, instead of guessing a grid from soft video.
107
Why it is good
October 4, 2026 briefing recipe read the briefing →
One palette per clip stops the flicker you get when every frame picks its own colors, which is what usually gives AI "pixel art" away.
Why it is good
One palette per clip stops the flicker you get when every frame picks its own colors, which is what usually gives AI "pixel art" away.
108
Where it breaks
October 4, 2026 briefing recipe read the briefing →
It is an early checkpoint, the card warns that "static-region stability and exact pixel boundaries may be inconsistent," local use needs DiffSynth-Studio and the full H3 download, and long clips are untested.
Where it breaks
It is an early checkpoint, the card warns that "static-region stability and exact pixel boundaries may be inconsistent," local use needs DiffSynth-Studio and the full H3 download, and long clips are untested.
109
The steps
October 4, 2026 briefing recipe read the briefing →
Install the AusBoss node pack (2.5.1 or newer) on ComfyUI 0.38 or newer and load the workflow. Load the add-on at strength 1.0, with no trigger word. Feed the person as image 1 and the outfit as image 2. Prompt: "Dress the person in image 1 in the [outfit description] shown in…
110
How it works
October 4, 2026 briefing recipe read the briefing →
The add-on learned from 66 hand-picked before-and-after pairs, each one chosen because the person and background stayed put, so it learned that a swap should touch only the clothes.
111
Why it is good
October 4, 2026 briefing recipe read the briefing →
A product shot that stays registered to the original can be layered over it in Photoshop and masked, instead of redone.
Why it is good
A product shot that stays registered to the original can be layered over it in Photoshop and masked, instead of redone.
112
Where it breaks
October 4, 2026 briefing recipe read the briefing →
Shoes, accessories that cross the garment, poses that suit the old clothes, untested sizes (the author sampled at about 2 megapixels), and the research license.
Where it breaks
Shoes, accessories that cross the garment, poses that suit the old clothes, untested sizes (the author sampled at about 2 megapixels), and the research license.
113
Read a line with a chosen emotion, using a [Humaneness Voice add-on](https://huggingface.co/laion/Humaneness-Voice-Small-Rank1-LoRAs)
October 4, 2026 briefing recipe read the briefing →
Humaneness Voice add-on ↗
114
The steps
October 4, 2026 briefing recipe read the briefing →
Download the base model and the add-on repo side by side. Run the card's inferrank1.py with --adapter-id emotion/Affection, a prompt of CAPTION: warmly affectionate... plus the quoted TRANSCRIPT:, --language en and a fixed --seed. Change only the add-on and compare takes.
115
How it works
October 4, 2026 briefing recipe read the briefing →
The base model already knows how to act; each add-on nudges it toward one voice, one mood or one trait.
116
Why it is good
October 4, 2026 briefing recipe read the briefing →
A fixed seed and a swapped add-on give you comparable takes of the same line, like asking an actor for a warmer read.
Why it is good
A fixed seed and a swapped add-on give you comparable takes of the same line, like asking an actor for a warmer read.
117
Where it breaks
October 4, 2026 briefing recipe read the briefing →
One add-on at a time, research-grade timing control, English and German only, and no consent check built into the voice files.
Where it breaks
One add-on at a time, research-grade timing control, English and German only, and no consent check built into the voice files.
118
Real Pixel Art From an AI Video Model: Draw Big, Snap to the Grid, Lock the Palette
October 4, 2026 deep dive read the piece →
AI video models fake pixel art because nothing forces them onto a grid or a fixed set of colors, and two new MiniMax H3 add-ons plus a three-command cleanup pass (shrink, one palette for the whole clip, enlarge without smoothing) turn their output into sprite animation that…
119
The steps
October 3, 2026 briefing recipe read the briefing →
Install with pip install "audiosliders[model,demo] @ git+https://github.com/takakhoo/audio-diffusion-control" audioboxaesthetics, then start the local slider app with python -m audiosliders.server --sliders hf:ace-step-1.5-xl-turbo/text --backbone ace-turbo. Write one…
120
How it works
October 3, 2026 briefing recipe read the briefing →
Each slider is a small add-on that nudges the model along one direction it learned, so the same seed produces the same piece moved along one axis.
121
Why it is good
October 3, 2026 briefing recipe read the briefing →
An editor gets three related cues that cut together, instead of three unrelated rerolls that share a genre and nothing else.
Why it is good
An editor gets three related cues that cut together, instead of three unrelated rerolls that share a genre and nothing else.
122
Where it breaks
October 3, 2026 briefing recipe read the briefing →
Short instrumentals only, sliders leak into each other, and quality drops past each slider's usable range (-1 to +1 for some, up to +2 for others). No vocals tested.
Where it breaks
Short instrumentals only, sliders leak into each other, and quality drops past each slider's usable range (-1 to +1 for some, up to +2 for others). No vocals tested.
123
The steps
October 3, 2026 briefing recipe read the briefing →
Clone both repos at --branch v0.6.0, run docker compose build and docker compose up -d inside ldraw-nova-docker, and open https://localhost:8443. Describe the model, review the plan, steer as it renders and corrects. Export the .glb into Blender for lighting, or keep the LDraw…
124
How it works
October 3, 2026 briefing recipe read the briefing →
The agent never draws pixels. It writes a script that places real parts by coordinates, renders the result, looks at it and fixes mistakes.
125
Why it is good
October 3, 2026 briefing recipe read the briefing →
The output is an editable parts list, not a picture of bricks, so you can change a color or swap a piece by hand afterward.
Why it is good
The output is an editable parts list, not a picture of bricks, so you can change a color or swap a piece by hand afterward.
126
Where it breaks
October 3, 2026 briefing recipe read the briefing →
The author says only expensive top models produce large, correct builds, setup needs Docker and a self-signed certificate for VR, and the README does not say whether builds hold together physically.
Where it breaks
The author says only expensive top models produce large, correct builds, setup needs Docker and a self-signed certificate for VR, and the README does not say whether builds hold together physically.
127
Draft a five-second talking clip in four steps with [DMAD](https://huggingface.co/ZhengmingYu/DMAD) (`dmad_minimax_h3_4step_full_critic.safetensors`)
October 3, 2026 briefing recipe read the briefing →
DMAD ↗
128
The steps
October 3, 2026 briefing recipe read the briefing →
Download MiniMax H3 with hf download MiniMaxAI/MiniMax-H3 --local-dir models/MiniMax-H3 --exclude "FL2VA/", (the README excludes a few more folders), then run inference.py from the DMAD repo with the base folder, the add-on file and a --prompt such as its own "A polar bear is…
DMAD repo ↗
129
How it works
October 3, 2026 briefing recipe read the briefing →
A training method taught a small add-on to jump to the finished clip in four moves instead of fifty.
130
Why it is good
October 3, 2026 briefing recipe read the briefing →
Picture and sound arrive together, fast enough to try several takes of a line.
Why it is good
Picture and sound arrive together, fast enough to try several takes of a line.
131
Where it breaks
October 3, 2026 briefing recipe read the briefing →
No ComfyUI node, a roughly 170 GB base download, a single 80 GB GPU per the README, no published timing, and fast versions usually trade some detail for speed.
Where it breaks
No ComfyUI node, a roughly 170 GB base download, a single 80 GB GPU per the README, no published timing, and fast versions usually trade some detail for speed.
132
ldraw-nova: Design a LEGO Model by Describing It to an AI Agent, Then Open It in Blender
October 3, 2026 deep dive read the piece →
ldraw-nova shows that an AI agent which writes an editable parts file, rather than generating a picture of bricks, gives designers something they can revise, render and price, but its own instructions admit a pretty render does not prove a model can be built, so you check the…
133
The steps
October 2, 2026 briefing recipe read the briefing →
Start from your full-resolution master. Make one change per pass ("make the bottle cobalt blue"). If the change belongs to one area, paint a mask (on Replicate's listing, black is edited and white is kept). Feed the output of pass one in as the source for pass two. On fal, set…
134
How it works
October 2, 2026 briefing recipe read the briefing →
Ideogram regenerates the region you asked about and then copies untouched pixels back from your original, so errors cannot compound in areas you never edited.
135
Why it is good
October 2, 2026 briefing recipe read the briefing →
Ten small edits stop being ten chances to wreck the skin texture. And because output comes back at source size, you are not upscaling a shrunken file at the end.
Why it is good
Ten small edits stop being ten chances to wreck the skin texture. And because output comes back at source size, you are not upscaling a shrunken file at the end.
136
Where it breaks
October 2, 2026 briefing recipe read the briefing →
The guarantee covers untouched pixels, not the seam: a recolored object can still pick up a halo where new meets copied. With a mask on fal you get three reference images, not four. No independent drift test exists yet, so check with a difference blend in Photoshop before…
Where it breaks
The guarantee covers untouched pixels, not the seam: a recolored object can still pick up a halo where new meets copied. With a mask on fal you get three reference images, not four. No independent drift test exists yet, so check with a difference blend in…
137
Compose a layout with boxes, then fix one element, using [FLUX 3 Image's layout editing](https://docs.bfl.ai/flux_3/flux3_image_layout)
October 2, 2026 briefing recipe read the briefing →
FLUX 3 Image's layout editing ↗
138
The steps
October 2, 2026 briefing recipe read the briefing →
In the BFL playground, pick an aspect ratio, draw a box for each element that matters (product, headline, prop), describe what goes in each, add a scene prompt, and generate. To revise, draw a box only around the thing that changes and describe the recolor, replacement or move.
BFL playground ↗
139
How it works
October 2, 2026 briefing recipe read the briefing →
Boxes use a 0 to 1000 grid on both axes, so the same layout works at any aspect ratio. The model treats everything outside your box as an anchor to keep.
140
Why it is good
October 2, 2026 briefing recipe read the briefing →
A poster becomes a set of slots you can revise one at a time instead of one prompt you rewrite from scratch.
Why it is good
A poster becomes a set of slots you can revise one at a time instead of one prompt you rewrite from scratch.
141
Where it breaks
October 2, 2026 briefing recipe read the briefing →
BFL's docs say shadows, reflections and nearby lighting "may still change," and "Very small boxes can fail. In our tests, a new element in a box about 40 × 25 pixels often did not appear." Give new elements generous boxes.
Where it breaks
BFL's docs say shadows, reflections and nearby lighting "may still change," and "Very small boxes can fail. In our tests, a new element in a box about 40 × 25 pixels often did not appear." Give new elements generous boxes.
142
Describe a ComfyUI graph and let [Comfy Agent](https://blog.comfy.org/p/comfy-agent-the-first-agent-for-craft) build it
October 2, 2026 briefing recipe read the briefing →
Comfy Agent ↗
143
The steps
October 2, 2026 briefing recipe read the briefing →
Open Comfy Cloud, open the Agent panel, leave run permission on "Ask," and describe the job ("image to video from this still, five seconds, slow push-in"). The agent finds templates, adds and connects nodes, validates, then asks before each run. Open any workflow tab yourself…
docs ↗
144
How it works
October 2, 2026 briefing recipe read the briefing →
A frontier language model (the docs price it as Opus 4.8) reads your environment's models and nodes and edits the canvas while you watch.
145
Why it is good
October 2, 2026 briefing recipe read the briefing →
The node-graph learning curve is the biggest wall in local creative AI, and this turns it into a conversation you can watch and learn from.
Why it is good
The node-graph learning curve is the biggest wall in local creative AI, and this turns it into a conversation you can watch and learn from.
146
Where it breaks
October 2, 2026 briefing recipe read the briefing →
Comfy's blog calls it a beta, it runs in Comfy Cloud only for now, and it bills tokens from the same Comfy Credits pool as your renders. "Auto" mode runs generations without asking, which is how you burn a month of credits on an iteration loop.
Where it breaks
Comfy's blog calls it a beta, it runs in Comfy Cloud only for now, and it bills tokens from the same Comfy Credits pool as your renders. "Auto" mode runs generations without asking, which is how you burn a month of credits on an iteration loop.
147
Comfy Agent: How to Build a ComfyUI Workflow by Describing It (and Why to Leave It on Ask)
October 2, 2026 deep dive read the piece →
Comfy Agent removes the node-graph wall by building ComfyUI workflows from a plain description while you watch, but the useful skill is directing it like an assistant (say what to keep, approve each run, save your house rules as skills), because its tokens and your renders come…
148
Draft a fight scene, check the choreography, then commit
October 1, 2026 briefing recipe read the briefing →
Files: 【WORK-FISHER】26-9-28-MINIMAXH3-武戏高动态工作流.json and 打戏提示词模版MiniMax15秒一镜到底v1.4.md.
【WORK-FISHER】26-9-28-MINIMAXH3-武戏高动态工作流.json and… ↗
149
The steps
October 1, 2026 briefing recipe read the briefing →
Install the six node packs at the versions listed on the card (the core one is comfyui-minimax-h3-audio-T8 v1.86.0). Download the 12 model files into the folders the card names, and rename the upscaler file after download as instructed. Load a character reference into the "参考图1"…
150
How it works
October 1, 2026 briefing recipe read the briefing →
The expensive part of video generation is detail at full resolution. Motion and composition get decided early and cheaply, so the workflow renders those small first, enlarges the rough version, and spends the heavy pass only on drafts you approved.
151
Why it is good
October 1, 2026 briefing recipe read the briefing →
Fast action is where AI video fails most, so being able to throw away a bad take after the cheap pass saves most of the wasted compute. It fits on a mid-range 12 GB card.
Why it is good
Fast action is where AI video fails most, so being able to throw away a bad take after the cheap pass saves most of the wasted compute. It fits on a mid-range 12 GB card.
152
Where it breaks
October 1, 2026 briefing recipe read the briefing →
43.9 GB of downloads, pinned node versions that can conflict with what you already have, and Chinese-language UI labels and prompt template. The int8 model, six add-ons from four different authors and a bridge model all carry their own licenses, so check each before client work.
Where it breaks
43.9 GB of downloads, pinned node versions that can conflict with what you already have, and Chinese-language UI labels and prompt template. The int8 model, six add-ons from four different authors and a bridge model all carry their own licenses, so check each…
153
Two-step H3 in stock ComfyUI
October 1, 2026 briefing recipe read the briefing →
File: minimaxh3pdmd2nfecomfyui.safetensors.
minimaxh3pdmd2nfecomfyui.safetensors ↗
154
The steps
October 1, 2026 briefing recipe read the briefing →
Load your normal H3 graph, add LoraLoaderModelOnly, point it at the file at strength 1.0, set steps to 2, use H3's own scheduler, and turn guidance off. Use the 4-step file (the v6 one; the converter says earlier uploads silently did nothing) when two steps looks too rough.
4-step file ↗
155
How it works
October 1, 2026 briefing recipe read the briefing →
A slow model was used to teach a student to land close to the same result in far fewer passes, and the add-on carries that student's changes.
156
Why it is good
October 1, 2026 briefing recipe read the briefing →
Short waits make it practical to try ten prompts instead of two.
Why it is good
Short waits make it practical to try ten prompts instead of two.
157
Where it breaks
October 1, 2026 briefing recipe read the briefing →
Any strength other than 1.0 changes the behavior, and there is no published quality comparison yet. Expect softer detail than a full-step render and judge it on your own shots.
Where it breaks
Any strength other than 1.0 changes the behavior, and there is no published quality comparison yet. Expect softer detail than a full-step render and judge it on your own shots.
158
Carry forward: the 360 headset workflow from September 30
October 1, 2026 briefing recipe read the briefing →
The H3 equirect add-on recipe from yesterday's briefing remains the best documented end-to-end pipeline this week; see that briefing for steps and limits.
H3 equirect add-on ↗
159
Draft Your AI Fight Scene Before You Render It: A Two-Pass MiniMax H3 Workflow
October 1, 2026 deep dive read the piece →
The most useful idea in the Wuxia Fight Workflow is not the fight choreography but the cheap low-resolution draft you can watch and cancel before the expensive detail pass, a draft-then-commit habit any local video creator can adopt, and its eight-beat prompt template shows that…
160
Make a 360 video and play it in a headset
September 30, 2026 briefing recipe read the briefing →
Files: h3-360-equirect-lora-v1.safetensors on fal's minimax/h3/text-to-video/lora endpoint.
h3-360-equirect-lora-v1.safetensors ↗
161
The steps
September 30, 2026 briefing recipe read the briefing →
Send a request with the add-on's direct file URL at strength 0.75, aspect ratio 21:9, and prompt expansion disabled. Start the prompt with the trigger word eqr360, then the layout sentence from the card, then your scene, describing what is ahead, to the sides, behind and…
162
How it works
September 30, 2026 briefing recipe read the briefing →
The add-on was trained on 181 real 360 clips that passed seam and pole checks, so H3 learns to draw the curved horizon and stretched top and bottom of a sphere laid flat.
