Creative AI Briefing: Saturday, August 29, 2026
A thirty-second continuous take, two characters who each keep their own voice all the way through, no visible cut at any shot boundary and one unbroken audio track, rendered on a single consumer graphics card. That workflow updated on Thursday. On the same day, the company that maintains the software it runs in became an official reseller of the license that lets you bill a client for the result.
Neither of those is a new model, and that is the shape of the week. Almost nothing shipped that can do what no model could do before, yet two real barriers came down: a 720p video model went from fifty denoising passes to four under Apache 2.0, and an open-weights video model got a commercial license you can buy. Step count and legal permission are doing the work new capability used to do. The third barrier, running a music generator on a processor with no graphics card, fell four weeks ago and nobody noticed.
New models
SANA-Video 2.0 5B 720p, four-step preview landed August 27 from NVIDIA's Efficient-Large-Model group, and it is the release worth your afternoon. About five seconds of 736 by 1280 video, 81 frames at 16 frames per second, from a text prompt in four passes instead of the base model's fifty. Distillation is the trick, and it means what it sounds like: a student network trained to reach a finished frame in a handful of jumps where its teacher needed the slow walk. 4.47 billion parameters, BF16, using Google's small Gemma 2 2B as text encoder. Apache 2.0, which puts no license restriction between the model and paid client work, and lets you fine-tune from it commercially.
The card is unusually honest. It publishes the exact four sigma values the student was trained on, and states plainly that motion, anatomy, text rendering and object permanence break down in crowded or fast scenes. Text-to-video only, so no first-frame conditioning; use the 50-step base release for that. Until the upstream pull request merges you install from a preview branch, and as of this morning PR 471 is still open (model card, free demo).
Video
ComfyUI added Gemini Omni 1.1 Flash as a partner node on August 28. One node covers text-to-video, image-to-video, reference-to-video, scene extension, and instruction-based editing of a clip you already have, up to 4K with a generated audio track on every result. Stacked edits keep context, so "now change the lighting" works without re-describing the shot. It bills through an API rather than running on your card, and the post names no price, so treat it as a hosted service living inside a local graph (Comfy).
The local route to speed is a LoRA, a small trained file that bolts a behavior onto a bigger model without replacing it. drbaph/MiniMax-H3-Turbo-Lora-ComfyUI added v1.1 four-step adapters at rank 28 and 64 on August 27; the eight-step 768p variants, for when four steps look too rough, landed a week earlier on August 20. 197,450 downloads, Apache 2.0, ships its own example graph (repo, demo).
Second-day work worth catching up on: FastVideo's FastH3 Preview v0.2 has been public since August 23, and only its acknowledgements changed this week. It is the same distillation trick on a bigger animal, MiniMax H3, a 33B model that generates picture and sound together, cut to four steps. Yesterday's briefing flagged v0.1 as checkpoint 1400 of 4000 and not worth installing. This is step 2900, and the card corrects its predecessor: v0.1 told you to pass num_inference_steps=4, which quietly makes the scheduler invent its own grid instead of using the four points the student trained on. Pass the trained ladder explicitly, dmd_denoising_steps=[999, 749, 500, 250]. The same mistake corrupted the in-training validation clips, so judge it on fresh samples. Still soft on fine motion. MiniMax H3 Community License, not Apache (v0.2, hosted demo).
Image
Recraft V4 Styles arrived August 26 and fixes style drift without training anything. One reference image is enough to start and you can add up to ten more, a brand book, a photoshoot, a screenshot you liked, and every later generation holds that look. Precise mode locks rendering technique, color, composition and lighting; Flexible loosens the grip. You skip the LoRA training run entirely, which for a small studio is the difference between a style lock today and one next week. Free tier, no card required. Recraft cites third-party blind testing run August 14 in which evaluators preferred it over seven other models 92% of the time; the post does not name the evaluator (Recraft).
Audio and music
Nothing new shipped in audio this week, and the two things worth your time both shipped earlier and went largely unread.
