Creative AI Briefing September 5, 2026 morning edition 18 min read

Nobody shipped a new generator this week, so the work went into output formats instead, and teaching a 33-billion-parameter video model an entirely new frame layout now costs twelve dollars and fifty cents.

VDN-H3 MiniMax H3 FLUX.2 Klein Foundry LLaDA-Image-Turbo Stable Audio 3 Small SFX video-gen image-gen audio-gen local-creative-ai lora-finetuning comfyui open-weights licensing-provenance

Creative AI Briefing: Saturday, September 5, 2026

Write a sentence about a temple courtyard at dawn, get back a five-second clip, run two ffmpeg commands, and put on a Quest to stand inside it. That path opened this week, twice, from two people who had never met, using the same rented trainer and about $12.50 each. Nobody released a new generator in this window. What shipped instead was output formats: adapters that teach an existing video model a headset's frame layout, a hybrid attention branch that makes the same model render faster than it plays, and a build of a 4-billion-parameter image model that runs inside a browser tab. The generator stopped being the interesting layer. The container it fills is.

New models

VDN-H3 is the release worth your morning, and it is the closest thing to a real model this week. Video DeltaNet weights landed on Hugging Face September 2. The idea is a hybrid attention architecture: bolt a frame-wise linear attention branch onto MiniMax H3, keep a softmax branch alongside it to hold visual quality, and let the cheap branch carry most of the sequence. Linear attention costs far less than softmax attention as your clip gets longer, which is exactly where video generation hurts.

The numbers are on the 768p, 14.4-second workload, and they are the most useful published this week because they include the honest single-GPU row. On one H200, dense MiniMax H3 at 8 denoising steps takes 4.4 minutes; VDN-H3 in FP8 takes 90.5 seconds. On one B200, 2.23 minutes becomes 51 seconds. The headline, eight B200s producing a 14.4-second clip in 11.23 seconds, is real but it is a datacenter claim, and the authors say plainly that it excludes model loading, warm-up, VAE decoding and MP4 encoding.

What makes this a creator story rather than an infrastructure story is the packaging. The checkpoint adds a separate linear branch and two small LoRA adapters that merge into the backbone at inference without touching the base weights, so nothing about your existing MiniMax H3 install is destroyed. The full download is about 82 GB, of which 72 GB is just the released base you may already have; the actual VDN parts are 4.3 GB for the 50-step model and 5.1 GB for the 8-step. Training code ships alongside inference code, all four stages, with the DMD stage documented as data-free.

The catch is the same one that has followed every MiniMax H3 derivative through this briefing. The code is Apache 2.0. The weights are not. They carry the MiniMax H3 Community License Agreement, whose applicable territory excludes the European Union, the United Kingdom, the Republic of Korea and the United States. Most people reading this in English are outside the territory the weights are licensed for. There is a free hosted Space if you want to hear that sentence and still see the output (weights, code, hosted demo).

Image

FLUX.2 Klein 4B, quantized for a browser tab, appeared September 5 and it is the most quietly ambitious thing in this briefing. The repo holds an int8 transformer as nine raw binary shards described by a manifest.json, an int4 group-wise set for phones, a 96 MB rank-64 LoRA that cuts generation to two steps, a distilled 0.6B text encoder in ONNX, and the FLUX.2 small decoder as fixed-shape graphs. The stated design is a hand-written WGSL transformer on WebGPU with onnxruntime-web handling the encoder and VAE, cached in the browser's origin private file system so a second visit starts in seconds. First-run download is about 5 GB on desktop and 3.2 GB on a phone. End-to-end parity against the f32 reference is reported at rel-L2 of roughly 1e-2.

Read that as a weights drop, not a product, because the two things that would make it usable did not resolve this run: the Hugging Face Space the card names as the demo returned nothing and is not listed among the repo's linked Spaces, and the flux-klein.js npm engine's GitHub repo returned "repo not found." The files are real and dated. The way in is not verifiable yet (weights).

Closer to usable today: LLaDA-Image-Turbo landed in ComfyUI on September 4 as an INT8 transformer plus a Q4_K_M GGUF text encoder. Four steps, guidance 1.0. The part worth knowing is that editing here is not img2img with a denoise slider. The same checkpoint runs generation_mode="editing", feeding your source image through LLaDA's own conditioning path, so you write an instruction rather than dial a strength value. Edit dimensions must be divisible by 32 (weights, nodes).