163
Why it is good
September 30, 2026 briefing recipe read the briefing →
A 360 establishing shot with sound for $2 and a few minutes, and the card tells you the one setting that trips people up: pass the direct file URL, because the repo-name form loaded differently and gave a different video on the same seed (tested September 29).
Why it is good
A 360 establishing shot with sound for $2 and a few minutes, and the card tells you the one setting that trips people up: pass the direct file URL, because the repo-name form loaded differently and gave a different video on the same seed (tested September 29).
164
Where it breaks
September 30, 2026 briefing recipe read the briefing →
A faint seam can appear directly behind the viewer on busy scenes. Longer renders can darken in the last few seconds, so keep the first ~7.5 s of a 10 s clip. Mono only, so no depth.
Where it breaks
A faint seam can appear directly behind the viewer on busy scenes. Longer renders can darken in the last few seconds, so keep the first ~7.5 s of a 10 s clip. Mono only, so no depth.
165
Direct a voice line by line with an open model
September 30, 2026 briefing recipe read the briefing →
Files: checkpoints/S3/modelbf16.pt and code/infer.py.
checkpoints/S3/modelbf16.pt and code/infer.py ↗
166
The steps
September 30, 2026 briefing recipe read the briefing →
Start with the caption format, because it gets the words right: a CAPTION: describing the voice, then an exact TRANSCRIPT: in quotes. Run infer.py --stage S3 with a frame budget (--frames, in 80 ms frames). Add --reference-wav with a clean clip under about three seconds for…
167
How it works
September 30, 2026 briefing recipe read the briefing →
The model was trained on every one of these prompt formats, so it treats a bracketed note like "(whispered panic, sharp inhale)" as instructions for delivery, not words to speak.
168
Why it is good
September 30, 2026 briefing recipe read the briefing →
CC BY 4.0 weights (check the codec and backbone licences too), and 18,720 sample takes to audition first.
Why it is good
CC BY 4.0 weights (check the codec and backbone licences too), and 18,720 sample takes to audition first.
169
Where it breaks
September 30, 2026 briefing recipe read the briefing →
Word errors in script mode, and the card says it has not shown reliable control over exactly when a laugh or gasp lands. It needs a CUDA machine and some patience.
Where it breaks
Word errors in script mode, and the card says it has not shown reliable control over exactly when a laugh or gasp lands. It needs a CUDA machine and some patience.
170
Write a two-voice scene with sound effects in one pass
September 30, 2026 briefing recipe read the briefing →
Source: the Eleven v4 page and its demo script.
Eleven v4 page ↗
171
The steps
September 30, 2026 briefing recipe read the briefing →
In ElevenCreative's text-to-speech, choose Eleven v4, add a second speaker, and write each line with its delivery in brackets in front of it and sound effects as their own tags, as in ElevenLabs' quiz-show example ([warm], [long pause], [Gong sounds], [crowd applause]).…
172
How it works
September 30, 2026 briefing recipe read the briefing →
v4 reads the whole scene, so a line's delivery can react to the previous one.
173
Why it is good
September 30, 2026 briefing recipe read the briefing →
A radio-play draft, ad read or game bark set without a booth, a mixer or a separate sound-effects pass.
Why it is good
A radio-play draft, ad read or game bark set without a booth, a mixer or a separate sound-effects pass.
174
Where it breaks
September 30, 2026 briefing recipe read the briefing →
You cannot export the effects as separate stems from the same generation as far as the launch material says, so a mixer who wants the gong on its own track will still need a separate pass. Free-plan output is personal use only.
Where it breaks
You cannot export the effects as separate stems from the same generation as far as the launch material says, so a mixer who wants the gong on its own track will still need a separate pass. Free-plan output is personal use only.
175
Make a 360 VR Video With Sound From a Text Prompt, Then Watch It in a Headset
September 30, 2026 deep dive read the piece →
A new add-on for MiniMax H3 turns a written scene into a look-around 360 video with sound for about two dollars, and the craft lies less in the generation than in writing for a viewer who can face any direction and in the two small finishing steps that make a headset recognise…
176
Turn one lucky glitch into a reusable effect
September 29, 2026 briefing recipe read the briefing →
Files: minimaxh3orb360cardspinstep50.safetensors, with the prompts in prompts/cardspincaption.txt and prompts/orbitrealscenecaption.txt.
minimaxh3orb360cardspinstep50.safetensors ↗
177
The steps
September 29, 2026 briefing recipe read the briefing →
In ComfyUI, load a MiniMax-H3 Ref2VA model and add the add-on with a standard Load LoRA node at strength 1.0 (model only). Give it one portrait as the reference and paste the card-spin prompt as the text. Keep 124 frames at 24 fps, because the prompt's timestamps (edge-on at 1.0…
178
How it works
September 29, 2026 briefing recipe read the briefing →
The team was testing an orbit add-on on real photographs and fed it an 1867 Julia Margaret Cameron portrait. The model treated the photograph itself as the object and spun it like a card. They captioned that single generated clip with timestamps and trained 50 more steps on it,…
179
Why it is good
September 29, 2026 briefing recipe read the briefing →
It is a documented recipe for keeping a happy accident. You do not need a new dataset for the effect. You need the clip, an honest caption with times in it, and a short top-up training run on an add-on that already works (here, a 750-step orbit add-on trained on four Blender…
Why it is good
It is a documented recipe for keeping a happy accident. You do not need a new dataset for the effect. You need the clip, an honest caption with times in it, and a short top-up training run on an add-on that already works (here, a 750-step orbit add-on trained…
180
Where it breaks
September 29, 2026 briefing recipe read the briefing →
Learned from one example, so the card always turns the same way at roughly the same speed, and some seeds commit less. The top-up run peaked at 73 GB of graphics memory, so training it needs a rented workstation card even though using it does not. MiniMax H3 Community License,…
Where it breaks
Learned from one example, so the card always turns the same way at roughly the same speed, and some seeds commit less. The top-up run peaked at 73 GB of graphics memory, so training it needs a rented workstation card even though using it does not. MiniMax H3…
181
Make any character perform a clip, in one node
September 29, 2026 briefing recipe read the briefing →
Files: comfyui/workflows/viggle-animate-workflow.json, the ComfyUI-SlimDiT node pack, and minimaxh3ref2vaslimditint8convrot.safetensors.
comfyui/workflows/viggle-animate-workflow.json ↗
182
The steps
September 29, 2026 briefing recipe read the briefing →
Clone the repo, copy ComfyUI-SlimDiT into customnodes, then copy the slimdit folder inside it, which is the step the authors say tripped them up. Download the three upstream files the card lists (a speed-up add-on, a text-conditioning file and the H3 video decoder), install…
183
How it works
September 29, 2026 briefing recipe read the briefing →
Viggle-Animate reads motion from the driving footage and appearance from your still. The node also picks an attention method for each call, based on clip length and free memory: a fast path for short clips, the standard one for long clips.
184
Why it is good
September 29, 2026 briefing recipe read the briefing →
The authors measured their choices. On an RTX 5090, a 124-frame render took 33.2 s on the fast path against 38.3 s standard. On a 372-frame clip, an in-place quantized attention kernel used 31.8 GiB and showed ghosting from around frame 124, while standard attention finished…
Why it is good
The authors measured their choices. On an RTX 5090, a 124-frame render took 33.2 s on the fast path against 38.3 s standard. On a 372-frame clip, an in-place quantized attention kernel used 31.8 GiB and showed ghosting from around frame 124, while standard…
185
Where it breaks
September 29, 2026 briefing recipe read the briefing →
Their own tip is the limit: the still needs to match the driving shot's pose, framing and light, ideally as a repainted frame of that shot, or the character drifts off-model. The checkpoint is 19.6 GiB. H3 licence applies.
Where it breaks
Their own tip is the limit: the still needs to match the driving shot's pose, framing and light, ideally as a repainted frame of that shot, or the character drifts off-model. The checkpoint is 19.6 GiB. H3 licence applies.
186
Swap one loader and run H3 at 4-bit
September 29, 2026 briefing recipe read the briefing →
Files: minimaxh3fl2va4stepquantfuncint4r128.safetensors or the Ref2VA 8-step file, via ComfyUI-QuantFunc.
ComfyUI-QuantFunc ↗
187
The steps
September 29, 2026 briefing recipe read the briefing →
Install or update ComfyUI-QuantFunc. Replace your model loader with the QuantFunc loader, pick the file that matches your mode, and use its matching step count and scheduler. Leave everything else in the workflow alone.
188
How it works
September 29, 2026 briefing recipe read the briefing →
The speed-up add-on and supporting parts are fused into the file, so there is nothing else to attach.
189
Why it is good
September 29, 2026 briefing recipe read the briefing →
A graph you already trust gets faster without being rebuilt. It supports every NVIDIA card from the RTX 20 series up.
Why it is good
A graph you already trust gets faster without being rebuilt. It supports every NVIDIA card from the RTX 20 series up.
190
Where it breaks
September 29, 2026 briefing recipe read the briefing →
The sealed format means no other loader will open it. The quality figure is QuantFunc's own internal comparison. Memory needs are not stated as a number, so test your own resolution before committing a job.
Where it breaks
The sealed format means no other loader will open it. The quality figure is QuantFunc's own internal comparison. Memory needs are not stated as a number, so test your own resolution before committing a job.
191
How to Turn an AI Video Glitch Into a Reusable Effect: The CardSpin Recipe
September 29, 2026 deep dive read the piece →
A single accidental clip, captioned honestly with timestamps and used for a half-hour top-up on an add-on that already works, is enough to turn a model's mistake into an effect you can apply to any photo, and the recipe matters more than the effect itself.
192
Refocus a photograph you already shot
September 28, 2026 briefing recipe read the briefing →
Files: itsmag11/AnyBokeh, stages in stage1/ and stage2/, driven by inferencefull.py from the GitHub repo.
itsmag11/AnyBokeh ↗
193
The steps
September 28, 2026 briefing recipe read the briefing →
Accept the FLUX.1-Fill-dev licence on its model page and sign in; the base model is gated. Run python focuspicker.py --imagepath yourphoto.jpg, open the local page it prints, click the point you want sharp, copy the two numbers. Then run inferencefull.py with --focusx, --focusy,…
194
How it works
September 28, 2026 briefing recipe read the briefing →
The first pass estimates how far every pixel sits from the sharp plane and roughly what is near and far. The second repaints the image from those maps plus your target, so blur follows the scene's geometry rather than a cut-out around the subject. No calibration pass, because…
195
Why it is good
September 28, 2026 briefing recipe read the briefing →
It fixes the one mistake Lightroom cannot. A Gaussian blur behind a cut-out subject has never fooled anyone who looks at photographs, and working from an optical model means edges and out-of-focus highlights behave the way a lens behaves.
Why it is good
It fixes the one mistake Lightroom cannot. A Gaussian blur behind a cut-out subject has never fooled anyone who looks at photographs, and working from an optical model means edges and out-of-focus highlights behave the way a lens behaves.
196
Where it breaks
September 28, 2026 briefing recipe read the briefing →
About 27 GB of graphics memory for the full run puts it on a rented card. S-Lab plus FLUX non-commercial keeps it out of paid work. Training code and dataset are unreleased, so nobody outside the lab can retrain or verify it. And the output is a re-render, not a correction:…
Where it breaks
About 27 GB of graphics memory for the full run puts it on a rented card. S-Lab plus FLUX non-commercial keeps it out of paid work. Training code and dataset are unreleased, so nobody outside the lab can retrain or verify it. And the output is a re-render,…
197
Measure whether your edit model is moving the picture, then stop it
September 28, 2026 briefing recipe read the briefing →
Files: qwen-image-2.1-consistency.safetensors (start with the step-1500 file, not the 2000).
qwen-image-2.1-consistency.safetensors ↗
198
The steps
September 28, 2026 briefing recipe read the briefing →
Run a baseline test first: one picture, ask Qwen-Image-2.1 for a watercolour, lay the result over the original in a difference blend, and see where a hard edge lands. Then add a LoraLoaderModelOnly node straight after the model loader at strength 1.0. Feed your picture into Text…
199
How it works
September 28, 2026 briefing recipe read the briefing →
Trained on pairs where the drift had already been measured and removed, so every target sits exactly on its source's frame. The set includes reverse pairs (turn this watercolour back into a photo), which is why it holds in both directions.
200
Why it is good
September 28, 2026 briefing recipe read the briefing →
Drift is why your before-and-after sliders never line up and your edit passes will not composite. It also stops the model repainting the storefront sign when you asked for a different cardigan. Every number is measured on held-out pictures with seeds and settings stated, and the…
Why it is good
Drift is why your before-and-after sliders never line up and your edit passes will not composite. It also stops the model repainting the storefront sign when you asked for a different cardigan. Every number is measured on held-out pictures with seeds and…
201
Where it breaks
September 28, 2026 briefing recipe read the briefing →
Sample on a latent of any size other than the encoder's output and the model zooms by the size ratio; nothing undoes that. Strengths below 1.0 let drift back. Comic and anime restyles redraw every outline, so shapes still move a few pixels; worst case 12.9 at a corner. Untested…
Where it breaks
Sample on a latent of any size other than the encoder's output and the model zooms by the size ratio; nothing undoes that. Strengths below 1.0 let drift back. Comic and anime restyles redraw every outline, so shapes still move a few pixels; worst case 12.9 at…
202
Put the camera somewhere else without losing the style
September 28, 2026 briefing recipe read the briefing →
Files: lilylilith/QI2.1AnyAngle.
lilylilith/QI2.1AnyAngle ↗
203
The steps
September 28, 2026 briefing recipe read the briefing →
Generate a rough 3D version of your scene with a splat or mesh generator, import it into Blender, place a camera where you want the new shot, render one frame. That render will look coarse, which is fine. Feed both pictures into Qwen-Image-2.1 with this add-on at strength 1.0…
204
How it works
September 28, 2026 briefing recipe read the briefing →
The rough render is the instruction, not the output. It carries the geometry of the new angle and the model repaints your original's style onto it, so the style holds where it drifts when you just ask for an angle in words.
205
Why it is good
September 28, 2026 briefing recipe read the briefing →
Apache 2.0, which is unusual here, and it gives arbitrary angles rather than a menu of presets.
Why it is good
Apache 2.0, which is unusual here, and it gives arbitrary angles rather than a menu of presets.
206
Where it breaks
September 28, 2026 briefing recipe read the briefing →
Everything rests on the 3D stand-in: a coarse or confused splat puts objects in the wrong place, and the card shows this happening to a table and a ponytail. Small or soft faces come back deformed. Minutes per shot, not seconds.
Where it breaks
Everything rests on the 3D stand-in: a coarse or confused splat puts objects in the wrong place, and the card shows this happening to a table and a ponytail. Small or soft faces come back deformed. Minutes per shot, not seconds.
207
Your AI Image Editor Is Moving the Picture. Here Is How to Measure It
September 28, 2026 deep dive read the piece →
Instruction-based image editors redraw the whole frame a few percent off its original position, which is why edit passes never composite cleanly, and the fix starts with a five-minute measurement anyone can run.
208
Watch the picture form instead of waiting for it
September 27, 2026 briefing recipe read the briefing →
Files: the seven decoders in vladmandic/MicroDecoder, named microdecoder-f1, -f2, -qwen, -qwen21, -sd, -sdxl and -wan21.
vladmandic/MicroDecoder ↗
209
The steps
September 27, 2026 briefing recipe read the briefing →
In SDNext, turn on the Micro preview type. It picks the right file for whatever model you have loaded, downloads it into models/Preview, and draws a live preview every step. Outside SDNext, the inference code is under eighty lines and takes the model's working data plus how far…
inference code ↗
210
How it works
September 27, 2026 briefing recipe read the briefing →
A small network trained only to turn working data into a rough picture, fast, with the render's progress fed in as an extra input so early steps do not come out as garbage. It deliberately blurs and then scales up, because a fast honest blur reads better than a fast wrong…
211
Why it is good
September 27, 2026 briefing recipe read the briefing →
Apache 2.0. One file per model family, and sizes are worked out from the file itself, so there is nothing to configure when you switch models. The payoff is not speed, it is that you can abandon a bad generation at step four.
