Creative AI Briefing: Monday, September 14, 2026
Edit the first frame of a clip in Photoshop, set it on fire, put a different face on the actor, retexture the grey clay model, and the change now propagates through the rest of the shot on its own, across camera cuts, with the original motion intact. That adapter went up Sunday afternoon. It is the second one a private individual has published for LTX 2.5 in three days, and it lands at the end of a week where Lightricks quietly moved seven of its own editing adapters onto the 2.5 base. Nothing here is a new model. The new thing is that footage you already shot has become the input.
Video
LTX Ripple propagates a single edited frame through a whole video, and it does it without a prompt. lolokukun/LTX-Ripple went up September 13 at 17:28 UTC: a 654,443,392-byte in-context LoRA for LTX-2.5, created by WepeNerd. The workflow is three steps. Feed it your video, replace frame zero with an edited version of frame zero, generate. The card calls the design First Frame All Frames, and the fifteen example clips shipped in the repo cover object replacement, outfit swaps, face swaps, archviz retexturing and style transfer applied to one actor and not the other. Weather changes and logo replacement appear on the card's list of intended uses but are not among the demonstrated clips. One example is worth the download on its own: because the edited first frame was sharper than the source, the entire video came back sharpened. The recommended strength is 1.35, higher than most adapters want, and the card gives the exact default prompt the workflow ships with, which begins "Use the reference video for motion, timing, camera movement, composition, and unchanged scene content." Raise strength when the edit is not carrying; lower it when things change that should not have. The honest gap is documentation: no rank, no training data, no step counts, no timings, nothing about which LTX 2.5 checkpoint it was trained against. You get weights, a workflow JSON and fifteen videos. Distribution falls under the LTX-2.x Community License.
Lightricks moved its editing library onto LTX 2.5 last week, and seven adapters did not make the trip. Between September 8 and September 10, eight new adapters appeared under the Lightricks account for the 2.5 base: Day-To-Night, Water-Simulation, Clean-Plate, Deblur, Colorization, Decompression, Ingredients and a Cinemagraph LoRA, joining the Pixel Spatial Upscaler that arrived August 11. Count the 2.3 shelf against the 2.5 shelf on the same account page and the gap is the story: Relight, In-Outpainting, Instant-Shave, Cross-Eyed, HDR, DubIt and the Foley video-to-audio LoRA all exist for 2.3 and have no 2.5 sibling as of this morning. If your pipeline depends on outpainting a frame, dubbing a performance or generating Foley, upgrading the base costs you the tool. Downloads tell you which ones people actually reach for: Ingredients leads the new batch at 490, Clean-Plate 124, Water-Simulation 113.
NVIDIA's two-stage acceleration pipeline now runs on a graphics card from 2018. coolthor/H3-Super-Acceleration-Turing, September 14 at 06:58 UTC, puts NVIDIA's Sol Engine H3 Super Acceleration configuration on a modified RTX 2080 Ti with 22 GB, compute capability 7.5. MiniMax H3 drafts at 672x384 in four steps, LTX 2.5 spends three refine steps at sigmas 0.78, 0.643, 0.546 and 0, and out comes 1344x768, 121 frames, 24 fps, 5.04 seconds, with audio. Warm, with models resident, that is 165 seconds, median of three runs with a 1.1 second spread. Cold off disk, 178. First run after a ComfyUI restart, 190, because Triton throws away its kernel cache and SageAttention recompiles six kernels. NVIDIA runs the same configuration on one DGX Spark in 56 seconds. The repo also does something unusual and correct: it refuses NVIDIA's own headline. The 6.852 seconds on the GB200 page belongs to a 896x512 draft line, and NVIDIA's benchmark note confirms that figure sums two separately measured stages with model loading and warmup excluded, so it is not a wall clock. Read the version pins before you try this. comfy-kitchen 0.2.31 will not import under torch 2.6, and torch 2.6 is what sageattention 1.0.6 and triton 3.2.0 are built against. Two catches, and they are large. The H3 community license excludes the EU, the UK, South Korea and the United States, which is why the repo is gated. And the gate is what hides TURING.md, the 13 KB working document the author describes as the actual point of the release, with six mandatory departures and five dead ends, three of which fail silently. The author states plainly that the documentation, workflow JSON, startup adapter and conversion script are his own and carry no restriction. You still cannot read them without clearing a licence gate on weights you may not be licensed to use.