Video

Two people taught MiniMax H3 to shoot for a headset, three days apart, without coordinating.

The VR180 stereoscopic LoRA shipped September 3. It makes the model produce one frame holding a left-eye view on the left and a right-eye view on the right, each a 180-degree hemisphere, 24 fps, with H3's native audio. The card reports 3 to 7 pixels of horizontal disparity per 768-pixel eye on near objects and roughly zero in the far field, which is what real VR180 footage looks like. Training data was 50 clips of about 5.3 seconds from 27 real VR180 YouTube videos, filtered through Google's Stereo4D index of validated stereo footage.

The 360 equirectangular LoRA shipped September 4 and goes after the other VR grammar. Instead of a stereo window you face, it produces a full spherical projection wrapping around you, horizon at the vertical middle. Its card frames the difference exactly right: places, not windows. It trained on 124 clips from 47 real 360 videos across three lanes, realism and psychedelic and cavern.

One caveat that belongs on every number from that second repo. Its card carries a note from the author saying it was written by GLM 5.3 Flash and "may not be entirely accurate or represent the authors intent." The weights, the file list, the training scripts and the sample clips are all real and checkable. The prose describing them was generated, so treat its figures as the author's machine summary rather than as hand-written documentation. The VR180 card carries no such note.

Both were trained on fal's hosted MiniMax H3 trainer, both rank 32 on the qkv_proj and out_proj attention projections, both 2500 steps at AdamW 2e-4, both about 3.9 hours on one H200, both US$12.50. Both use the same trick: fit a 2:1 frame into H3's 896x384 21:9 bucket so the platform's cover-and-crop becomes a no-op, then un-squeeze back to 2:1 at packaging time. Both ship their full pipeline in the repo and neither redistributes a single training frame.

Both are honest about failure too, which is rare enough to name. The 360 card reports that the wrap seam almost closes and not quite, because the trainer has no circular-convolution option, and that H3's dolly-and-pan prior bleeds through and invents camera movement that drags distortion across the projection. The VR180 author reports dark porthole vignettes at the outer edge of each eye and a first-person-outdoor-walk bias from one source channel supplying 20 of 50 clips. Both say the same closing line: evaluated on a handful of prompts and one seed, put it in a headset before trusting it (VR180, 360).

Audio and music

Foundry is a 16-pad drum machine for macOS where every pad renders from a text prompt, and the model ships inside the app. It went up September 2 as a signed, notarized 3,391,504,096-byte disk image with a Developer ID and a stapled Apple notarization. No account, no API key, and the card states there is no network access after download.

The inference contract is published in full: stable-audio-3-small-sfx unmodified, T5Gemma as the text encoder, fp32 on MPS with CPU fallback, 8 steps, CFG 1.0, two seconds clamped between one and three, 44.1 kHz. It retries up to three times against a spectral-flatness floor of 0.55, which is a sensible way of saying "that came back as noise, try again." Around the model sits an actual instrument: a morph pad blending four corner prompts bilinearly, per-pad trim and pitch envelope and transient and choke groups, alternate tunings including any .scl file you hand it, CoreMIDI in on notes 36 through 51, and export to 24-bit WAV or an Ableton .adg Drum Rack with a DrumCell per pad.

Note what the bundling does. The upstream Stability checkpoint is gated on Hugging Face behind a form. Foundry redistributes it unmodified inside the app, under the Stability AI Community License, with both that license and the Gemma terms shipped in the bundle. Installing means accepting both, including Gemma's prohibited use policy. Foundry's own source is MIT and covers none of that (download and card, upstream model).

Also September 5, a Q8_0 GGUF of HeartMuLa, the Apache 2.0 open music foundation family, packaged as one self-contained 7.66 GB file for Apple Silicon. Two details the filename hides. It is the happy-new-year themed fine-tune of the 3B model, not the plain checkpoint, so expect its bias. And it exists to feed one specific app: MindMuse, a local macOS music and speech studio whose model manager fetches this repo automatically. The upstream generates complete songs from style tags and structured lyrics with vocal conditioning in Chinese, English, Japanese, Korean and Spanish, and the quantization keeps the upstream Apache 2.0 (weights, upstream, paper).