Why it is good
Apache 2.0. One file per model family, and sizes are worked out from the file itself, so there is nothing to configure when you switch models. The payoff is not speed, it is that you can abandon a bad generation at step four.
212
Where it breaks
September 27, 2026 briefing recipe read the briefing →
The author calls it low quality reconstruction and means it: this is a preview, never an output. There is no file for architectures the routing table does not cover, though the training script covers that if you can spend fifteen minutes and a few hundred images. And the example…
Where it breaks
The author calls it low quality reconstruction and means it: this is a preview, never an output. There is no file for architectures the routing table does not cover, though the training script covers that if you can spend fifteen minutes and a few hundred…
213
Write the prompt with the model's own rewriter before you generate
September 27, 2026 briefing recipe read the briefing →
Free hosted: hugging-apps/qwen-image-2-1-prompt-enhancer.
hugging-apps/qwen-image-2-1-prompt-enhancer ↗
214
The steps
September 27, 2026 briefing recipe read the briefing →
Type your short request in any language. Take back a long English prompt and a recommended aspect ratio. Paste both into whatever you are generating with. To run it locally, pull Qwen/Qwen-Image-2.1-PE-T2I and the systemprompt.txt that ships with it.
Qwen/Qwen-Image-2.1-PE-T2I ↗
215
How it works
September 27, 2026 briefing recipe read the briefing →
A vision-language model fine-tuned on the specific job of turning intent into the phrasing this image model responds to, returning structured output rather than prose, so the aspect ratio is a decision the rewriter makes rather than one you guess.
216
Why it is good
September 27, 2026 briefing recipe read the briefing →
It is the enhancer behind the official demo, so you are matching the conditions the sample images were made under. WanPE's cross-generator results this week suggest this is generally worth doing, and worth doing with a good enhancer rather than the default one.
Why it is good
It is the enhancer behind the official demo, so you are matching the conditions the sample images were made under. WanPE's cross-generator results this week suggest this is generally worth doing, and worth doing with a good enhancer rather than the default…
217
Where it breaks
September 27, 2026 briefing recipe read the briefing →
Qwen Research licence on the weights, so local use is evaluation only. The Space is free and shared, so it queues. And a rewriter that expands your six words into two hundred will invent specifics you did not ask for; read what comes back before you generate.
Where it breaks
Qwen Research licence on the weights, so local use is evaluation only. The Space is free and shared, so it queues. And a rewriter that expands your six words into two hundred will invent specifics you did not ask for; read what comes back before you generate.
218
Get ACE-Step 1.5 singing on a Mac instead of babbling
September 27, 2026 briefing recipe read the briefing →
219
The steps
September 27, 2026 briefing recipe read the briefing →
Install mlx-audio from the pc/add-ace branch at commit 1e8264a. Pull roman220220/ACE-Step1.5-sft-MLX-8bit. Then apply the two fixes from the model card: pass uselm=False, and patch the guidance path to use the model's own trained blank input rather than all zeros. The card has…
roman220220/ACE-Step1.5-sft-MLX-8bit ↗
220
How it works
September 27, 2026 briefing recipe read the briefing →
The toolkit assumed every version of this model wants a planner feeding it a rough sketch first. This version does not, and choking on that sketch is what produces noise. Removing it is both correct and about fifteen seconds faster per track.
221
Why it is good
September 27, 2026 briefing recipe read the briefing →
MIT on both halves, licensed training data, 48 kHz stereo, and the author measured intelligibility rather than claiming it.
Why it is good
MIT on both halves, licensed training data, 48 kHz stereo, and the author measured intelligibility rather than claiming it.
222
Where it breaks
September 27, 2026 briefing recipe read the briefing →
Every number is one person on one M5. Word error rate measures whether you can hear the words, not whether the mix is any good, and the author says plainly that the faster sibling sounds fuller and more finished even though the words come through worse. Twelve test tracks is a…
Where it breaks
Every number is one person on one M5. Word error rate measures whether you can hear the words, not whether the mix is any good, and the author says plainly that the faster sibling sounds fuller and more finished even though the words come through worse.…
223
Your Local Music Model Is Not Broken. The Wrapper Around It Is.
September 27, 2026 deep dive read the piece →
A music model that sang gibberish on a Mac turned out to be receiving a message in the wrong format, and the two-line fix is a pattern worth recognising because it fails silently and sounds like the model's fault.
224
Run the six-step Qwen student in ComfyUI without losing the speedup to a bad loader
September 26, 2026 briefing recipe read the briefing →
Files and workflows are in comfyui/: viggleturbo.py goes in ComfyUI/customnodes/, then drag in Qwen-Image-2.1-viggle-turbo-t2i.json or Qwen-Image-2.1-viggle-turbo-edit.json.
comfyui/ ↗
225
The steps
September 26, 2026 briefing recipe read the briefing →
Copy the node file, restart ComfyUI, drag in a workflow, download the four model files the workflow links (int8 transformer 7.3 GB, int8 text encoder 9.4 GB, VAE 0.7 GB, rank-128 adapter 0.7 GB), generate.
226
How it works
September 26, 2026 briefing recipe read the briefing →
Two custom nodes replace stock parts. One supplies the six-value schedule with the resolution-dependent adjustment the reference pipeline applies. The other applies the add-on file at runtime instead of baking it into the weights.
227
Why it is good
September 26, 2026 briefing recipe read the briefing →
Baking it in loses precision. The author measured it: the merge keeps about 70% of the adjustment on average, and as little as 40% in some early layers. The unmerged node costs 10% to 25% more time per step and gets you visibly closer to the reference output.
Why it is good
Baking it in loses precision. The author measured it: the merge keeps about 70% of the adjustment on average, and as little as 40% in some early layers. The unmerged node costs 10% to 25% more time per step and gets you visibly closer to the reference output.
228
Where it breaks
September 26, 2026 briefing recipe read the briefing →
The author says outright that the port was "mostly vibe-coded" with an AI assistant and that they do not use ComfyUI day to day. Tested against ComfyUI 0.37.0. With every model resident it peaks at 26 GB of memory on a 1248x832 image. On an NVIDIA driver older than 580 the…
Where it breaks
The author says outright that the port was "mostly vibe-coded" with an AI assistant and that they do not use ComfyUI day to day. Tested against ComfyUI 0.37.0. With every model resident it peaks at 26 GB of memory on a 1248x832 image. On an NVIDIA driver…
229
Get Ming-Image's design-and-type generation onto a consumer card
September 26, 2026 briefing recipe read the briefing →
Pull mingimage0.1designint4convrot.safetensors into ComfyUI/models/diffusionmodels/.
mingimage0.1designint4convrot.safetensors ↗
230
The steps
September 26, 2026 briefing recipe read the briefing →
Download the 6.3 GB file. Then, and this is the part the file listing does not tell you, pull the mllm/, connector/, mlp/ and vae/ folders from the original repository, because the converted repo ships none of them.
the original repository ↗
231
How it works
September 26, 2026 briefing recipe read the briefing →
The heavy image-making component is stored at reduced precision, layer by layer, with the parts most sensitive to rounding left alone. The components that read your prompt and write out the final picture are untouched.
232
Why it is good
September 26, 2026 briefing recipe read the briefing →
MIT on both halves. You can ship what comes out.
Why it is good
MIT on both halves. You can ship what comes out.
233
Where it breaks
September 26, 2026 briefing recipe read the briefing →
Untested by anyone but the author, who says so. Needs a recent ComfyUI with both Ming-Image and INT4 ConvRot support, which not every install has. And since the type-heavy work is what this model is for, small text is exactly where reduced precision tends to hurt.
Where it breaks
Untested by anyone but the author, who says so. Needs a recent ComfyUI with both Ming-Image and INT4 ConvRot support, which not every install has. And since the type-heavy work is what this model is for, small text is exactly where reduced precision tends to…
234
Generate foley and ambience as separate stems
September 26, 2026 briefing recipe read the briefing →
Mondo Sónico in Google Flow. Open it, describe the scene, export the individual tracks into your edit rather than a single mixed file. Duplicate the tool and change what it asks for if the defaults do not suit your work. Where it breaks: it is a Flow tool, so it lives inside…
Mondo Sónico ↗
235
Google Shipped Six Creative Tools It Didn't Design. One of Them Hands You Separate Audio Stems.
September 26, 2026 deep dive read the piece →
The six new Google Flow Tools matter less as products than as proof that describing the tool you need now beats bending your work to fit someone else's software, and the sound tool proves it by delivering the one thing generated audio almost never does.
236
Take a flat design apart into editable layers
September 25, 2026 briefing recipe read the briefing →
Repo: inclusionAI/Ming-Image. Model: inclusionAI/Ming-Image-0.1-Design-Layer. Sample files: assets/layersamples/cardmakinginput.png and assets/layersamples/cardmakingprompt.txt.
Why it is good
Layer counts are specified rather than guessed, so the output slots into a deck or a mockup without a reconciliation pass. Aspect ratio is preserved. MIT licence on the model.
Where it breaks
The tool always returns one extra composite image first, and the script silently skips it, so if you call the model directly rather than through infer.py your layer numbering will be off by one. Resolution is bucketed to 512 or 1024 only, so fine type in a…
237
Build a speaker-labelled transcript from a two-model pipeline
September 25, 2026 briefing recipe read the briefing →
Guide: ASRINTEGRATIONGUIDE.md in the model repo.
Why it is good
You can swap either half. If you already have transcripts from your editing suite with timecode, you only need the diarization half.
Where it breaks
NVIDIA says this plainly and you should believe them: the midpoint rule is a heuristic. When two people talk at once it marks the word overlap/ambiguous rather than guessing, and words outside detected speech come back unassigned. Cross-talk at the top of a…
238
How to Get a Speaker-Labelled Transcript From Two Models That Never Talk to Each Other
September 25, 2026 deep dive read the piece →
The useful thing about NVIDIA's new speaker model is not its benchmark ranking but its size and its licence, and the pipeline it enables joins two separate models on nothing more than a shared clock.
239
Transparent PNGs out of Qwen-Image-2.1, locally, with no selection tool
September 24, 2026 briefing recipe read the briefing →
The graph is imageqwenimage21t2i.json in Comfy's official template repo, and the edit graph beside it is imageqwenimage21imageedit.json.
imageqwenimage21t2i.json ↗
240
The steps
September 24, 2026 briefing recipe read the briefing →
Drop qwenimage2.1int8convrot.safetensors in ComfyUI/models/diffusionmodels/, qwen3vl8bint8convrot.safetensors in ComfyUI/models/textencoders/, and qwenimage2.1vaebf16.safetensors in ComfyUI/models/vae/. Load the template. Then write your prompt in this exact frame, which is…
241
How it works
September 24, 2026 briefing recipe read the briefing →
The transparency is not a post-process and not a mask node. The model's decompressor carries an alpha channel natively, and the sentence you wrap the prompt in is what tells it to use it. Everything downstream is ordinary: one loader per part, a single text encode node, a…
242
Why it is good
September 24, 2026 briefing recipe read the briefing →
A cutout that comes out of the model has no halo to clean up and no edge to feather, and for repetitive asset work, icons, product shots on white, sticker sheets, it removes an entire step you were doing by hand. The top level is five nodes, two of which are documentation notes,…
Why it is good
A cutout that comes out of the model has no halo to clean up and no edge to feather, and for repetitive asset work, icons, product shots on white, sticker sheets, it removes an entire step you were doing by hand. The top level is five nodes, two of which are…
243
Where it breaks
September 24, 2026 briefing recipe read the briefing →
The template ships at 25 steps while its own note says the official pipeline runs about 40 to 50 with euler, so out of the box you are at roughly half, and thin or wispy edges are where you will see it. Keep cfg at 1 on this path; the note says the negative prompt does nothing…
Where it breaks
The template ships at 25 steps while its own note says the official pipeline runs about 40 to 50 with euler, so out of the box you are at roughly half, and thin or wispy edges are where you will see it. Keep cfg at 1 on this path; the note says the negative…
244
Teaching one model to play two different things from recorded demonstrations
September 24, 2026 briefing recipe read the briefing →
The recipe is examples/games/README.md with the settings documented in docs/embodiments.md.
examples/games/README.md ↗
245
The steps
September 24, 2026 briefing recipe read the briefing →
Record episodes with each game's own scripted bot, which writes one index.json plus one .npz per episode holding frames at 256 by 256 and actions at 15 Hz. Point TrainConfig.dataset at examples.games.dataset:build. Edit the paths in configs/games/train.json, then run torchrun…
246
How it works
September 24, 2026 briefing recipe read the briefing →
One window is 33 frames and 32 actions: one frame you observed, then the frames your actions led to. The caption is the switch. Write the shooter's name and the model behaves like the shooter; write the racer's and it steers. Where the two games have different numbers of…
247
Why it is good
September 24, 2026 briefing recipe read the briefing →
This is the clearest published example of caption-as-router that I have seen, and the idea generalises well past games. If you have recorded sequences of a thing happening and the inputs that caused it, you can train one model over several different kinds of that thing and pick…
Why it is good
This is the clearest published example of caption-as-router that I have seen, and the idea generalises well past games. If you have recorded sequences of a thing happening and the inputs that caused it, you can train one model over several different kinds of…
248
Where it breaks
September 24, 2026 briefing recipe read the briefing →
Eight cards in the published command, and the docs warn that the default warmup phases of 1000/2000/1000 fit a 30,000-step run; at 3,000 steps you must switch to 200/600/200, because with the defaults, in the docs' own words, "the trunk never unfreezes" and stays frozen for the…
Where it breaks
Eight cards in the published command, and the docs warn that the default warmup phases of 1000/2000/1000 fit a 30,000-step run; at 3,000 steps you must switch to 200/600/200, because with the defaults, in the docs' own words, "the trunk never unfreezes" and…
249
Qwen-Image-2.1 Will Hand You a Transparent PNG, But Only If You Say the Magic Sentence
September 24, 2026 deep dive read the piece →
The official ComfyUI templates for Qwen-Image-2.1 can produce genuine transparent cutouts with no selection tool, but the sentence that switches transparency on is hidden inside a note in the workflow file and is missing from the documentation page most people will read instead.
250
Block a shot with rectangles instead of adjectives
September 23, 2026 briefing recipe read the briefing →
Direct link: the Layout condition section of the 2.0 model card, driving examples/minimaxh3fun/predictv2vcontrol.py in VideoX-Fun.
Why it is good
Staging is the thing prompts are worst at. "On the left" is a coin flip; a box at x=230 is not. This also gives you something to show a client before any pixels exist.
Where it breaks
Boxes carry no pose, so a subject can be in the right place doing the wrong thing. The output resolution and aspect ratio are taken from your control video, and frame count snaps down to the largest value the decoder accepts, capped at 15 seconds at 24 fps.…
251
Regenerate one region of a shot and keep the rest
September 23, 2026 briefing recipe read the briefing →
Direct link: the inpainting section and examples/minimaxh3fun/predictv2vcontrolinpaint.py.
Why it is good
This is the video equivalent of a paint-out, and it does not require you to re-render the whole shot to change a sign, a garment or a piece of set dressing.
Where it breaks
Masks must be authored by hand or by a separate segmentation pass, and the model card publishes no numbers on edge fidelity. A soft or shaky mask edge will show. Nothing here preserves audio.
252
Train a style dial in an afternoon on a laptop-class budget
September 23, 2026 briefing recipe read the briefing →
Direct link: the training formulation and reproduction notes.
Why it is good
No image collection, no licensing question about training material, no rights clearance. The output is a small file you can turn up and down.
Where it breaks
The author says plainly that the dial can shift unrelated objects, that the compressed half-size version weakens the effect, and that the fast timing was one run on one card and is not a benchmark. Supra is a 256-pixel model, so faces, fine detail and…
253
Train a Style Control With Six Sentences and Zero Images
September 23, 2026 deep dive read the piece →
A published recipe trains a reusable style control from six pairs of sentences and no training pictures at all, which removes the dataset from the middle of custom style work and turns the style into something you can turn up and down.