Also new, unverified beyond its own card: JOKER141/BUNNY_H3_Conditioning_Bridge (September 13, 23:03 UTC), a conditioning adapter for MiniMax H3 aimed at action logic and motion continuity in combat shots, and zshyang1106/CoaG-Wan2.2-Fun-A14B-Control-LoRA (September 13, 20:24 UTC, Apache 2.0), a layout and camera control adapter for Wan 2.2 Fun. Neither publishes a measurement yet.
Image
The Krea 2 fine-tune wave is the quietest big thing on the Hub right now. Nobody announced anything, but in the last twenty-four hours alone at least eight new LoRAs appeared against krea/Krea-2-Raw and krea/Krea-2-Turbo from eight different accounts, alongside a Krea-Edit LoRA (September 14, 08:11 UTC). The base is not new, September 14 is not its birthday: Krea-2-Raw has a createdAt of June 18 and now carries 720 likes and roughly 75,000 downloads. What changed is that the training path got easy, and the Spaces list on that model page now includes several community LoRA trainers. The model is gated behind the Krea 2 Community License with a click-through agreement and an acceptable use policy, so read before you build a business on it. Replicate's June 23 write-up remains the most useful plain-language guide to what the model is actually good at.
Quantization crews moved to whatever was left. JoaoZaokk/Krea-2-Turbo-W4A4-ConvRot (September 14, 05:58 UTC) and Qwen-Image-2512-W4A8-ConvRot (September 13, 10:24 UTC, Apache 2.0) continue the run covered here yesterday. On Apple Silicon, dushandz/FLUX.2-klein-4B-CoreAI landed at 08:41 UTC this morning, and toxicdog published six MLX and mflux conversions inside four minutes on September 13 covering FLUX.2 Klein at 4 and 8 bit, Krea 2 Turbo and Z-Image Turbo. None of these publish a quality comparison. Treat them as convenience, not as measured work.
ERNIE Image Turbo now runs through Vulkan, which means it runs on almost anything. akashimio/ERNIE-Image-Turbo-ncnn and Coderdw/ernie-image-ncnn-vulkan, both September 13, convert Baidu's Apache 2.0 model to ncnn graphs. ncnn needs no PyTorch and no CUDA, so the target is Intel and AMD integrated graphics, older discrete cards and phones. Same idea, different medium: fszontagh/LLaDA-Image-Turbo-GGUF (September 13, 16:54 UTC, Apache 2.0) puts a diffusion language image model into stable-diffusion.cpp.
Audio and music
YuE2 is being ported everywhere at once, and the Mac builds arrived overnight. Five days after m-a-p/YuE2-3B shipped, the conversion count is remarkable: smcleod/YuE2-3B-int8-ar for Apple Silicon MPS (September 14, 01:32 UTC), vanch007/mlx-Yue2-3B and npario and ahmadw MLX builds, and the runaway leader, audio-cpp/Yue2-3B-GGUF at 37 likes and 14,077 downloads. All of it inherits CC BY-NC 4.0 from the base, so this is your own music, not a client's. One note on the MLX builds: vanch007/mlx-Yue2-3B declares Apache 2.0 in its card metadata while the upstream weights are CC BY-NC 4.0. The upstream licence is the one that governs.
Local TTS moved onto the Neural Engine. FluidInference/chatterbox-nano-coreml (September 13, 19:23 UTC, MIT) and FluidInference/moss-tts-nano-coreml (20:44 UTC, Apache 2.0) are Core ML conversions targeting iOS and macOS, the MOSS build advertising streaming voice cloning across twenty languages. If you narrate anything on a Mac, these are the cheapest thing to try this week.
Open and local
The most interesting repository here has no stars at all, which says more about how fresh it is than how good.
- Lightricks/ComfyUI-LTXVideo (4.1k stars): the node pack that supplies the
LTX IC-LoRA Loader Model Onlynode. Why now: every adapter in this briefing loads through it. - Lightricks/LTX-2 (9.4k stars): the reference implementation and the home of the LTX-2.x Community License file every derivative points at. Why now: two community adapters shipped against it in three days.