Open and local

The local story is that VDN-H3 became loadable in ComfyUI one day after the weights appeared. t8star/Vdn-Minimax-H3-Comfy went up September 3 as a drop-in bundle already using ComfyUI's folder names, pinned to a specific OpenVDN revision. It is the first HF repo in weeks to gate itself on the territory question directly: the access form asks your country, asks you to confirm you accept the MiniMax H3 Community License, and asks you to confirm you are not requesting from an excluded territory.

Star totals below are cache-busted shields.io reads taken September 5.

OpenVDN/vdn-minimax-h3: the hybrid-attention speedup, with inference and all four training stages open, so you can retrain the branch rather than only run it (261) (repo)
T8mars/comfyui-minimax-h3-audio-T8: the node pack that now carries OpenVDN workflows in examples/workflows/10-speed alongside its WASD and lip-sync routes (945) (repo)
RealRebelAI/LLaDa-Image_ComfyUI: nodes for diffusion-language image generation and instruction editing from one checkpoint (repo)
google/spatial-media: the unglamorous dependency both VR LoRAs need, because a 2:1 mp4 without spherical metadata is just a squashed rectangle (repo)
inclusionAI/LLaDA-Image: upstream source and documentation for the editing path the ComfyUI nodes wrap (repo)

One warning about that bundle. Its sources section credits the OpenVDN implementation to NVlabs/OpenVDN under Apache-2.0. That repository does not exist; a cache-busted shields.io read returns "repo not found." The real code is at OpenVDN/vdn-minimax-h3, whose own README carries the Apache-2.0 grant for the code and states outright that the weights are not covered by it. If you are checking provenance before you run 82 GB of someone's checkpoint, check it against the upstream repo, not the bundle's link.

Creative workflows

1. Generate a 360 video and get it into a headset tonight. LoRA at shamanic/minimax-h3-equi360-lora. Weights file: h3-equi360-lora-step2500.safetensors. Packaging tool: google/spatial-media.

The steps. POST to https://queue.fal.run/minimax/h3/text-to-video/lora with loras pointing at the resolve URL for that safetensors file at scale 1.0, aspect_ratio: "21:9", resolution: "768P", duration: 5, and prompt_expansion_mode: "disabled". Start the prompt with the trigger equirect360, follow it with the layout sentence from the card ("Equirectangular 360-degree panoramic video: the frame is a full spherical projection..."), then describe your scene including its sound. Un-squeeze and tag the result:

ffmpeg -i out.mp4 -vf "scale=1536:768:flags=lanczos,setsar=1" -c:v libx264 -crf 16 -pix_fmt yuv420p -c:a copy tmp.mp4
python spatialmedia -i --v2 --stereo=none -p equirectangular tmp.mp4 out_360.mp4

Sideload out_360.mp4 and it plays as mono 360 in Quest, DeoVR or Skybox. Four finished, already-tagged samples sit in the repo's samples/ if you want to check your headset pipeline before spending a generation.

How it works. The adapter never learns geometry. It learns a frame layout. Training captions all carried the same prepended layout sentence, so at inference that sentence plus the trigger token puts the model in the projection it was taught. The 21:9 bucket is not aesthetic; it is the closest shape fal's trainer offers to 2:1, and everything else is a squeeze on the way in and a stretch on the way out.

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/, Node.js, needing only yt-dlp, ffmpeg, zip and a fal key.

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 camera, locked tripod" in the prompt helps. Stay on the minimax/h3 route, because the 8-step h3-max turbo route was not evaluated, and fal's 2K and 4K outputs are SeedVR upscales of the 768p pass rather than native. And the base weights carry the MiniMax territory restriction that excludes the US, UK, EU and South Korea.

2. Run the 8-step VDN speedup locally in ComfyUI. Bundle at t8star/Vdn-Minimax-H3-Comfy. Nodes at T8mars/comfyui-minimax-h3-audio-T8. Workflows: examples/workflows/10-speed.