254
Re-shoot a clip from a different angle without reshooting it
September 22, 2026 briefing recipe read the briefing →
Adapter: CrossView-Warp v1, file MiniMax-H3Ref2VA-LoRA-CrossView-Warpv13500.safetensors. Node: ComfyUI-CrossViewWarp.
CrossView-Warp v1 ↗
255
The steps
September 22, 2026 briefing recipe read the briefing →
Update to ComfyUI v0.37.0 so Run MoGe Inference is present. Load your clip, run MoGe on it, and wire the mogegeometry output into the CrossView Warp node rather than a relative depth map. Run the graph once to cache the clip. Now orbit the live preview by dragging the picture;…
256
How it works
September 22, 2026 briefing recipe read the briefing →
The node reprojects every pixel of your footage into the new viewpoint using measured geometry, then hands that reprojection to the model as an instruction. Identity and lighting come from the untouched clip on a separate input. You are not describing a camera move, you are…
257
Why it is good
September 22, 2026 briefing recipe read the briefing →
The live preview calls the same code as the final render: measured against a full run, at most 4 pixels out of 147,456 differ. What you audition is what you get, without queueing a job.
Why it is good
The live preview calls the same code as the final render: measured against a full run, at most 4 pixels out of 147,456 differ. What you audition is what you get, without queueing a job.
258
Where it breaks
September 22, 2026 briefing recipe read the briefing →
Stay inside the green zone, about 45 degrees of azimuth and plus 30 to minus 20 of elevation. Yellow runs to 90 azimuth and plus 45 to minus 35, and sideways that is real coverage; vertically it is not, with below minus 20 at 3.1 percent of the training set. The distance control…
Where it breaks
Stay inside the green zone, about 45 degrees of azimuth and plus 30 to minus 20 of elevation. Yellow runs to 90 azimuth and plus 45 to minus 35, and sideways that is real coverage; vertically it is not, with below minus 20 at 3.1 percent of the training set.…
259
Get a usable depth and normal pass out of a single frame
September 22, 2026 briefing recipe read the briefing →
Template: utilitymoge3geometryestimation, shipped with ComfyUI v0.37.0. Weights from Comfy-Org/MoGe, file moge3vitgfp16.safetensors in models/geometryestimation/.
Comfy-Org/MoGe ↗
260
The steps
September 22, 2026 briefing recipe read the briefing →
Load one image, run the template, set refinesteps on the Run MoGe Inference node (default 3, maximum 8, 0 turns the refiner off), and take the depth map and normal map out the other side. For a 360 image, the panorama node splits an equirectangular frame into twelve perspective…
261
How it works
September 22, 2026 briefing recipe read the briefing →
Instead of guessing depth pixel by pixel in the flat image, it builds a rough three-dimensional shell and sharpens the estimate there, which is why thin things stop smearing.
262
Why it is good
September 22, 2026 briefing recipe read the briefing →
No install beyond the weights, because the sparse kernels were already in ComfyUI for other work. The node also reports estimated field of view, which is what you need to match a virtual camera to a real plate.
Why it is good
No install beyond the weights, because the sparse kernels were already in ComfyUI for other work. The node also reports estimated field of view, which is what you need to match a virtual camera to a real plate.
263
Where it breaks
September 22, 2026 briefing recipe read the briefing →
refinesteps does nothing on MoGe 1 or MoGe 2 checkpoints, which the same node still loads, so a wrong file in the dropdown gives you a silently worse result with no error. One image in, one geometry pass out; temporal consistency across a moving shot is not something this…
Where it breaks
Refinesteps does nothing on MoGe 1 or MoGe 2 checkpoints, which the same node still loads, so a wrong file in the dropdown gives you a silently worse result with no error. One image in, one geometry pass out; temporal consistency across a moving shot is not…
264
Change a few notes inside a finished recording
September 22, 2026 briefing recipe read the briefing →
Tool: SpanSynth-Edit, Apache 2.0.
SpanSynth-Edit ↗
265
The steps
September 22, 2026 briefing recipe read the briefing →
Install with the sparse checkout in the README, which skips the demo audio. Align MIDI to your track, edit the notes you want changed, then run spansynth-edit edit --audio track.mp3 --midi revised.mid --edit-start 6.40 --edit-end 14.08 --cfg 2.0 --output results/. If you have…
266
How it works
September 22, 2026 briefing recipe read the briefing →
The model generates only the selected span, conditioned on your MIDI and on the surrounding audio, which is what supplies instrument timbre and room so the patch matches.
267
Why it is good
September 22, 2026 briefing recipe read the briefing →
Outside the edited region the output is bit-identical to the input. That is a guarantee, not a hope, and it makes the tool safe to put in a real session.
Why it is good
Outside the edited region the output is bit-identical to the input. That is a guarantee, not a hope, and it makes the tool safe to put in a real session.
268
Where it breaks
September 22, 2026 briefing recipe read the briefing →
One run covers at most 20.48 seconds, so anything longer is a stitching job you do yourself. Edit boundaries round outward to 40 milliseconds. Output is mono at 48 kHz, so this is a source for a track, not a master. Audio and MIDI must be aligned or offset correctly, and the…
Where it breaks
One run covers at most 20.48 seconds, so anything longer is a stitching job you do yourself. Edit boundaries round outward to 40 milliseconds. Output is mono at 48 kHz, so this is a source for a track, not a master. Audio and MIDI must be aligned or offset…
269
How to Get a Second Camera Angle Out of Footage You Already Shot
September 22, 2026 deep dive read the piece →
The new re-camera workflow is not prompt writing, it is a measuring pass on your own footage plus art direction on the parts the original camera never saw.
270
Generate a transparent sticker sheet with no cutting out
September 21, 2026 briefing recipe read the briefing →
Open the free demo or load the ComfyUI template at templates/imageqwenimage21t2i.json.
free demo ↗
271
The steps
September 21, 2026 briefing recipe read the briefing →
Write the prompt in the three-part shape the team documents, because the model was trained on it: open with "This is an RGBA image with transparency," then your subject, then close with "The image has alpha channel and the background is transparent." Set 40 steps. Pick a size…
272
How it works
September 21, 2026 briefing recipe read the briefing →
The compressor that turns the model's internal picture back into pixels has a fourth channel for opacity, so transparency is something the model generates rather than something you subtract afterwards. That is why the edges hold on hair and glass, where background removal…
273
Why it is good
September 21, 2026 briefing recipe read the briefing →
It removes a whole step and the artefacts that come with it. No halo, no matte line, no cleanup pass.
Why it is good
It removes a whole step and the artefacts that come with it. No halo, no matte line, no cleanup pass.
274
Where it breaks
September 21, 2026 briefing recipe read the briefing →
The opening and closing sentences are load-bearing; drop them and you will get an opaque square. The licence is research and evaluation only, so a sticker pack you sell is out until someone buys a commercial licence. And the ComfyUI download is large because the prompt-reading…
Where it breaks
The opening and closing sentences are load-bearing; drop them and you will get an opaque square. The licence is research and evaluation only, so a sticker pack you sell is out until someone buys a commercial licence. And the ComfyUI download is large because…
275
Edit by drawing on the picture instead of masking it
September 21, 2026 briefing recipe read the briefing →
Template: templates/imageqwenimage21imageedit.json.
templates/imageqwenimage21imageedit.json ↗
276
The steps
September 21, 2026 briefing recipe read the briefing →
Load your image. Draw a rough circle around each region you want changed, or paint over it, straight onto the input. Write one instruction per region in plain language. Queue once. The model card's worked example changes three things in a single pass: removes a watch, changes…
277
How it works
September 21, 2026 briefing recipe read the briefing →
The prompt-reading model takes text and images through the same path, so a circle on the picture is an instruction in the same sense a word is. You are not supplying a precise selection, you are pointing.
278
Why it is good
September 21, 2026 briefing recipe read the briefing →
Circles take seconds. Clean masks take minutes, and for three regions you would otherwise be doing three passes.
Why it is good
Circles take seconds. Clean masks take minutes, and for three regions you would otherwise be doing three passes.
279
Where it breaks
September 21, 2026 briefing recipe read the briefing →
Pointing is not selecting. Where two regions overlap, or where the circle covers something you wanted untouched, you get a judgement call rather than a boundary, and you will not know which until you look. Multiple reference images are also where memory goes: the prompt-reading…
Where it breaks
Pointing is not selecting. Where two regions overlap, or where the circle covers something you wanted untouched, you get a judgement call rather than a boundary, and you will not know which until you look. Multiple reference images are also where memory goes:…
280
Compress the encoder, not the model
September 21, 2026 briefing recipe read the briefing →
Reference implementation: FastWan2.2-TI2V-5B-MLX-q8, file t5encoderq8.safetensors.
FastWan2.2-TI2V-5B-MLX-q8 ↗
281
The steps
September 21, 2026 briefing recipe read the briefing →
Check what your pipeline loads and how big each piece is. If a full-precision prompt-reading model is in there, swap in an 8-bit one. In the video case above that is 6.0 GB against 11.4 GB on disk and 22.7 GB resident. On the image side, Comfy-Org ship the equivalent files…
282
How it works
September 21, 2026 briefing recipe read the briefing →
The encoder runs once per generation and is then released. Rounding its weights harder costs you a little precision in a step that happens once, not in the step that repeats forty times.
283
Why it is good
September 21, 2026 briefing recipe read the briefing →
It is the single largest memory saving available in most local setups, and unlike compressing the generation model it does not change the picture's trajectory at all, only the prompt's reading of it.
Why it is good
It is the single largest memory saving available in most local setups, and unlike compressing the generation model it does not change the picture's trajectory at all, only the prompt's reading of it.
284
Where it breaks
September 21, 2026 briefing recipe read the briefing →
Four bits destroys it. Cosine 0.911 means the model is answering a different prompt. And a loader that is not expecting compressed weights will fail on shapes rather than on a missing file, which is exactly why the author renamed the file.
Where it breaks
Four bits destroys it. Cosine 0.911 means the model is answering a different prompt. And a loader that is not expecting compressed weights will fail on shapes rather than on a missing file, which is exactly why the author renamed the file.
285
Your Picture Model Is Not What Fills Your Graphics Card. The Prompt Reader Is.
September 21, 2026 deep dive read the piece →
Two independent measurements published on the same weekend found that the model reading your prompt is the largest thing resident in a local generation pipeline, and shrinking that file costs nothing visible while shrinking the picture model costs you quality for a fraction of…
286
Chain three video shots into one take that holds the same face and voice
September 20, 2026 briefing recipe read the briefing →
Direct link: LaDruid/MiniMax-H3-Multishot-Workflow. Start with workflows/H3SeamlessChainCORE.json, which needs nothing but the pack and stock ComfyUI.
LaDruid/MiniMax-H3-Multishot-Workflow ↗
287
The steps
September 20, 2026 briefing recipe read the briefing →
Copy ComfyUI-H3-Multishot into ComfyUI/customnodes/ and restart. Pull the H3 GGUF checkpoint at the quality your card takes, plus the text encoder and both VAE files (there are two, video and audio) from Comfy-Org/MiniMax-H3. Open the CORE graph. Type your shots into the…
Comfy-Org/MiniMax-H3 ↗
288
How it works
September 20, 2026 briefing recipe read the briefing →
Each shot conditions on the tail of the one before it, so the model is handed the previous shot's ending rather than starting fresh. Identity lives in that conditioning, not the seed, which is why varying the seed per shot holds the face while reusing one seed everywhere drifts…
289
Why it is good
September 20, 2026 briefing recipe read the briefing →
Every number is attached to a decision. Setting memoryframes to 0 drops brightness drift from 1.055 to 1.022 per join and colour drift from 1.086 to 1.039, and stops the drift accelerating. The memory reserve sizes itself: a hand-set value that suited shot 1 was wrong for shot…
Why it is good
Every number is attached to a decision. Setting memoryframes to 0 drops brightness drift from 1.055 to 1.022 per join and colour drift from 1.086 to 1.039, and stops the drift accelerating. The memory reserve sizes itself: a hand-set value that suited shot 1…
290
Where it breaks
September 20, 2026 briefing recipe read the briefing →
Texture ratchets about 1.3x at every join and keeps climbing, with no honest fix after the fact, because blur is the only lever and it destroys real detail. Set chaingaincontrol to flatten past five shots and accept one house texture. Resolution cannot change mid-chain. Per-shot…
Where it breaks
Texture ratchets about 1.3x at every join and keeps climbing, with no honest fix after the fact, because blur is the only lever and it destroys real detail. Set chaingaincontrol to flatten past five shots and accept one house texture. Resolution cannot change…
291
Find out whether your image add-on is actually loaded
September 20, 2026 briefing recipe read the briefing →
Direct link: flux2-klein-peft/ in dptlab, specifically scripts/mergeandexport.py.
flux2-klein-peft/ ↗
292
The steps
September 20, 2026 briefing recipe read the briefing →
Load your base model, load the add-on, render one image. Then set the adapter strength to zero and render the same prompt at the same seed. Identical images mean the add-on was never loaded. For any format that is not LoRA, skip the test and run mergeandexport.py first, which…
293
How it works
September 20, 2026 briefing recipe read the briefing →
The standard loader looks for a specific key naming scheme in the file. LoRA files have it. DoRA, LoHa, LoKr, OFT and BOFT do not, so the loader finds nothing, logs a line most people never read, and proceeds with the untouched model. You get an image. It looks fine. It is not…
294
Why it is good
September 20, 2026 briefing recipe read the briefing →
Two renders, and it catches a class of failure that produces no error at all. It also explains a forum complaint you have seen: "my trained file does nothing."
Why it is good
Two renders, and it catches a class of failure that produces no error at all. It also explains a forum complaint you have seen: "my trained file does nothing."
295
Where it breaks
September 20, 2026 briefing recipe read the briefing →
The zero-strength trick only works on formats the loader can read in the first place, which is the joke. And merging costs you the thing that made add-ons useful, because a merged model is a full copy per subject rather than a small file you swap per job.
Where it breaks
The zero-strength trick only works on formats the loader can read in the first place, which is the joke. And merging costs you the thing that made add-ons useful, because a merged model is a full copy per subject rather than a small file you swap per job.
296
Chain Three AI Video Shots Into One Take, and the Drift Fix Everyone Gets Backwards
September 20, 2026 deep dive read the piece →
Chained AI video shots drift because each shot regenerates from the previous shot's own output, and the per-shot colour and brightness corrections most people reach for cannot fix it, because they operate on the picture you see rather than the data being handed forward.
297
Train your own music style on one consumer graphics card
September 19, 2026 briefing recipe read the briefing →
Direct link: the MLTNT model card, which documents its training run, its failed runs and its prompt testing in more detail than most music model cards bother with.
MLTNT model card ↗
298
The steps
September 19, 2026 briefing recipe read the briefing →
Collect songs in the style you want, as FLAC, trimmed under 320 seconds with a fade. Caption each one as a single descriptive sentence in a fixed order: language, genre, vocal, instruments, mood, BPM, production. Measure the BPM with librosa.beat.beattrack rather than guessing.…
FSAudio Suite ↗
299
How it works
September 19, 2026 briefing recipe read the briefing →
YuE2 has two halves. One writes the score and decides the arrangement. The other turns that plan into sound. One training run produces adapters for both, in one file. That matters when you prompt: the decoder dial changes how the record sounds, the planner dial changes what gets…
300
Why it is good
September 19, 2026 briefing recipe read the briefing →
Everything here is testable. The author states that pushing decoder strength from 1.0 to 1.5 makes the sound darker and rougher without touching the writing, and that pushing planner strength to 1.25 collapses the vocal, with the receipt: one sung bar against 424 rests.