- AkaneTendo25/musubi-tuner: the
ltx-2-devbranch that can train an LTX 2.5 IC-LoRA. Why now: it is the trainer behind the first fully documented community IC-LoRA. - kohya-ss/musubi-tuner (2k stars): the upstream trainer that fork is based on. Why now: the fork's recipe transfers back.
- mingshi2333/seedvr2-ncnn-vulkan (0 stars): a native C++20 and Vulkan runtime for ByteDance's SeedVR2 restoration model, no PyTorch anywhere in the inference path. Why now: akashimio/SeedVR2-3B-ncnn published the converted packages on September 13, and the card is unusually honest about what they are worth.
- madebyollin/taehv: the tiny decoders NVIDIA's two-stage pipeline uses. Why now: TAEH3 decode at 0.028 s against the official H3 VAE's 3.427 s is most of where NVIDIA's Stage 1 win comes from.
A caution on that SeedVR2 conversion, because the card earns it. The packages are roughly 20.44 GB for image and 21.10 GB for video, capped at 512 pixels on the long edge for stills and 17 frames at 128 pixels for video, and the author reports that on three retained development videos both his build and the official FP32 reference scored below plain bicubic on the fixed quality metrics. His words: numerical parity does not imply improved restoration quality. That is a conversion project reporting its own negative result, and it is worth more than most launch posts.
Creative workflows
1. Change one frame and let the change travel through the shot. From the LTX Ripple card, using the shipped Workflow_example/LTX_Ripple_FFAF_Edit_v11.json.
The steps. Put LTX25_Ripple_v11.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 in beside the untouched video. Set LoRA strength to 1.35. Leave the default prompt alone on the first attempt, and add one sentence describing the change only if the edit is not carrying. How it works. An in-context LoRA learns a relationship between two videos rather than a look. Here the pair is your source clip plus its edited first frame on one side, and the edited clip on the other, so the adapter reads your image edit as the instruction and the source footage as the specification for motion, timing and everything you did not touch. Why it is good. The instruction is a picture, not a sentence, so you can be exact about a colour, a logo or a garment in a way no prompt reaches, and you do it in a tool you already know. Where it breaks. The card names four failure modes: an unclear or incoherent edited frame, very large changes, heavy occlusion and fast motion, and anything you ask for that does not appear anywhere in the source video. Strength is a two-sided dial, and too much of it changes things you wanted left alone. Nothing is published about speed or VRAM, so budget your first run as an experiment.
2. Paint flat colour blobs, get smoke and fire that follows them. From AInVFX's September 13 write-up and the ainvfx-fluid card, with a 45-minute video walkthrough.
The steps. 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 last frame at index 120, black everywhere between. Any extra keyframe must land on a multiple of 8 plus 1, so frame 1, 9, 17 and so on up to 113. Paint flat colours, one blob per volume, light grey for smoke, white for steam, orange with a yellow core for fire. Prompt ainvfxfluid, smoke plume. Generate on the distilled transformer at 8 steps, CFG 1, euler_ancestral. How it works. The LTX VAE groups frames into blocks of eight, so a keyframe sitting on that grid gets read cleanly and one placed off it gets smeared across a block. The control is encoded at half resolution, which is why the multiple-of-64 rule bites. Why it is good. Roughly 50 seconds for a five-second 512x512 element on a laptop 4090, and the input is the drawing a supervisor already scribbled on the frame. Shading carries: a darker grey along one edge of your painted plume comes back as shading rather than a flat white cloud. Where it breaks. Smoke, steam and fire only; water and ink are hit and miss. It is not a plate matcher, so generate over black and composite, or run two generations, one on the plate for perspective and one on black to patch. Off the grid, quality falls fast. And the LTX-2.x Community License means a company with 10M USD or more in annual revenue needs a paid agreement with Lightricks before shipping anything made with it.
3. Check whether an upscaler is actually helping before you trust it. Carried forward as method, prompted by the SeedVR2 ncnn card's own reporting.
The steps. 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 percent. How it works. Restoration models are usually demonstrated on degradation they were trained to reverse. Your footage has different degradation, so the demo tells you very little. Why it is good. It costs twenty minutes and it catches the case the SeedVR2 conversion author caught in his own work, where a correct, numerically verified implementation still lost to bicubic on his test clips. Where it breaks. Fixed-target metrics disagree with eyes often enough that you should not let the number alone decide. Use it to reject the clearly worse option, then judge the rest by looking.