The steps. Clone the node pack into ComfyUI/custom_nodes as minimax-h3-audio-T8. Request access on the bundle repo, answering the country and territory questions honestly. Then hf auth login and hf download t8star/Vdn-Minimax-H3-Comfy --local-dir ComfyUI/models, which lands everything in ComfyUI's own folder names with nothing to rearrange. Restart, open a workflow from 10-speed, and pick your base in UNETLoader: either the full 34.04 GB minimax_h3_fl2va_int8_convrot.safetensors or the 20.97 GB pruned variant.

How it works. The Composer node reads the base checkpoint's structure and picks the matching adapter itself. That matters because the published Turbo adapter's 51 AdaLN targets were trained for a 2688-column input, and the pruned base has eight columns; T8 supplies a separately identified curve-projected adapter that represents those 51 modules as eight-column factors plus 51 bias residuals, 310 applied patches in total. Selection is bound to a content hash of the raw adaln_t_table, not to a filename, so an unknown pruned checkpoint fails before sampling instead of silently running the wrong adapter. Renaming a file does not get you past it.

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 passed strict H.264/AAC decoding with finite non-clipped PCM.

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 of why the memory is tight. The maintainer's phrasing is the right one: this proves compatibility, not 16 GB safety. Do not stack another Turbo or EMA LoRA, SLA, VSA, Sol-Attn, BlockCache or a second attention patch on the OpenVDN model. DMD is fixed at 8 Euler steps with video and audio shifts of 12 and 3, and those are requirements. The bundle does not claim its INT8/ConvRot base is byte-identical to OpenVDN's declared BF16 upstream. And its own source list points at a GitHub repo that does not exist.

3. Teach a video model your own frame format for about $12.50. Reference pipelines: VR180 (training/pipeline_config.yaml, training/fal_trainer_request.json) and 360 (scripts/train.mjs, training/trainer-config.json).

The steps. Collect 50 to 125 real clips of about five seconds each in the layout you want the model to learn. Scale every frame exactly into fal's 896x384x124 bucket at 24 fps constant frame rate so the platform's cover-and-crop does nothing, strip letterboxing, keep the 32 kHz stereo audio. Caption each clip with a vision model that can hear the audio, one paragraph, then prepend the identical layout sentence to every caption and set a trigger phrase. Zip, submit to minimax/h3/t2v/trainer at rank 32 on qkv_proj and out_proj, AdamW 2e-4 with linear decay, batch 1, 2500 steps. Evaluate same-seed A/B at LoRA scale 0 against scale 1.

How it works. You are not teaching the model new content. You are teaching it a spatial convention, and the constant layout sentence is what binds that convention to a token you can type later. This is why 50 clips is enough where a style LoRA would need more: the target is one repeated geometric arrangement, not a distribution of subjects.

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 converging on the same recipe is stronger evidence than either one alone.

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 skip the pilot. Neither repo redistributes training frames, so your data problem is entirely yours, and YouTube's terms are yours to read. Both authors evaluated on a handful of prompts and one seed each.

Worth testing

The VDN-H3 Space, free, running on a zero-A10G, for 8-step video with a soundtrack. Tradeoff: shared ZeroGPU hardware tells you nothing about the H200 and B200 numbers that are the entire point, and the license question follows you into the browser.
The 360 LoRA's four tagged sample clips on a headset before you generate anything. Tradeoff: they are the author's best four, chosen after evaluation, so treat them as a ceiling.
Foundry, if you have an Apple silicon Mac and a drum machine shaped hole. Tradeoff: 3.4 GB, two-second pads, and a bundled license stack you are accepting by installing.
LLaDA-Image-Turbo's editing mode for instruction edits without a strength slider, on upstream weights that are Apache 2.0. Tradeoff: edit dimensions must divide by 32, and the quantized repo never restates that license, so you have to go up the chain to learn you are covered.
VDN-H3's training code, not its weights, if you have hardware and curiosity. Tradeoff: PyTorch 2.13 and prerelease FlashAttention 4 are stated requirements, and the DMD stage initializes from a specific community Turbo LoRA you must fetch separately.

What actually matters from today's signal

The cost of teaching a frontier-scale video model a genuinely new capability is now $12.50 and four hours, and two strangers proved it independently in the same 48 hours. That number should reset your sense of what is worth attempting. A LoRA that produces headset-native footage is not a research project any more; it is a Saturday and a rented H200. Whatever spatial or format convention your work needs, and nobody has bothered to train, is probably now within reach of one person with 50 reference clips.