Why it is good
Everything here is testable. The author states that pushing decoder strength from 1.0 to 1.5 makes the sound darker and rougher without touching the writing, and that pushing planner strength to 1.25 collapses the vocal, with the receipt: one sung bar against…
301
Where it breaks
September 19, 2026 briefing recipe read the briefing →
Three places, all documented. The trainer names its best checkpoint by the planner's score, which keeps improving while the sound quality peaks early and then degrades, so the file the trainer recommends can be the burnt one. Pick by the sound curve instead. Automatic lyric…
Where it breaks
Three places, all documented. The trainer names its best checkpoint by the planner's score, which keeps improving while the sound quality peaks early and then degrades, so the file the trainer recommends can be the burnt one. Pick by the sound curve instead.…
302
Halve an image model's work without changing what it makes
September 19, 2026 briefing recipe read the briefing →
Direct link: radames/FLUX.2-klein-Sana-Sprint.
radames/FLUX.2-klein-Sana-Sprint ↗
303
The steps
September 19, 2026 briefing recipe read the briefing →
Load black-forest-labs/FLUX.2-klein-4B in diffusers 0.37 or later. Call pipe.loadloraweights("radames/FLUX.2-klein-Sana-Sprint"). Generate at numinferencesteps=2. Ignore the guidance setting; the base ignores it too. For editing, pass reference images alongside the prompt.
304
How it works
September 19, 2026 briefing recipe read the briefing →
The adapter was trained to jump the whole distance in one move rather than four small ones, using the technique from the SANA-Sprint paper. Because it sits on top rather than replacing anything, setadapters(["default0"], [0.0]) gives you the stock model back without reloading.
SANA-Sprint paper ↗
305
Why it is good
September 19, 2026 briefing recipe read the briefing →
Two passes instead of four, on a model that already fits modest hardware, with editing preserved. And the same file runs in a browser through WebGPU, so you can hand someone a link instead of a build.
Why it is good
Two passes instead of four, on a model that already fits modest hardware, with editing preserved. And the same file runs in a browser through WebGPU, so you can hand someone a link instead of a build.
306
Where it breaks
September 19, 2026 briefing recipe read the briefing →
Above 512 pixels it gets inconsistent, so this is an iteration tool, not a final-render tool. And the warmer color cast stays unless you blend the adapter down to 0.7, which costs you some of the speed.
Where it breaks
Above 512 pixels it gets inconsistent, so this is an iteration tool, not a final-render tool. And the warmer color cast stays unless you blend the adapter down to 0.7, which costs you some of the speed.
307
Your Music Model Trainer Saves a File Called best. Do Not Use It.
September 19, 2026 deep dive read the piece →
The number a music style trainer optimises has almost nothing to do with whether the result sounds like the band, and two independent sources now say the file it labels best is usually the wrong one.
308
Try the faster H3 on someone else's GPU before you commit 82 GB and a licence problem
September 18, 2026 briefing recipe read the briefing →
309
The steps
September 18, 2026 briefing recipe read the briefing →
Open akhaliq/vdn-minimax-h3-demo, which runs the 8-step VDN-H3 checkpoint and returns video with a soundtrack. Write your prompt the way MiniMax asks for it: the VDN card explicitly recommends rewriting the prompt through H3-Context-IR or the official prompt-writing skills…
akhaliq/vdn-minimax-h3-demo ↗
310
How it works
September 18, 2026 briefing recipe read the briefing →
The team's own measurement is that plain attention eats more than 85% of H3's runtime. Their fix splits the video's attention in two. Nearby frames get exact, full-strength attention in a sliding window, which is what keeps local detail and short-term stability. Distant frames…
311
Why it is good
September 18, 2026 briefing recipe read the briefing →
It is a bolt-on. The checkpoint adds a separate branch plus two small adapters that merge into the backbone at render time without touching the base weights, so this is not a fork of H3 you have to maintain. And the training code shipped with the weights, which is rarer than it…
Why it is good
It is a bolt-on. The checkpoint adds a separate branch plus two small adapters that merge into the backbone at render time without touching the base weights, so this is not a fork of H3 you have to maintain. And the training code shipped with the weights,…
312
Where it breaks
September 18, 2026 briefing recipe read the briefing →
The headline nine seconds needs eight B200s. On one card you are at 51 seconds (B200) or 90.5 seconds (H200) for the 8-step version, which is still a different working day from 14 minutes but is not interactive. The published figures exclude loading, warm-up and the decode step.…
Where it breaks
The headline nine seconds needs eight B200s. On one card you are at 51 seconds (B200) or 90.5 seconds (H200) for the 8-step version, which is still a different working day from 14 minutes but is not interactive. The published figures exclude loading, warm-up…
313
Pick your LTX-2.5 file by your card, not by the download button
September 18, 2026 briefing recipe read the briefing →
314
The steps
September 18, 2026 briefing recipe read the briefing →
Go to comatto/LTX-2.5-Distilled-GGUF and read the size table before you click anything. Grab one video file: LTX-2.5-Distilled-Q4KM.gguf at 15.7 GB is the maintainer's recommended baseline, Q3KS at 12.6 GB is the floor, Q80 at 23.6 GB is described as nearly indistinguishable…
comatto/LTX-2.5-Distilled-GGUF ↗
315
How it works
September 18, 2026 briefing recipe read the briefing →
LTX-2.5 is a split-component model: the generator, the part that reads your prompt, and the parts that turn numbers back into pixels and audio all live in separate files. The GGUF files here only compress the generator. That is why the quantized file size is not your memory…
316
Why it is good
September 18, 2026 briefing recipe read the briefing →
Seven tiers with published sizes means you can size the download to the card you actually have. The included workflow files skip the usual hour of wiring loaders by hand. And the LTX-2.x community licence permits commercial and production use at no cost for entities under $10M…
Why it is good
Seven tiers with published sizes means you can size the download to the card you actually have. The included workflow files skip the usual hour of wiring loaders by hand. And the LTX-2.x community licence permits commercial and production use at no cost for…
317
Where it breaks
September 18, 2026 briefing recipe read the briefing →
Budget for the encoder and the decoders or you will run out of memory with a correctly sized generator file loaded. The repo's own links point at a different account, so double-check you are pulling from the repo you think you are. And the model card's own limitations list is…
Where it breaks
Budget for the encoder and the decoders or you will run out of memory with a correctly sized generator file loaded. The repo's own links point at a different account, so double-check you are pulling from the repo you think you are. And the model card's own…
318
Your LTX-2.5 Download Is Only a Third of the Bill
September 18, 2026 deep dive read the piece →
The compressed LTX-2.5 video file everyone links to is one of four pieces you need, and the piece nobody mentions is the same size again, which is why sizing the download by your card still runs you out of memory.
319
Write an Ideogram brief instead of an Ideogram prompt
September 17, 2026 briefing recipe read the briefing →
model card ↗
320
The steps
September 17, 2026 briefing recipe read the briefing →
The full schema is published at ideogram-oss/ideogram4/docs/prompting.md. Three top-level fields, only the third required: highleveldescription, a sentence or two; styledescription, holding aesthetics, lighting, medium, either a photo or an artstyle field, and an optional…
ideogram-oss/ideogram4/docs/prompting.md ↗
321
How it works
September 17, 2026 briefing recipe read the briefing →
The guide says it outright: the model was trained exclusively on these structured captions, and "passing in plain-text prompts directly to the model will not work and will likely trigger a safety warning." The bounding boxes are the part worth sitting with. You are not…
322
Why it is good
September 17, 2026 briefing recipe read the briefing →
It rewards what a designer already does: separate content from treatment, list the pieces, place them. It re-rolls cleanly, because changing one element's description leaves the rest of the composition mostly intact. And the palette control is real typographic control, not a…
Why it is good
It rewards what a designer already does: separate content from treatment, list the pieces, place them. It re-rolls cleanly, because changing one element's description leaves the rest of the composition mostly intact. And the palette control is real…
323
Where it breaks
September 17, 2026 briefing recipe read the briefing →
Small text is still imperfect, and the 4-bit build is slightly worse at it than the original, so set type at poster scale. Hex must be uppercase full-length RRGGBB; shorthand is rejected. The safety filter returns a grey "Image blocked by safety filter" card, and Ideogram say…
Where it breaks
Small text is still imperfect, and the 4-bit build is slightly worse at it than the original, so set type at poster scale. Hex must be uppercase full-length RRGGBB; shorthand is rejected. The safety filter returns a grey "Image blocked by safety filter" card,…
324
Get 11 GB back on your card by swapping out a video model's text encoder
September 17, 2026 briefing recipe read the briefing →
ClipProj ↗
325
The steps
September 17, 2026 briefing recipe read the briefing →
Install the node from github.com/nicolab28/ComfyUI-ClipProj (151 stars). Put the projection files in ComfyUI/models/clipprojections/. Start with mmh3-4b-ClipProj-v3.1, 26 MB of projection plus about 4.6 GB for its encoder. Before you trust anything, run the two control files,…
github.com/nicolab28/ComfyUI-ClipProj ↗
326
How it works
September 17, 2026 briefing recipe read the briefing →
Every Qwen3-VL shares one tokenizer, so a prompt breaks into the same pieces at the same positions in a 4B and a 32B. That makes it possible to fit a plain position-by-position translation between what the small model produces and what the big one would have produced. The main…
327
Why it is good
September 17, 2026 briefing recipe read the briefing →
It is 11 GB of headroom on the exact card you already own, it is MIT, and the author published the failures alongside the wins, including three of his own claims he later corrected. The best of those corrections is a lesson for anyone who reads model cards: a change that moved…
Why it is good
It is 11 GB of headroom on the exact card you already own, it is MIT, and the author published the failures alongside the wins, including three of his own claims he later corrected. The best of those corrections is a lesson for anyone who reads model cards: a…
328
Where it breaks
September 17, 2026 briefing recipe read the briefing →
Named people were the original weakness and the -celeb files fix most of it, but characters identified by a mask or a costume still come out as a stranger in the right outfit, and people famous before photography come out generic. The measured similarity to the original…
Where it breaks
Named people were the original weakness and the -celeb files fix most of it, but characters identified by a mask or a costume still come out as a stranger in the right outfit, and people famous before photography come out generic. The measured similarity to…
329
Ideogram 4 Does Not Want Your Sentence. It Wants a Layout Spec.
September 17, 2026 deep dive read the piece →
Ideogram 4 was trained only on structured captions with coordinates and hex colours, so typing a sentence at it is not prompting badly, it is using the wrong interface entirely.
330
Turn your own line drawing into a finished render with Krea 2
September 16, 2026 briefing recipe read the briefing →
ComfyUI-Krea2-Ostris-Edit ↗
331
The steps
September 16, 2026 briefing recipe read the briefing →
Install ComfyUI-Krea2-Ostris-Edit and comfyuicontrolnetaux (4.2k stars). Download one of the two files from Cierpliwy/krea2lineartedit into ComfyUI/models/loras/: krea2linearteditv1.safetensors follows your lines closely, krea2lineartcoarseeditv1.safetensors interprets them…
ComfyUI-Krea2-Ostris-Edit ↗
332
How it works
September 16, 2026 briefing recipe read the briefing →
The adapters were trained on lines produced by the Realistic Lineart preprocessor, so they expect that specific line character. Your drawing carries the structure; the prompt carries everything else. Starting from a photo rather than a sketch, push it through that same…
333
Why it is good
September 16, 2026 briefing recipe read the briefing →
You keep the composition. A sketch is a far more precise instruction than a paragraph, and this is a free way to give Krea 2, which people picked specifically for its look, a drawing to follow. I have not found another on the hub.
Why it is good
You keep the composition. A sketch is a far more precise instruction than a paragraph, and this is a free way to give Krea 2, which people picked specifically for its look, a drawing to follow. I have not found another on the hub.
334
Where it breaks
September 16, 2026 briefing recipe read the briefing →
Line style is the whole ball game: hand-drawn lines with a different weight or texture than the Realistic Lineart output will drift, and the loose variant drifts on purpose. The weights sit under the Krea 2 Community License, not an open one, so check the terms before commercial…
Where it breaks
Line style is the whole ball game: hand-drawn lines with a different weight or texture than the Realistic Lineart output will drift, and the loose variant drifts on purpose. The weights sit under the Krea 2 Community License, not an open one, so check the…
335
Pull a relighting stack out of a single photograph
September 16, 2026 briefing recipe read the briefing →
336
The steps
September 16, 2026 briefing recipe read the briefing →
Open the ShadeNet-2 Space and upload a plate, or run it locally: pip install onnxruntime pillow numpy, grab onnx/modelquantized.onnx (28 MB) from the repo, feed a [1, 3, H, W] array in the range minus one to one at a size that is a multiple of 32, and read back [1, 8, H, W].…
ShadeNet-2 Space ↗
337
How it works
September 16, 2026 briefing recipe read the briefing →
The shading channel has no ground truth anywhere in training. Its only supervision is the requirement that albedo multiplied by shading comes back out looking like the input. That one constraint forces a split between what colour a surface is and how much light is landing on it,…
338
Why it is good
September 16, 2026 briefing recipe read the briefing →
Your normal and depth maps arrive in the same pass and the same registration as the albedo, so you can drop a new light on a photographed object without rebuilding geometry. All Apache 2.0, all of it in 28 MB.
Why it is good
Your normal and depth maps arrive in the same pass and the same registration as the albedo, so you can drop a new light on a photographed object without rebuilding geometry. All Apache 2.0, all of it in 28 MB.
339
Where it breaks
September 16, 2026 briefing recipe read the briefing →
Depth is relative, so do not read distances off it. Any strongly coloured light source, a sunset, a neon sign, a practical with a gel, bleeds into the albedo instead of staying in the shading layer. That is the failure that will bite you first on real footage. Normals go soft in…
Where it breaks
Depth is relative, so do not read distances off it. Any strongly coloured light source, a sunset, a neon sign, a practical with a gel, bleeds into the albedo instead of staying in the shading layer. That is the failure that will bite you first on real…
340
A 28 MB File Separates the Light From the Paint in Any Photograph
September 16, 2026 deep dive read the piece →
ShadeNet-2 pulls a relighting stack out of one photo on a CPU for free, and it works because its lighting layer is trained on nothing but the requirement that colour times light has to equal the original picture.
341
Teach Krea 2 an edit it was never trained to do
September 15, 2026 briefing recipe read the briefing →
From the krea2 architecture source in ai-toolkit and the ComfyUI-Krea2-Ostris-Edit README.
krea2 architecture source ↗
342
The steps
September 15, 2026 briefing recipe read the briefing →
Build a paired dataset: source image as control, edited result as target, matching filenames. Set the architecture to krea2 and model.modelkwargs.edit: true. Leave kvcache off. No Krea 2 example config exists in config/examples on main, so start from…
343
How it works
September 15, 2026 briefing recipe read the briefing →
Reference images enter twice. Through the Qwen3-VL text encoder next to the prompt, downscaled to fit 384x384 total pixels, because the language side only needs a coarse read. And separately as clean latents, capped at one megapixel, appended to the image sequence at timestep…
344
Why it is good
September 15, 2026 briefing recipe read the briefing →
The instruction is a picture, on a base model chosen for how it looks rather than what it can edit. Two working adapters appeared in a day.
Why it is good
The instruction is a picture, on a base model chosen for how it looks rather than what it can edit. Two working adapters appeared in a day.
345
Where it breaks
September 15, 2026 briefing recipe read the briefing →
Three silent traps. Stock Krea 2 in ComfyUI ignores reference latents entirely, so without the patch node you get a normal render and no warning. The text encoder checkpoint must include the Qwen3-VL vision weights or your images cannot be encoded at all. And the kvcache toggle…
Where it breaks
Three silent traps. Stock Krea 2 in ComfyUI ignores reference latents entirely, so without the patch node you get a normal render and no warning. The text encoder checkpoint must include the Qwen3-VL vision weights or your images cannot be encoded at all. And…
346
Run a three-step audio-video adapter without the runtime it was built for
September 15, 2026 briefing recipe read the briefing →
From the TaoMate-H3 3-step ComfyUI card.
TaoMate-H3 3-step ComfyUI card ↗
347
The steps
September 15, 2026 briefing recipe read the briefing →
Put TaoMate-H3-3step-ComfyUI.safetensors (1.24 GB) into ComfyUI/models/loras/. Load MiniMax H3 FL2VA as the base. Load the adapter with the standard loader at strength 1.0. Set steps to 3, CFG to 1.0, sampler Euler, scheduler simple. Confirm the console prints 208 patches…
hosted Space ↗
348
How it works
September 15, 2026 briefing recipe read the briefing →
The adapter was distilled so three passes land where many more used to, which is why guidance has to be off. The distilled path has already absorbed the guidance behaviour; pushing more on top of it breaks the trajectory.