Since August 2, Stable Audio 3 has had CPU-AMX engines: compiled C++ that runs the whole music and sound-effects pipeline on a processor with no PyTorch, no MLX, no graphics card, just shared libraries and numpy. A three-second clip takes about six seconds through the medium model at eight steps, and text-to-audio, audio-to-audio and inpainting all work. Read the catch first. These builds target Intel Xeon chips with the AMX matrix unit, Sapphire Rapids and Emerald Rapids, Linux x86-64 only. On any other processor the prebuilt engines are no faster than plain fp32 and you would rebuild from source. So this is cheap server rendering, not your laptop. The engineering notes are the best part: the team states that the sophisticated quantization methods made the diffusion transformer worse than naive rounding, and explains why, though it publishes no head-to-head number for that specific comparison. Stable Audio Community License, which terminates above $1M annual revenue (engine notes).
Xiaomi's ControlFoley released back in April, but its official Hugging Face Space only went up on August 14, which is the first time most people could try it without an install. It takes silent video and generates sound that lands on the action, steered either by a text prompt or by a reference clip whose timbre it copies while keeping timing locked to your footage. It is built for the case where your words and your picture disagree, which is the hard part of a foley tool. audio.cpp v0.7 added native C++ inference with GGUF packages this month. Weights are CC BY-NC 4.0, so personal and spec work only (model, try it free).
Open and local
The most consequential item of the week is a contract. On August 27, Comfy became the first official reseller of commercial-use licenses for MiniMax H3. Open weights let you download and run a model; they do not automatically let you sell what comes out. H3 generates 2K video with native stereo audio and, by Comfy's own account, runs on hardware as accessible as a 3060, and until Thursday the local path was legally murky for paid work. Professional and Enterprise tiers now cover local commercial use, LoRA training on your own characters, and downstream agency work; Enterprise includes the undistilled weights. Pricing lives on the license page, not the post. Reselling H3 access as your own inference platform stays excluded (Comfy).
Star totals verified via shields.io:
- comfyanonymous/ComfyUI: the node graph every local model here targets, and now the licensing counterparty too (131k stars) (repo)
- Lightricks/LTX-2: picture and sound generated in one pass, still the reference implementation for that idea (9.3k) (repo)
- NVlabs/Sana: home of the four-step 720p video model above (8.9k) (repo)
- hao-ai-lab/FastVideo: the lab turning 50-step video models into 4-step ones (4.1k) (repo)
- city96/ComfyUI-GGUF: loads compressed model files, which is how most people fit H3 on a consumer card (4k) (repo)
- kijai/ComfyUI-KJNodes: the utility pack most community video workflows depend on (3.2k) (repo)
- 0xShug0/audio.cpp: portable C++ audio generation, added ControlFoley in v0.7 (2.1k) (repo)
- Stability-AI/stable-audio-3: the runtime that pulls those CPU engines on first use (704) (repo)
Creative workflows
1. Render a multi-shot scene as one continuous take. 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).
The steps. Install ComfyUI 0.30.0 or newer, then the ComfyUI-H3-Multishot node pack. Load H3_Seamless_Chain_CORE.json first, because it needs zero third-party nodes. Weights are not in this repo: get GGUF quants from joeygambino/MiniMax-H3-GGUF and the text encoder plus both VAEs from Comfy-Org/MiniMax-H3. Sizing from the quant repo: Q8_0 for 32GB cards, Q5_1 for 24 to 32GB, Q4_0 for 16GB. Run once with every switch off and the memory reserve at 0 before touching anything, because the pack measures your card and sizes its own pool. For a long take, open H3_Extend_Take.json and set take_seconds.
How it works. H3 natively generates blocks of roughly 10 to 15 seconds. The pack renders those blocks in sequence and carries continuity across the joins two ways: boundary plates that pin the visual state, and a context pin that lets H3 continue speech in its own voice rather than handing off to a text-to-speech engine. Set voice_ref_2 and voice_ref_3, new in 2.7.0, and each character keeps their own voice across the chain.
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 cannot produce.
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 qwen3:14b as its script writer, which competes with H3 for the same card on anything under 32GB; use the CORE workflow or a remote endpoint instead. And H3's license is the reason the Comfy reseller news above matters.
2. Put sound on silent footage. Code at xiaomi-research/controlfoley, nodes at YJX-Research/comfyui-controlfoley-official.