Worth testing
- LTX Ripple, free, ungated adapter, gated base. Watch the fifteen example videos on the card before downloading 654 MB. Tradeoff: the most capable-looking video editing adapter of the week ships with no training documentation at all, so you cannot predict where it will fail except by running it.
- ainvfx-fluid, free, 654 MB, with a tutorial and every training number published. Tradeoff: smoke, steam and fire only, and the revenue threshold in the base licence is real.
- Krea 2 Turbo through one of the community LoRA trainer Spaces linked on the model page. Tradeoff: gated, click-through licence, and an acceptable use policy that governs what you produce.
- audio-cpp/Yue2-3B-GGUF for a full song locally with no Python stack. Tradeoff: CC BY-NC 4.0 travels from the base, so nothing you make is for a paying client.
- FluidInference/chatterbox-nano-coreml, MIT, on-device narration for macOS and iOS. Tradeoff: no benchmark, no comparison against the PyTorch original, and the card is a day old.
What actually matters from today's signal
The unit of release changed. For most of this year the thing worth waiting for was a checkpoint, and the question a creator asked was whether the new model was better than the old one. This week the thing worth waiting for was an adapter, and the question became what you can now do to material you already own. Ripple, ainvfx-fluid and Lightricks' own eight all take footage as the input and return the same footage altered. That is a different relationship to a generative model than prompting one, and it fits a working day in a way prompting never has, because the thing you are changing is a shot that already has your composition, your talent and your camera move in it.
The second thing is cost. ainvfx-fluid published its whole bill: 7 hours 47 minutes on one rented RTX PRO 6000, 52 clips off Pexels, rank 32, three resolution stages, peak 24 GB. That is a weekend and a two-figure GPU rental for a capability that did not exist. The recipe is model-agnostic in the way that matters: swap the dataset, keep the control video rules. If the effect you need is not smoke, nobody is coming to build it, and now nobody has to.
The counter-signal is documentation, and it cuts in a direction that should make people uncomfortable. The most impressive adapter this week publishes almost nothing about how it was made, and the most thoroughly documented one is a teaching demo from a training course. Those are not coincidences, they are what each thing is for. The practical consequence is that an undocumented adapter is an experiment, not a dependency. And the Turing repository is the sharpest version of the problem: a set of files the author explicitly gives away, sealed behind a licence gate on weights that exclude most of the people who would read English coverage of it. The work is free. Getting to it is not.
Source access notes: An adversarial fact-check pass ran against this draft before publication. It caught two things, both corrected above: the Ripple paragraph originally credited the fifteen example clips with demonstrating weather changes and logo replacement, which appear only on the card's separate list of intended uses, and a Krea-2-Raw download figure that had already drifted. Everything else it tested held, including the adapter gap counted exhaustively against the full Lightricks listing, the exact Ripple byte count, every ainvfx-fluid training number against both the model card and the blog post, and the NVIDIA benchmark-basis correction against NVIDIA's own page. Every ship date above comes from the Hugging Face API createdAt field with ?cb= cache busting, never a listing "Updated" timestamp. Star counts are cache-busted img.shields.io JSON read September 14. TURING.md in coolthor/H3-Super-Acceleration-Turing could not be retrieved: the repo is gated, and raw fetches, the blob page and a browser session all returned an unauthenticated 401, so every claim about that repo here comes from its README, which the resolve endpoint does serve, plus NVIDIA's own Sol Engine page, which independently confirms the benchmark-basis point. The 2.3-versus-2.5 adapter gap was counted from the Lightricks author listing read September 14 and is accurate as of that read. huggingface.co/api/daily_papers?date=2026-09-13 returned an empty array. No lab published a creative-model release inside the window: fal, Replicate, ElevenLabs and Hugging Face's own blog were checked and their most recent relevant items are dated where cited. Adobe, OpenAI, DeepMind, Runway, Black Forest Labs, Stability, Luma, Midjourney, Kling and Suno produced nothing inside the 24-hour window, so no claim here rests on them. blog.comfy.org and civitai.com were not reachable in usable form. WepeNerd's author API endpoint returned an empty body, so nothing is claimed about that account beyond the credit line on the Ripple card.