349
Why it is good
September 15, 2026 briefing recipe read the briefing →
Synchronized audio and video from one generation, at 3 steps, on a single card, from a file you can download in a minute.
Why it is good
Synchronized audio and video from one generation, at 3 steps, on a single card, from a file you can download in a minute.
350
Where it breaks
September 15, 2026 briefing recipe read the briefing →
Everything the runtime did beyond speed is gone: no chunked streaming, no prompt-per-five-seconds, no continuity across prompt boundaries, no long-form. The licence excludes four major markets. The conversion publishes no speed or memory measurement of its own, so the 11.45x…
Where it breaks
Everything the runtime did beyond speed is gone: no chunked streaming, no prompt-per-five-seconds, no continuity across prompt boundaries, no long-form. The licence excludes four major markets. The conversion publishes no speed or memory measurement of its…
351
Catch a compressed model that renders fine and is wrong anyway
September 15, 2026 briefing recipe read the briefing →
Carried forward as method, prompted by the Z-Image iQ4.5 card's own warning.
352
The steps
September 15, 2026 briefing recipe read the briefing →
Before trusting any compressed model served through a long-running local server, send three visibly different prompts in sequence in one session, same seed, then send the same three to a freshly restarted server. If the sequential run drifts toward the earlier prompts, the…
353
How it works
September 15, 2026 briefing recipe read the briefing →
Servers cache prompt embeddings to avoid re-encoding. If the cache key is wrong or missing, later requests reuse earlier conditioning. Nothing errors.
354
Why it is good
September 15, 2026 briefing recipe read the briefing →
Ten minutes, and it catches the exact bug the mlx-serve fix addresses, in a class of failure that produces a perfectly plausible image.
Why it is good
Ten minutes, and it catches the exact bug the mlx-serve fix addresses, in a class of failure that produces a perfectly plausible image.
355
Where it breaks
September 15, 2026 briefing recipe read the briefing →
It catches conditioning leaks only. Whether the compression itself hurt quality needs a separate side-by-side against the full-precision weights.
Where it breaks
It catches conditioning leaks only. Whether the compression itself hurt quality needs a separate side-by-side against the full-precision weights.
356
Krea 2 Cannot Edit Images. Here Is How People Are Teaching It To Anyway
September 15, 2026 deep dive read the piece →
A reference picture is a more precise instruction than any sentence, and the new Krea 2 edit path lets you teach the model one specific edit from a folder of before-and-after pairs, with three silent failure modes standing between you and a working result.
357
Change one frame and let the change travel through the shot
September 14, 2026 briefing recipe read the briefing →
From the LTX Ripple card, using the shipped Workflowexample/LTXRippleFFAFEditv11.json.
LTX Ripple card ↗
358
The steps
September 14, 2026 briefing recipe read the briefing →
Put LTX25Ripplev11.safetensors in ComfyUI/models/loras/. Install ComfyUI-LTXVideo through the Manager and accept the gated LTX 2.5 files. Load the shipped workflow. Export frame zero of your clip, edit it in whatever image editor you already own, and load the edited frame back…
ComfyUI-LTXVideo ↗
359
Paint flat colour blobs, get smoke and fire that follows them
September 14, 2026 briefing recipe read the briefing →
From AInVFX's September 13 write-up and the ainvfx-fluid card, with a 45-minute video walkthrough.
September 13 write-up ↗
360
The steps
September 14, 2026 briefing recipe read the briefing →
Put ainvfx-fluid.safetensors (654 MB) in ComfyUI/models/loras/ and load it with LTX IC-LoRA Loader Model Only at strength 1.0. Build a control video the same size and length as your shot: dimensions a multiple of 64, 121 frames, your painted first frame at index 0, your painted…
361
Check whether an upscaler is actually helping before you trust it
September 14, 2026 briefing recipe read the briefing →
Carried forward as method, prompted by the SeedVR2 ncnn card's own reporting.
362
The steps
September 14, 2026 briefing recipe read the briefing →
Before adopting any restoration or upscaling model, build a three-way comparison on your own footage: the source, the model's output, and plain bicubic resize to the same target size. Score all three against the same reference with the same metric, and look at them at 100…
363
Paint Six Flat Colours, Get Fire: The Drawing Is the Real Prompt
September 14, 2026 deep dive read the piece →
In this kind of tool the painted control frames carry the art direction and the text prompt carries almost nothing, and the team published enough of the recipe that the method transfers to any effect you need.
364
Get a real ten-second LTX 2.5 video out of a 24 GB card
September 13, 2026 briefing recipe read the briefing →
Download ltx-2.5-22b-distilled-transformer-w4a8.safetensors from JoaoZaokk/LTX-2.5-22B-distilled-W4A8-ConvRot into ComfyUI/models/diffusionmodels/, paired with the gemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot encoder the measurement used.
JoaoZaokk/LTX-2.5-22B-distilled-W4A8-ConvRot ↗
365
The steps
September 13, 2026 briefing recipe read the briefing →
Launch with --disable-dynamic-vram. Set 249 frames at 25 fps, 512×512, 3 steps, cfg 1.0, euler, and enter the sigmas manually: 0.909375, 0.725, 0.421875, 0.0. Frame count must satisfy 8n+1, so 250 is rejected and 249 is the legal neighbour. Expect roughly 400 seconds on a 3090.…
366
Train a control type your model does not have yet, for about ten GPU-hours
September 13, 2026 briefing recipe read the briefing →
The Anima Canny adapter card publishes the whole recipe, which makes it a template rather than an artifact.
Anima Canny adapter card ↗
367
The steps
September 13, 2026 briefing recipe read the briefing →
Clone the feat/anima-vace-controlnet branch of TaihoC/sd-scripts-animaCN. Assemble roughly 15,000 captioned images at 1024×1024. Generate edge maps with cv2.Canny at fixed thresholds 100 and 200, not randomized. Attach four control blocks to base blocks 0, 7, 14 and 21. Train…
TaihoC/sd-scripts-animaCN ↗
368
Swap a faster autoregressive checkpoint into YuE2 without forking the pipeline
September 13, 2026 briefing recipe read the briefing →
From the YuE2-3B-NVFP4 card.
YuE2-3B-NVFP4 card ↗
369
The steps
September 13, 2026 briefing recipe read the briefing →
pip install "yue2infer[fast]", run the repo's patchfast.py against your installed yue2/fast.py to add a YUE2ARCHECKPOINT environment override, download the quantized checkpoint, point that variable at it, and call YuE2Pipeline.frompretrained("m-a-p/YuE2-3B", backend="vllm",…
370
Ten Hours on One GPU Buys You a ControlNet: The Anima Canny Recipe, Printed in Full
September 13, 2026 deep dive read the piece →
The release worth copying is the hyperparameter list, not the safetensors file, because it turns a missing control type from a lab problem into a ten-hour errand.
371
Clean up a bad recording locally with AuK in ComfyUI, and know in advance which fixes will hold
September 12, 2026 briefing recipe read the briefing →
Install ComfyUI-AuK into customnodes/, install its requirements against your ComfyUI Python, then pull three files from drbaph/AuK-comfyui: one diffusion model (aukbasebf16.safetensors at 3.062 GB, or aukbasew4a8.safetensors at 0.880 GB if you are tight), one encoder…
ComfyUI-AuK ↗
372
The steps
September 12, 2026 briefing recipe read the briefing →
Load your audio, type the instruction, set seconds to 0 to match the source, render. How it works. The Qwen2.5-Omni encoder listens to the clip directly, so you never supply a transcript; the instruction and the audio become one conditioning tensor that the diffusion transformer…
cookbook ↗
373
Change a scene mid-shot instead of cutting to a new one
September 12, 2026 briefing recipe read the briefing →
Evoke's segment mode conditions each chunk separately, so the prompt can change while the rollout is running. From a cloned Evoke repo with weights and ViGeo in place, run MODE=segment NUMCHUNKS=6 MAXCASES=0 bash scripts/inference/inferpostdistill.sh.
Evoke ↗
374
The steps
September 12, 2026 briefing recipe read the briefing →
Six chunks at 36 frames each gives 213 frames, roughly 8.9 seconds, with the prompt switching at chunk 3. The bundled cases include an aurora igniting over a frozen tundra and a meteor shower opening over a still lake. How it works. The world state bank keeps the geometry you…
375
Generate a full song locally on a mid-range card
September 12, 2026 briefing recipe read the briefing →
Build the dev branch of audio.cpp, download the YuE2-3B GGUF directory including sidecars/, and point --model at the directory, not at a single file. The steps. audiocppcli --task gen --family yue2 --model --backend cuda --threads 8 --lyrics $'[Verse]\n...' --request-option…
audio.cpp ↗
376
Change the Prompt Mid-Shot With Evoke, and Watch the One Number That Says When It Stopped Remembering
September 12, 2026 deep dive read the piece →
Evoke can repaint a scene mid-rollout because the geometry lives outside the model in a point-cloud bank, and that same architecture prints a per-chunk coverage number that is the only reliable signal that the model has stopped remembering the place and started inventing it.
377
Cover an existing song by editing its score, not its audio
378
The steps
September 10, 2026 briefing recipe read the briefing →
Transcribe the source recording with SheetSage2 and save the melody ABC without chord symbols as melody.abc. Get the lyrics separately, either transcribed with Qwen3-ASR or found and checked by hand, organized into sections that match the recording, saved as coverlyrics.txt.…
379
How it works
September 10, 2026 briefing recipe read the briefing →
Melody-only planning tells the model to honor the pitch contour you supplied and invent everything else. The published SHS100K numbers show what that buys: with a full score, CLEWS Hit@1 is 71.3 percent; without chords, 67.3 percent; without any score at all, 0.3 percent. Song…
380
Why it is good
September 10, 2026 briefing recipe read the briefing →
You are editing notation, which is diffable, versionable and reviewable by a musician who has never touched a model.
Why it is good
You are editing notation, which is diffable, versionable and reviewable by a musician who has never touched a model.
381
Where it breaks
September 10, 2026 briefing recipe read the briefing →
cot="melody" does not strip chord symbols for you, so a score exported with harmony will silently constrain the arrangement you were trying to replace. The transcription tools want their own Python environments. And the same table shows the cost of control: dropping the score…
Where it breaks
Cot="melody" does not strip chord symbols for you, so a score exported with harmony will silently constrain the arrangement you were trying to replace. The transcription tools want their own Python environments. And the same table shows the cost of control:…
382
Restyle a shot you already filmed, using a keyframe you painted
September 10, 2026 briefing recipe read the briefing →
Carried forward deliberately: this is July work (HF createdAt July 15, code pushed July 29) that today's mirror surfaced again, and it is the most complete worked example of the editable-intermediate idea anyone has published. (ID-V2V restylization recipe)
ID-V2V restylization recipe ↗
383
The steps
September 10, 2026 briefing recipe read the briefing →
Build a sample directory holding source.mp4, stylizedfirstframe.png and prompt.txt, optionally with keyframes/40.png and keyframes/80.png for mid or end anchors. Run SAMPLEDIR=... SAMPROMPT=person bash scripts/preprocess.sh, then bash scripts/examples/restylization.sh.
384
How it works
September 10, 2026 briefing recipe read the briefing →
Preprocessing uses SAM3 to segment the person and gray out the background, writing preprocessing/origpixel.mp4. That foreground-on-gray video is the single VACE control condition. The model relights whatever is inside the mask and regenerates everything outside it from your…
385
Why it is good
September 10, 2026 briefing recipe read the briefing →
The output includes fliptest.mp4, which plays your result and freezes at up to five evenly spaced frames to alternate generated against source. A built-in A/B is a small thing that changes how fast you can tell whether it worked.
Why it is good
The output includes fliptest.mp4, which plays your result and freezes at up to five evenly spaced frames to alternate generated against source. A built-in A/B is a small thing that changes how fast you can tell whether it worked.
386
Where it breaks
September 10, 2026 briefing recipe read the briefing →
The published hardware is 8× A100-80GB. Checkpoints run about 96 GB and SAM3 needs a gated-repo login. Pick the wrong one of the two .pth files and nothing errors, you just get worse video. And the authors call it a research exploration released "for demonstration and…
Where it breaks
The published hardware is 8× A100-80GB. Checkpoints run about 96 GB and SAM3 needs a gated-repo login. Pick the wrong one of the two .pth files and nothing errors, you just get worse video. And the authors call it a research exploration released "for…
387
ID-V2V Lets Your Keyframe Be Wrong
September 10, 2026 deep dive read the piece →
ID-V2V treats identity preservation as a relighting problem, which is why a mismatched stylized keyframe still works and why the workflow is reproducible from one video.
388
Generate fifteen seconds of video with native audio on one AMD card, then verify you are not being sold a number
September 9, 2026 briefing recipe read the briefing →
Package at EliovpAI/MiniMax-H3-W4A8-Paiton-RDNA4. Files: launch.sh, run.sh, comfyui/workflows/turbo8-15s.json, checkpoints.lock.json, BENCHMARKS.md.
EliovpAI/MiniMax-H3-W4A8-Paiton-RDNA4 ↗
389
The steps
September 9, 2026 briefing recipe read the briefing →
Clone github.com/Eliovp-BV/paiton-vllm-plugin, then run ./models/MiniMax-H3/launch.sh. Open 127.0.0.1:8190/?paiton=1, which loads the Turbo8 fifteen-second workflow with the engine toggle already in the graph. Edit the prompt, fix the seed, click Run. For terminal work, stop the…
390
How it works
September 9, 2026 briefing recipe read the briefing →
The compiled denoiser fuses Turbo residuals, paired SwiGLU and RDNA4 INT8 attention into one artifact rather than replacing the model. Both engines load the exact same pinned W4A8 checkpoint and Turbo adapter, so the comparison isolates the runtime. Frame counts snap upward to…
391
Why it is good
September 9, 2026 briefing recipe read the briefing →
Fifteen continuous seconds with a coherent subject and a generated soundtrack, on one card, with no per-generation billing and nothing leaving the machine. And the evidence is unusually complete: raw per-run timings, file hashes, sampled GPU energy, disk reads per phase and…
Why it is good
Fifteen continuous seconds with a coherent subject and a generated soundtrack, on one card, with no per-generation billing and nothing leaving the machine. And the evidence is unusually complete: raw per-run timings, file hashes, sampled GPU energy, disk…
392
Where it breaks
September 9, 2026 briefing recipe read the briefing →
Linux with Docker and working /dev/kfd and /dev/dri only, one specific GPU, and 60 GB of disk before anything renders. Audio clips at full scale on some prompts, as measured above. Four-step pouring duplicates the bottle; eight-step pouring overfoams. The card states outright…
Where it breaks
Linux with Docker and working /dev/kfd and /dev/dri only, one specific GPU, and 60 GB of disk before anything renders. Audio clips at full scale on some prompts, as measured above. Four-step pouring duplicates the bottle; eight-step pouring overfoams. The…
393
Turn a hosted editing model you pay for into a small local LoRA you own
September 9, 2026 briefing recipe read the briefing →
Worked example at DirtScan/trail-delighter, Apache 2.0 over FLUX.2 Klein 4B. Teacher endpoint: fal's Qwen Image Edit lighting-restoration LoRA.