The steps. Try it free in the browser first at the official Space. Locally: Python 3.10, PyTorch 2.5.1, CUDA 11.8, ffmpeg, then huggingface-cli download YJX-Xiaomi/ControlFoley --local-dir model_weights. Plain video-to-audio is python demo.py --video clip.mp4 --duration 6.0. Add --prompt "skateboarding" to steer it with words, or --audio reference.wav to copy a specific timbre. For ComfyUI, install ControlFoley from the Comfy Registry, which brings workflows for all five modes.
How it works. Two visual encoders run together, CLIP for what the picture means and CAV-MAE-ST for how sound and vision correlate in time. Reference audio is stripped down to timbre only, with temporal cues suppressed, which is why copying a sound's character does not drag its rhythm along with it.
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 cutting AI-generated or archival silent footage, this is the missing half.
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.
Worth testing
- SANA-Video 2.0 four-step demo: 720p video in four passes, Apache 2.0, no install. Tradeoff: text-to-video only, and crowded or fast scenes fall apart.
- ControlFoley Space: add synchronized sound to a silent clip in your browser. Tradeoff: non-commercial license, so this is for testing whether it works on your footage, not for shipping.
- FastH3 v0.2: four-step video with matching audio from a 33B model. Tradeoff: a preview at step 2900 of 4000, visibly soft on fine motion.
- Recraft V4 Styles: lock a visual style from reference images with no training run. Tradeoff: hosted and metered, and the quality claim rests on unnamed evaluators.
- H3 Turbo LoRA demo: the four-step adapter most people are actually using. Tradeoff: locally you still need the base H3 weights and a card to run them on.
What actually matters from today's signal
Spend this week on the license page, not the model list. The gap between what open video models can do and what you are permitted to sell has been the quiet tax on local creative AI for two years, and Comfy's reseller deal is the first clean, purchasable answer a freelancer has had for a major open video model. Apache 2.0 releases like SANA-Video sidestep the question entirely, which is why a 4.47B model that fits your card is worth more to you than a 33B one that does not.
The other number to track is the step count, because it decides whether a tool fits your working rhythm. Four steps versus fifty is not an abstract 12x, it is the difference between queueing a render and trying an idea. Every open-weights video release in this window shipped a distilled variant beside the full one. Assume that is permanent and stop planning around overnight renders.
The counter-signal is that the speed is bought with quality, and the honest releases say so out loud. SANA's card lists what falls apart. FastH3 corrects its own previous instructions and warns that the earlier validation clips were misleading. The H3 multishot pack measures its texture drift at 13% per join and tells you where it becomes visible. That disclosure is the reason to trust these projects and the reason to test before committing a deliverable to one. The vendor announcements this week, by contrast, offered a 92% preference number with unnamed evaluators and a 4K claim with no price.
Source access notes: Hugging Face API responses fetched without cache-busting came back roughly two weeks stale on this run (Stable Audio 3 showed a July lastModified and 1,913 downloads against a real August 28 and 13,224). Every date and count here comes from a ?cb= cache-busted request. Cache-busting protects the number but not the attribution: an adversarial check of this draft caught two items whose repo-level lastModified was genuinely recent while the feature being reported was weeks old. FastH3 v0.2 shipped August 23, not August 28, whose only commit was acknowledgements; the Stable Audio CPU-AMX engines shipped August 2, and this week's commits to that repo were TensorRT housekeeping. Both are dated correctly above and framed as earlier work. The Comfy licensing post was edited after publication: its URL slug and og:title still read "the only official reseller," while the rendered body reads "the first official reseller." The body wording is used here. runwayml.com/blog, the bfl.ai homepage, the blog.comfy.org homepage and about.ideogram.ai are JavaScript walls or undated; reached via changelog, /blog and /archive paths. stability.ai/news serves a stale legacy page with no dates; stability.ai/news-updates is live. deepmind.google/blog shows month-only granularity and carried no Veo, Imagen, Lyria or Flow item for August. Suno's release-notes page renders without dates, so an undated entry in the window cannot be ruled out. Nothing in window from Runway (latest Aug 24), Black Forest Labs (Aug 20), Midjourney (Aug 20), Pika (Aug 18), ElevenLabs, Udio, Kling or Seedance. Stability's $76M Series B dates August 25, one day outside the window, and is omitted. Writing is omitted: no text-AI release met the creation-relevant filter. Recraft's 92% figure and ControlFoley's benchmark wins are each party's own reported results, labeled inline.