The reason it works is the part most coverage will skip. Neither adapter taught the model anything about geometry. Each taught it a frame layout, bound to a trigger token by a sentence repeated across every caption. That is the whole mechanism, and it generalizes: if you can describe the arrangement in one consistent sentence and show it 50 times, the adapter learns to reach for it. The corollary is that everything downstream is plumbing, which is why both workflows end in ffmpeg and a Google metadata tool from a decade ago rather than in anything clever.

The counter-signal is that the fastest, cheapest and most interesting work in this entire briefing sits on top of weights whose license excludes the four places most of its readers live. VDN-H3, both VR LoRAs and the ComfyUI bundle all inherit the MiniMax H3 Community License and its excluded territories. The bundle at least asks you the question at the download gate, which is more than the rest do. Set against that, the two items here that no territory clause touches are the ones nobody will write about: an Apache 2.0 music model quietly converted to GGUF, and a drum machine that fits in a dmg. Foundry still hands you a license stack, Stability's community terms plus Gemma's prohibited-use policy, and you accept both by installing. That is a real encumbrance. It is just one you can read in an afternoon and then get on with your day, which is not true of a clause that says your country is out.

Source access notes: Crawled September 5, 2026, covering roughly September 2 to September 5. Every Hugging Face date here is the API createdAt field read with cache busting, never a listing "Updated" timestamp: VDN-H3 weights September 2, its ComfyUI bundle September 3, the VR180 LoRA September 3, the 360 LoRA September 4, LLaDA-Image-Turbo for ComfyUI September 4, Foundry September 2, the FLUX.2 Klein browser weights September 5, the HeartMuLa GGUF September 5, and stable-audio-3-small-sfx itself May 17. Star counts are cache-busted shields.io reads. GitHub file contents were read via raw.githubusercontent.com with cache busting.

Two claims are deliberately flagged rather than asserted. The FLUX.2 Klein browser demo Space named on its model card returned an empty response and is not listed among that repo's linked Spaces, and the radames/flux-klein.js npm engine repo returns "repo not found" from shields.io, so that item is reported as a weights drop with no verified way in. And the t8star bundle's source list credits NVlabs/OpenVDN, which also returns "repo not found"; the real code is OpenVDN/vdn-minimax-h3.

An adversarial fact-check pass ran against this draft before publication and caught five things, all now fixed. The largest: the 360 LoRA's model card carries an author's note saying it was written by GLM 5.3 Flash and may not be accurate, which this draft had been quoting as hand-written documentation; that caveat is now stated in the Video section, and the "places, not windows" line is attributed to the card rather than to a person. It also found that the HeartMuLa GGUF does have a full card, which named it as the happy-new-year themed fine-tune tied to a specific macOS app, rewriting that item entirely; that LLaDA-Image-Turbo's upstream is explicitly Apache 2.0 in its API metadata, so the draft was manufacturing license uncertainty; that T2VA and I2VA generate audio as well as video and should not be glossed as text-to-video and image-to-video; and that calling Foundry "unambiguously" usable undercut the licensing paragraph two sections above it. Everything else held, including all six VDN-H3 timing figures, the checkpoint layout, both star counts, Foundry's exact byte count and inference contract, every FLUX.2 Klein size, the t8star patch and VRAM figures, both VR LoRAs' independently matching hyperparameters, every ffmpeg and fal API body against its card, and all three flagged non-existent repositories.

Blocked or stale this run: github.com/trending served a page of 2018-era repositories and was discarded entirely rather than sanity-checked item by item. The Hugging Face models API served a cached page nine weeks stale on the first uncached text-to-image query and returned current results only after cache busting, which is a reminder that the ?cb= habit applies to the listing endpoints and not just the per-model ones. Vendor blogs were checked and were quiet inside the window: OpenAI's newsroom last touched creative tooling before September 3, blog.adobe.com's most recent posts are a September 3 Premiere Emmy award and a September 2 Slack integration rather than model news (note for future runs: news.adobe.com is a separate feed whose September 3 item is a CEO succession announcement), fal's blog last posted August 27, Replicate's August 4, ElevenLabs' most recent entry is a September 2 executive hire, and Black Forest Labs' most recent post is August 20. The Writing section is omitted because nothing surfaced.