DirtScan/trail-delighter ↗
394
The steps
September 9, 2026 briefing recipe read the briefing →
Collect a set of your own source images that share the exact problem you keep fixing by hand. Run a selection of them through the hosted model that already solves it well, and keep both the input and the output for each one. That paired set is your teacher set. Train a LoRA on…
395
How it works
September 9, 2026 briefing recipe read the briefing →
You are not distilling the teacher's weights, which you do not have. You are distilling its behavior on your specific distribution of images, through a dataset. The hosted model is a labeling service you rent for one afternoon. The student never needs the teacher's capacity…
396
Why it is good
September 9, 2026 briefing recipe read the briefing →
It converts a recurring per-image API cost into a fixed one-time cost plus a file you keep. It works for any repetitive edit you already outsource, whether that is delighting, background cleanup, colour matching or product straightening. And it gives you a local, offline path…
Why it is good
It converts a recurring per-image API cost into a fixed one-time cost plus a file you keep. It works for any repetitive edit you already outsource, whether that is delighting, background cleanup, colour matching or product straightening. And it gives you a…
397
Where it breaks
September 9, 2026 briefing recipe read the briefing →
Every error the teacher makes becomes a rule the student learns, and there is no correction step in this loop. The published example ships no dataset size, no rank, no learning rate and no evaluation, so the recipe is sound but the settings are yours to find. The result is…
Where it breaks
Every error the teacher makes becomes a rule the student learns, and there is no correction step in this loop. The published example ships no dataset size, no rank, no learning rate and no evaluation, so the recipe is sound but the settings are yours to find.…
398
Turn a Hosted Image Editor Into a Local LoRA You Own
September 9, 2026 deep dive read the piece →
Renting a large hosted editing model for one afternoon to label your own images, then training a small local adapter on those pairs, converts a recurring per-image API bill into a fixed cost plus a file you keep, and the real constraint is the service terms rather than the model…
399
Generate a 36-second continuous shot on two consumer GPUs using a keyframe storyboard
September 8, 2026 briefing recipe read the briefing →
Pipeline at pottokao/minimax-h3-2x16gb. Files: pipeline/storylong.sh, pipeline/prompts.json, serveturbo4p.sh.
pottokao/minimax-h3-2x16gb ↗
400
The steps
September 8, 2026 briefing recipe read the briefing →
Clone with git lfs install first, because the weights are in the repo. Provision swap before anything else if you are on 32 GB: sudo fallocate -l 24G /swapfile, mkswap, swapon, then add it to fstab. Build with bash build.sh or install the bundled pre-patched engine with pip…
401
How it works
September 8, 2026 briefing recipe read the briefing →
Chaining first-last segments naively drifts, because each one pins only its opening frame and the character degrades over a few hops. This pipeline generates one short "fast-forward" clip of the entire action first, samples N keyframes out of it in a single pass so they are…
402
Why it is good
September 8, 2026 briefing recipe read the briefing →
The published example is seven segments at four steps: a 107-second montage plus seven 73-second segments, about 619 seconds of compute for 36 seconds of finished video, with character identity holding head to tail. That is roughly seventeen seconds of compute per second of…
Why it is good
The published example is seven segments at four steps: a 107-second montage plus seven 73-second segments, about 619 seconds of compute for 36 seconds of finished video, with character identity holding head to tail. That is roughly seventeen seconds of…
403
Where it breaks
September 8, 2026 briefing recipe read the briefing →
Audio is discontinuous at every seam and the fix is not shipped. Blackwell sm120 only. The 4/8-step distilled plus W4A4 plus 768p combination has a fidelity ceiling the card states directly. Sixty gigabytes of disk before you generate anything. And the text encoder is an…
Where it breaks
Audio is discontinuous at every seam and the fix is not shipped. Blackwell sm120 only. The 4/8-step distilled plus W4A4 plus 768p combination has a fidelity ceiling the card states directly. Sixty gigabytes of disk before you generate anything. And the text…
404
Tune the speed-quality dial on a Mac with measurements instead of vibes
September 8, 2026 briefing recipe read the briefing →
Runtime at antirez/h3.c. Requires FFmpeg and FFprobe on PATH plus a local MiniMax-H3 snapshot.
antirez/h3.c ↗
405
The steps
September 8, 2026 briefing recipe read the briefing →
make -j8, then ./h3 --info -d ./MiniMax-H3 to check layout and print the selected Metal device without mapping weights. Start at the validated balanced preset: 512×512, 22 frames, --steps 20 --layers 45 --reuse 2 --profile. When you need a fast look, drop to --steps 4 --layers…
406
How it works
September 8, 2026 briefing recipe read the briefing →
Four independent dials, each documented with what it costs. Denoising passes is literal: --steps N runs exactly N. Whole-denoiser reuse evaluates the first and last pass plus every requested interval and extrapolates the rest, so 20 steps at --reuse 2 is 11 fresh model…
407
Why it is good
September 8, 2026 briefing recipe read the briefing →
Every tradeoff comes with a number. The four-pass path measured 0.556 full-video SSIM against a 29-pass reference on the fox test and 0.547 on an independent surfer test, at 3.5 seconds of denoise on an M5 Max versus 26.4 for the reference. Token reduction took a 512-square…
Why it is good
Every tradeoff comes with a number. The four-pass path measured 0.556 full-video SSIM against a 29-pass reference on the fox test and 0.547 on an independent surfer test, at 3.5 seconds of denoise on an M5 Max versus 26.4 for the reference. Token reduction…
408
Where it breaks
September 8, 2026 briefing recipe read the briefing →
The docs name a specific poisoned combination: do not add --token-reduction to both --layers 40 and --reuse 3, which produced chromatic ringing, outlines and ghosted limbs despite acceptable latent norms. SSD streaming cannot be combined with --use-int8-row-fc2. Frames snap…
Where it breaks
The docs name a specific poisoned combination: do not add --token-reduction to both --layers 40 and --reuse 3, which produced chromatic ringing, outlines and ghosted limbs despite acceptable latent norms. SSD streaming cannot be combined with…
409
MiniMax H3 on Apple Silicon: Every Shortcut in h3.c Ships With the Number It Cost
September 8, 2026 deep dive read the piece →
Publishing the exchange rate for every speed shortcut, rather than only the win, is what turns an optimization document into a tool you can actually tune to your own quality floor.
410
Build a drum kit from text prompts and export it straight into Ableton
September 7, 2026 briefing recipe read the briefing →
App at petalsfm/foundry. File: Foundry.dmg, 3,391,504,096 bytes, Apple silicon.
petalsfm/foundry ↗
411
The steps
September 7, 2026 briefing recipe read the briefing →
Install and open it; no account, no API key, no network. Write a prompt into a pad, one per sound: "tight room snare, close mic, short decay", "dusty vinyl kick". Render, then use the per-pad editor to trim the waveform, set the pitch envelope, shape the transient and assign a…
412
How it works
September 7, 2026 briefing recipe read the briefing →
Each pad is an independent eight-step Stable Audio 3 Small SFX generation at CFG 1.0, two seconds by default, rendered at fp32 on MPS with CPU fallback. The morph pad is the interesting one: four corner prompts, bilinearly blended, so you get a continuous field between "brushed"…
413
Why it is good
September 7, 2026 briefing recipe read the briefing →
The whole loop from idea to a mapped drum rack happens on one machine with no per-generation cost and no upload, and the output is a normal sample you own the file of. Three retries on a spectral-flatness floor of 0.55 means the obvious failure, a pad that comes back as hiss,…
Why it is good
The whole loop from idea to a mapped drum rack happens on one machine with no per-generation cost and no upload, and the output is a normal sample you own the file of. Three retries on a spectral-flatness floor of 0.55 means the obvious failure, a pad that…
414
Where it breaks
September 7, 2026 briefing recipe read the briefing →
Two seconds, clamped to a three-second maximum, so this is percussion and one-shots and nothing else. Apple silicon only. The Stability Community License terminates above one million dollars in annual revenue and asks you to register for commercial use, and the Gemma Terms of…
Where it breaks
Two seconds, clamped to a three-second maximum, so this is percussion and one-shots and nothing else. Apple silicon only. The Stability Community License terminates above one million dollars in annual revenue and asks you to register for commercial use, and…
415
Generate sound design on a Mac from the command line, with the settings that actually work
September 7, 2026 briefing recipe read the briefing →
Assets at arraypress/stable-audio-3-small-sfx-coreai. Files: small-sfx.aimodel (2.9 GB), tokenizer/ (34 MB), example.py (self-contained, roughly 120 lines).
arraypress/stable-audio-3-small-sfx-coreai ↗
416
The steps
September 7, 2026 briefing recipe read the briefing →
pip install coreai-torch tokenizers numpy, which brings the Core AI runtime and nothing else; no stable-audio-tools, no PyTorch. Then python3 example.py "loud crackling campfire with crickets" out.wav --seconds 24. To continue or repair existing audio, pass --init-audio in.wav…
417
How it works
September 7, 2026 briefing recipe read the briefing →
The DiT accepts any length; the encoder and decoder are compiled at fixed sizes, so the asset ships a ladder of decoder graphs and you pick the smallest one your clip fits. That matters more than it sounds: decoding the same 24 seconds costs 0.30 s and 0.5 GB through the 47.6 s…
418
Why it is good
September 7, 2026 briefing recipe read the briefing →
Twenty-four seconds of 44.1 kHz stereo in about 2.21 seconds on an M-series Mac, roughly twice as fast as the same model under MLX with 8-bit weights, and the audio-in path gives you continuation and inpainting rather than only text-to-audio. Every fidelity claim in the card…
Why it is good
Twenty-four seconds of 44.1 kHz stereo in about 2.21 seconds on an M-series Mac, roughly twice as fast as the same model under MLX with 8-bit weights, and the audio-in path gives you continuation and inpainting rather than only text-to-audio. Every fidelity…
419
Where it breaks
September 7, 2026 briefing recipe read the briefing →
Compile time on first load is brutal and grows fast: on the 2 B decoder, 23.8 seconds of audio compiles in 12 s, 47.6 s in 44 s, 95.1 s in 8.5 minutes, and 190.2 s in over an hour. Precompiling produces an .aimodelc that runs only on the architecture it was built for. Memory,…
Where it breaks
Compile time on first load is brutal and grows fast: on the 2 B decoder, 23.8 seconds of audio compiles in 12 s, 47.6 s in 44 s, 95.1 s in 8.5 minutes, and 190.2 s in over an hour. Precompiling produces an .aimodelc that runs only on the architecture it was…
420
This Kokoro Core ML Conversion Hands You a Duration for Every Phoneme. That Is the Whole Subtitle Problem, Solved for Free.
September 7, 2026 deep dive read the piece →
Splitting a TTS model at its duration predictor turns narration timing from a second forced-alignment pass into a free byproduct of synthesis, and the conversion that does it also publishes what a broken conversion sounds like.
421
Shoot a 1980s broadcast day, with the correct base checkpoint
September 6, 2026 briefing recipe read the briefing →
LoRA at KennethFal/vh5tape-vhs-lora-minimax-h3. Files: vh5tape.safetensors, or vh5tape-comfyui.safetensors for the version with explicit alpha tensors. Prompts and seeds for every sample: PROMPTS.md.
KennethFal/vh5tape-vhs-lora-minimax-h3 ↗
422
The steps
September 6, 2026 briefing recipe read the briefing →
Drop the file in ComfyUI/models/loras/. Load a base that is not rotated: minimaxh3fl2vabf16.safetensors is the recommendation, minimaxh3fl2vaprunedfp8scaled.safetensors is acceptable. Wire UNETLoader into LoraLoaderModelOnly into MiniMaxH3ImageToVideo. Strength 1.0, useful range…
423
How it works
September 6, 2026 briefing recipe read the briefing →
Trigger and damage phrase were both trained through H3's own caption format, so the tier phrases behave like a dial rather than keywords you hope land. Because H3 generates audio in the same pass, describing the sound ("muffled mono audio", "canned laughter", "cheesy jingle")…
424
Why it is good
September 6, 2026 briefing recipe read the briefing →
No conversion step, all 104 modules load without warnings, and the aesthetic is one phrase you can turn up or down mid-project instead of a grade you apply afterward.
Why it is good
No conversion step, all 104 modules load without warnings, and the aesthetic is one phrase you can turn up or down mid-project instead of a grade you apply afterward.
425
Where it breaks
September 6, 2026 briefing recipe read the briefing →
Stacked with a turbo LoRA, drop this one to 0.4 to 0.6: the two perturbations add and a four-step path has no headroom, and four-step turbo softens the tape grain, which is the opposite of the point. Stay near 1 MP, single shot per prompt. The base-checkpoint rule above is not…
Where it breaks
Stacked with a turbo LoRA, drop this one to 0.4 to 0.6: the two perturbations add and a four-step path has no headroom, and four-step turbo softens the tape grain, which is the opposite of the point. Stay near 1 MP, single shot per prompt. The base-checkpoint…
426
Find your own four-bit threshold before you trust a converted checkpoint
September 6, 2026 briefing recipe read the briefing →
Bench at JoaoZaokk/comfy-quant-bench. Worked example with all three builds published: Wan2.1-VACE-1.3B-W4A4-ConvRot.
JoaoZaokk/comfy-quant-bench ↗
427
The steps
September 6, 2026 briefing recipe read the briefing →
Convert your model at more than one promotion threshold. Render the same prompt at the same six seeds through each build and through the unquantized reference. Look at the pictures. Run probequantdispatch.py --forward-only to confirm the quantized math is executing rather than…
428
How it works
September 6, 2026 briefing recipe read the briefing →
"Effective error" is measured per layer, on the real activations that layer saw during sampling, against the format that layer actually received. For Wan 2.1 VACE 1.3B, a build at 0.0546 (2 of 300 layers at four bits, 2.15 GiB) matches the FP16 reference; 0.0793 keeps the…
429
Why it is good
September 6, 2026 briefing recipe read the briefing →
It kills the idea that a converter default is safe. The author's version: --promote-error 0.15, chosen once on a 6B model and carried everywhere since, "writes a file here that loads cleanly, dispatches natively, passes every structural check, and renders a smear."
Why it is good
It kills the idea that a converter default is safe. The author's version: --promote-error 0.15, chosen once on a 6B model and carried everywhere since, "writes a file here that loads cleanly, dispatches natively, passes every structural check, and renders a…
430
Where it breaks
September 6, 2026 briefing recipe read the briefing →
One prompt, three to six seeds, 480x480, one scheduler, one card, no perceptual metric, and the size monotonicity rests on three data points. The page also carries two dated self-corrections from September 1, and they are the most instructive part: the earlier renders ran at cfg…
Where it breaks
One prompt, three to six seeds, 480x480, one scheduler, one card, no perceptual metric, and the size monotonicity rests on three data points. The page also carries two dated self-corrections from September 1, and they are the most instructive part: the…
431
There Is No Safe Setting for 4-Bit Quantization. There Is One Per Model.
September 6, 2026 deep dive read the piece →
A converted checkpoint can load cleanly, dispatch natively, pass every structural check and still render a smear, so the only setting you can trust is the one you measured on your own model.
432
Generate a 360 video and get it into a headset tonight
September 5, 2026 briefing recipe read the briefing →
LoRA at shamanic/minimax-h3-equi360-lora. Weights file: h3-equi360-lora-step2500.safetensors. Packaging tool: google/spatial-media.
Why it is good
Prompt-expansion off and scale 1.0 are stated, not guessed. The card publishes a same-seed A/B showing that the base model prompted identically produces a poor approximation, so the adapter is doing real work. And the entire training pipeline is in scripts/,…
Where it breaks
Four places. The seam at the 180-degree wrap can show a faint discontinuity, because the trainer offers no circular convolution. The base model's camera-motion prior invents dolly and pan on some prompts, which smears distortion across the sphere; "static…
433
Run the 8-step VDN speedup locally in ComfyUI
September 5, 2026 briefing recipe read the briefing →
Bundle at t8star/Vdn-Minimax-H3-Comfy. Nodes at T8mars/comfyui-minimax-h3-audio-T8. Workflows: examples/workflows/10-speed.
Why it is good
It is the rare case where a community port is stricter than the thing it ports. The validation matrix ran all nine multimodal modes serially on an RTX 4060 Ti 16 GB at 320x192x39, logged zero LoRA errors, applied 310 of 310 curve-Turbo patches every time, and…
Where it breaks
Minimum free VRAM across that pruned matrix was 290 MiB, and only T2VA and I2VA, the text-to-video-and-audio and image-to-video-and-audio modes, cleared the project's own 512 MiB margin. Every mode in that matrix generates audio alongside video, which is part…
434
Teach a video model your own frame format for about $12.50
September 5, 2026 briefing recipe read the briefing →
Reference pipelines: VR180 (training/pipelineconfig.yaml, training/faltrainerrequest.json) and 360 (scripts/train.mjs, training/trainer-config.json).
Why it is good
Both authors published their loss curves, their bucket math, their pilot failures and their costs. The 360 run reports loss falling from 1.12 to 0.43 over 2500 steps and about $0.30 of non-GPU spend on captions and test generations. Two independent runs…
Where it breaks
The VR180 author's pilots are the warning: 1000 steps at 21:9 produced a usable stereo pair on one prompt in four, and 1000 steps at 16:9 on zero in four. Aspect bucket and step count are both load-bearing, and you will spend $12.50 discovering that if you…
435
Two People Taught a Video Model to Shoot VR for $12.50 Each. The Recipe Is One Sentence.
September 5, 2026 deep dive read the piece →
A format LoRA does not teach a video model geometry, it teaches a frame layout bound to a trigger by one caption sentence repeated across every training clip, which is why fifty clips and twelve dollars are enough.
436
Steer a character and a camera through five seconds of generated video with WASD, in ComfyUI
September 4, 2026 briefing recipe read the briefing →
Node pack at T8mars/comfyui-minimax-h3-audio-T8. LoRA at t8star/Minimax-H3-World-Comfy. Workflow file: examples/workflows/26-h3-world/2026-09-04H3WorldI2VA832x480124f50stepAdvanced.json.
Why it is good
The file is 131 MB and byte-identical to the researchers' own upload, verified by a published SHA-256, so you are not trusting a community requantization of anything. It adds no new pip dependencies. And a blind test on a fixed parking-lot sample had users…
Where it breaks
Four places. The contract is fixed, so 832x480 and 124 frames are not suggestions; a first frame at another aspect ratio gets proportionally covered and center-cropped, which cuts the edges off your subject without warning you. You cannot stack this with…
437
Watermark and C2PA-sign a render without leaving the graph, offline
September 4, 2026 briefing recipe read the briefing →
Node at CreativeMayhemLtd/ComfyUI-Provcheck, CLI at CreativeMayhemLtd/provcheck.
Why it is good
It never leaves your machine, it needs no account, and it fails closed: a missing binary or a missing identity passes your media through with a console note rather than killing a long render. includefullworkflow is off by default so prompts and paths do not…
Where it breaks
The image path needs ONNX Runtime 1.22.x, and if the runtime is missing or mismatched the image mark silently does nothing, which is the worst possible failure mode for a provenance tool. When signing, the artifact is the file written to output/, not the…
438
Generate 48 kHz foley from a text prompt, free, with a license you can actually use
September 4, 2026 briefing recipe read the briefing →
Model at OpenMOSS-Team/MOSS-SoundEffect-v2.0. Try it first at the hosted Space.
Why it is good
Apache 2.0 on both the code and the weights, which is rare enough in generative audio to be the headline. 48 kHz output is delivery quality, not preview quality. And the hosted Space means you can audition it before committing to an 11 GB download.
Where it breaks
100 solver steps is a lot for iteration, so expect to wait between takes. The pipeline wraps the DiT in torch.compile plus a Triton CUDA graph, which means the first call can take several minutes to compile; set TORCHDYNAMODISABLE=1 if you hit Dynamo or…
439
Provcheck Signs Your ComfyUI Renders From Inside the Graph. Check Two Settings First.
September 4, 2026 deep dive read the piece →
A provenance node that runs inside ComfyUI can attest the generation graph itself, which no downstream signer can do, but two of its defaults will hand a first-time user an unmarked file or someone else's brand id.
440
Generate MiniMax H3 video and audio in eight steps instead of many, inside ComfyUI
September 2, 2026 briefing recipe read the briefing →
Node pack at Deno2026/comfyui-deno-custom-nodes. Weights at alibaba-pai/MiniMax-H3-Acc-LoRAs. Note on versions: the README states a ComfyUI 0.30.0 floor for the pack's two MiniMax H3 reference-image nodes, and states no version floor at all for the Acc LoRA Loader itself, so you…
Why it is good
It uses Alibaba's official file unconverted, so you are not trusting a community requantization. Sampler, scheduler and step count stay in stock ComfyUI nodes you already know. And the repo hosts side-by-side result videos against both the un-accelerated…
Where it breaks
Three places. If you run a Comfy-Org pruned checkpoint, which is exactly what low-VRAM users grab, the loader needs a matching non-pruned checkpoint present under models/diffusionmodels/ to derive an AdaLN bridge; without one it warns once and silently skips…
441
Roto a moving subject with local models, no account and no cloud
September 2, 2026 briefing recipe read the briefing →
OpenShot 4.0, free, GPLv3.
Why it is good
Your footage never leaves the machine, there is no AI account, and there is no subscription gate on the feature. The mask is reusable rather than baked, so one roto pass feeds several effects.
Where it breaks
OpenShot publishes no VRAM or timing figures for masking, and says plainly that performance depends on your computer and the chosen model. The point-prompt loop is interactive, so a subject that changes shape often will cost you a lot of clicking. The…
442
MiniMax H3 Now Generates Video in Eight Steps Inside ComfyUI. The License Excludes the United States.
September 2, 2026 deep dive read the piece →
Alibaba's official eight-step acceleration adapters for MiniMax H3 became loadable in ComfyUI five days after release, and the workflow is genuinely good, but the base model's community license names the United States as an excluded territory and the pruned checkpoint most…
443
Render a 1080p video longer than your model's ceiling, on 16 GB
September 1, 2026 briefing recipe read the briefing →
Node pack at hradec/ComfyUI-HR-Endless-Sampler. MiniMax H3 only for now; LTX-2.5 is described as planned.
Why it is good
You keep the identity references across every chunk, you can hover a chunk in the preview to read the exact prompt Gemma wrote for it, and debugstartchunk lets you re-render one late shot without resampling everything before it.
Where it breaks
Four things, all stated by the author. Chunk two costs more VRAM than chunk one because it carries the continuation tail, so a setting that survives the first chunk can still die on the second. Chunking only reduces the temporal part of the memory cost, so…
444
Drive a camera through a generated scene with a locked character
September 1, 2026 briefing recipe read the briefing →
Repo at Riemann-Dynamics/Matrix-Game-3.5.
Why it is good
Memoryvisualization.mp4 is the rare debugging artifact that tells you why a generation drifted rather than just showing you that it did.
Where it breaks
40 GB VRAM floor, Linux only, 64 GB system RAM. Each block is 80 frames and consumes 84 camera poses, and a trajectory shorter than 1 + 84 x numblocks poses is padded by holding the last pose, which will read as a dead camera at the end of your shot. You…
445
The Fix for AI Video That Forgets Is an Editable Text File
September 1, 2026 deep dive read the piece →
HR Endless Sampler renders arbitrarily long MiniMax H3 video on a 16 GB card by running Gemma 4 as a continuity supervisor between chunks, and because the entire supervision policy ships as an editable prompt file, continuity has become a text problem you can debug.
446
Chain generations into a scene instead of a shot
August 31, 2026 briefing recipe read the briefing →
Node pack at sandpies/ComfyUI-Hand-Tie-Clips, renamed from ComfyUI-H3-Ref-Chain on August 29 with three point releases on August 30. MIT.
Why it is good
H3 tops out at 15 seconds in one pass and resets the character every time you start again. This carries identity and voice across the join, so you get a scene rather than a pile of unrelated shots.
Where it breaks
Three honest failures from the author. Audio drifts against picture by about 40 milliseconds per hop, so a six-hop chain is a quarter second out by the end. "Texture still ratchets on long chains. Stay around 3-5 hops until that is handled." And a graph saved…
447
Get HD out of a draft-resolution generation
August 31, 2026 briefing recipe read the briefing →
Workflow at zuanfilm/H3HD2KDetailer, file zuanfilmH3HD2Kdetailerv4.json, MIT, updated August 30.
Why it is good
Generating at full resolution directly costs memory you probably do not have and time you definitely do not want to spend. This turns a cheap draft into a deliverable, with the audio riding along untouched.
Where it breaks
No VRAM figure appears anywhere and the author calls it GPU-intensive; on an out-of-memory error, "reduce the working resolution and/or frame count before increasing the final output resolution." Sampler choice is not free either. He tested all of them, found…
448
Generate video with sound from a terminal, no Python
August 31, 2026 briefing recipe read the briefing →
Docs at stable-diffusion.cpp/docs/ltx2.md, quants at vantagewithai/LTX-2.5-GGUF.
Where it breaks
Grab the wrong decoder and nothing works: the default ltx-2.5-video-vae-bf16.safetensors is a diffusion decoder the project lists as not implemented, and you need the -conv- variant. The temporal upscaler and --auto-duration are also unimplemented, so pass…
449
Generate video with synchronized audio on a Mac, no NVIDIA card
August 30, 2026 briefing recipe read the briefing →
Go to the rapid-mlx video docs and pick ltx-2.5-mlx-q8, which the docs call "the recommended starting point" and the only one of the nine checkpoints that returns audio.
Why it is good
The only local path that gives you picture and sound together on consumer hardware, on a laptop you can close and carry. No rental, no queue, no uploading client footage to somebody's endpoint.
Where it breaks
Speed. The docs record CogVideoX-Fun q4 at roughly 338 seconds for one second of 672 by 384 video on an M3 Ultra. Even granting LTX-2.5's distilled path is faster, plan in minutes per clip. The 14B Wan checkpoints want 64 to 96 GB and are out of reach on most…
450
Put a real image model on your iPhone
August 30, 2026 briefing recipe read the briefing →
Start from apple/ml-stable-diffusion and grab a finished conversion at mahmudplx/coreml-absolutereality-6bit rather than converting anything yourself.
Why it is good
A gigabyte and a phone. No server, no credits, no network, nothing leaving the device.
Where it breaks
512 by 512 and nothing else. The text encoder stays at fp16 because CLIP's causal mask contains negative infinity values a k-means palettizer cannot cluster, so a third of the bundle refuses to compress. Apple's docs note the runtime can spike over 2 GB of…
451
Render a globally illuminated 3D scene on a Mac with 8GB
August 30, 2026 briefing recipe read the briefing →
Clone microsoft/renderformer, install microsoft/renderformer-v1.1-swin-large (483M parameters), and set scenes up with the Blender extension.
Why it is good
Preview-quality global illumination in under a second per frame on hardware that cannot run a serious path trace, MIT licensed so the output is unencumbered.
Where it breaks
Hard limits, stated plainly in the repo. Up to 8 light sources. Training covered 4,096 triangles, with 8,192 "usually still works" at inference, so detailed meshes need remeshing first through the included sceneprocessor/remesh.py. Camera distance must sit…
452
Text-to-MIDI Generation: How to Get a Score You Can Edit Instead of a Song You Cannot
August 30, 2026 deep dive read the piece →
The commercial music generators still have no score underneath, so any MIDI they hand you is transcribed back out of audio they already rendered; a small set of open models emit the notes directly, and that is what lets you finish the music yourself.
453
Render a multi-shot scene as one continuous take
August 29, 2026 briefing recipe read the briefing →
Files at joeygambino/MiniMax-H3-Multishot-Workflow, updated August 27 (the card says v2.7.0; the repo head is v2.7.2, committed the same evening).
Why it is good
Every other route to a multi-shot AI scene means generating clips separately and hiding the seams in an edit. This returns one master video and one master audio track, and continuous dialogue across a shot boundary is something the standard workflow simply…
Where it breaks
Texture ratchets up about 13% per join at 736 by 1280 with anti-drift on. The author measures it as slight under about four windows, 30 to 40 seconds, and visibly sharpening by seven, and calls a fix in progress. The full workflow wants a local Ollama running…
454
Put sound on silent footage
August 29, 2026 briefing recipe read the briefing →
Code at xiaomi-research/controlfoley, nodes at YJX-Research/comfyui-controlfoley-official.
Why it is good
It handles your text and your video disagreeing, which most video-to-audio models do not, and the authors report it outperforming CondFoleyGen, a specialist in-domain baseline, on the reference-audio task. Those are the authors' own evaluations. For anyone…
Where it breaks
CC BY-NC 4.0 on the weights. Non-commercial only, so no client work, no matter how good it sounds. Default duration is 8 seconds and it is built for short cues, not scores.
455
ControlFoley Adds Sound to Silent Video. The Mode Worth Your Time Is the One Nobody Demos.
August 29, 2026 deep dive read the piece →
ControlFoley's reference-audio mode, which copies a sound's character without copying its rhythm, is the one capability that justifies the pipeline, and a free browser demo has existed since April, so nobody needs to download anything to find out whether it works.
456
Generate video on a Mac in three passes instead of fifty
August 28, 2026 briefing recipe read the briefing →
Start at the FastMetal collection and pick the tier your machine can hold.
Why it is good
Video generation on a Mac has been an emulation story until now. This is a native path under Apache 2.0, which means output goes into client work without a license review.
Where it breaks
Apple Silicon and Python 3.11+ only, so no Windows and no Intel Macs. The 14B is locked to 448 by 832 and 77 frames with no audio track. Its weights are about 15GB, so a 16GB Mac will not hold it no matter how patient you are. These have been public since…
457
Run LTX-2.5 in ComfyUI using the community workflow pack
August 28, 2026 briefing recipe read the briefing →
The files live at RuneXX/LTX-2.3-Workflows, which despite the name now covers 2.5.
Why it is good
LTX-2.5 generates picture and sound in one pass, so footsteps in the shot are the footsteps in the audio. This is the shortest route from download to a working graph, and the models stay compatible with most LTX-2.3 adapters.
Where it breaks
The author has not shipped dedicated 2.5 workflows yet and tells you to reuse the 2.3 ones, so expect to fix a node or two. Both custom node packs must be on their latest versions for LTX-2 support, and a stale install fails in confusing ways. Commercial use…
458
Put an image model on a phone
August 28, 2026 briefing recipe read the briefing →
The conversion recipe is public at john-rocky/hf-to-litertlm, and the finished graphs are at litert-community.
Why it is good
No network, no account, no per-image cost, and output bit-identical to a desktop run. For anyone building a creative app, this is the difference between a server bill and a download.
Where it breaks
Slow. Bonsai takes about 64 seconds on an iPhone 17 Pro, klein five minutes on a Pixel 8a. Klein is fixed at 256 by 256 with one reference image, and changing either means re-exporting. Two settings silently corrupt output: compute must be FP32, since the…
459
FastMetal-QAD: Local Video Generation on a Mac, and Why the Biggest Model Is the Wrong Pick
August 28, 2026 deep dive read the piece →
FastMetal-QAD makes five-second video generation practical on a 16GB Mac, and the workflow that works inverts two instincts: take the 5B over the 14B, and treat prompt caching as the main speed lever rather than model size.
460
Add sound to silent footage
August 27, 2026 briefing recipe read the briefing →
ControlFoley ComfyUI nodes
Why it is good
It produces a WAV and a finished MP4 with the audio muxed in, so a silent generated clip becomes a deliverable in one pass. Defaults are tuned for mid-range GPUs at 25 steps.
Where it breaks
NVIDIA GPU only, no Mac support. First run downloads several dependency models and can take minutes. Weights are non-commercial, so this is a personal or spec-work tool until the license changes.
461
Fast video with matching audio
August 27, 2026 briefing recipe read the briefing →
MiniMax H3 Turbo workflow
Why it is good
Preview speed changes how you work. You stop composing one careful prompt and start trying twelve.
Where it breaks
The audio. Video and sound run on separate schedules, and a wrong audio sigma shift setting produces noise rather than sound. Start at 4 to 6, keep video sigma shift at 12, and if audio blows out, suspect the scheduler config before blaming the model. You can…
462
Score a video without a subscription
August 27, 2026 briefing recipe read the briefing →
ACE-Step on mlx-serve
Why it is good
Unlimited generation at zero marginal cost, nothing leaves your machine, and the license permits commercial release. For anyone scoring short pieces or building temp tracks, that removes an entire line item.
Where it breaks
The 4-bit build is slightly softer than 8-bit, most noticeable on dense vocal mixes. Take the 8-bit if you have the disk. No stems, no arrangement control, so this composes cues rather than replacing a DAW.
463
Shrink Your LoRA Adapters by 80 Percent With an SVD Rank Resize
August 27, 2026 deep dive read the piece →
Rank resizing turns oversized LoRA adapters into small ones with a documented, reproducible loss, and the fidelity numbers printed on model cards measure weight reconstruction rather than anything you can see.
Method: recipes are parsed from the “Creative workflows” section of every briefing, plus any analysis piece filed as a workflow. Links point at the original model card, repo, or workflow file — not at a mirror.