Creative AI Briefing: Tuesday, September 15, 2026
Open a browser tab, type a sentence, wait, and get five seconds of video with a soundtrack generated on the same timeline, in three denoising passes, on a shared A10G, for free. That demo went up yesterday afternoon. The model underneath it was published eight days ago by Alibaba's TaoLive team as part of a streaming runtime that needs eight H20 cards to run as designed. Somebody pulled the one file that mattered out of it, renamed 208 tensors, and now it loads in a stock ComfyUI graph. That is the shape of the whole week: the interesting capabilities are arriving as extractions from vendor systems, not as products.
New models
A three-step adapter for MiniMax H3 escaped its datacenter runtime and landed in ComfyUI. TaoLiveAIGC/TaoMate-H3 went up September 7 (Hugging Face API createdAt 07:31 UTC): the step-3000 generator EMA adapter, rank 128, alpha 128, for Alibaba's low-latency streaming audio-video runtime. Read the card and the product is obvious: it generates synchronized audio and video in small chunks, takes one prompt per five-second block through a --prompt-json file with per-block seeds, holds identity and voice across prompt boundaries with a clean KV cache, and does 480p, 768p and aligned 1080p. The published numbers are from one node of 8x H20 96 GB with TP2 and Ulysses4: pure denoiser time 14.810 s against MiniMax H3's 169.572 s on the same 10-second 480x864 job, and the first chunk reaching its final state in 6.148 s against 170.052 s. Peak memory barely moves, 31.37 GiB against 32.03. Hopper-class GPUs only, validated on eight of them.
Almost nobody reading this has eight H20s. So on September 13 at 07:50 UTC, CZMartin22/TaoMate-H3-3step-ComfyUI republished just the adapter, converted out of PEFT naming into ComfyUI's tensor layout, alpha scalars embedded per module, cast from FP32 to BF16. The file is 1,240,539,224 bytes against the original repo's 2.48 GB, and the card promises all 208 patches attach without a missing-key warning. Settings are non-negotiable and stated as such: 3 steps, CFG exactly 1.0, LoRA strength 1.0, Euler, simple or linear flow-match scheduler, on the FL2VA variant of the base. Guidance above 1.0 burns a distilled sampler; the card says so in capitals. Asirus/TaoMate_H3_3_Step_LoRA followed September 14 at 20:38 UTC.
The adoption gap is the story. Read this morning, the official Alibaba repo shows 175 downloads and 64 likes. The ComfyUI conversion shows 1,596 downloads and 22 likes in roughly two days. What survived the extraction is the speed. What did not survive is everything the runtime was actually for: the chunked streaming, the per-five-second prompt switching, the clean-KV continuity across prompt boundaries, the long-form generation. You get a fast adapter, not a live pipeline. Both inherit the MiniMax H3 Community License, which still excludes the EU, the UK, South Korea and the United States.
Free hosted demo, running now: akhaliq/taomate-h3-3step-demo (created September 14 14:30 UTC, RUNNING on zero-a10g) and its sibling hugging-apps/taomate-h3-3step-demo. Five seconds of audio-video, no install. The Space declares the base, the official adapter, the ComfyUI conversion and multimodalart/qwen3vl-conditioner as its model dependencies, which tells you exactly which four pieces the extraction needed.
Image
Krea 2 has no image editing. Somebody added it anyway, and the community started shipping adapters within a day. Krea 2 is a text-to-image model: a single-stream diffusion transformer, 28 layers, a Qwen3-VL-4B-Instruct text encoder and the Qwen-Image autoencoder. No reference-image path. The krea2 architecture in ostris/ai-toolkit (12k stars) now carries a model_kwargs.edit flag that adds one, and ComfyUI-Krea2-Ostris-Edit (329 stars) is the node pack that runs the result. Two edit adapters appeared in the last 24 hours: Cierpliwy/krea2_lineart_edit (September 14, 20:37 UTC, 8 likes), which turns a line drawing into a finished image and whose four-line card says it works best with the realistic lineart preprocessor from comfyui_controlnet_aux, and WarmBloodAban/Krea-Edit-lora (September 14, 08:11 UTC), a four-view character turnaround adapter that ships its ComfyUI workflow in the repo and runs at 8 steps, CFG 1.0.
One warning on that second one. Its card declares license: apache-2.0, and Krea 2 is not Apache 2.0. It is the Krea 2 Community License with a separate acceptable use policy. An adapter cannot relicense the model it needs. The upstream terms govern.
Mixed-bit image models with the calibration written down. justintime47/Z-Image-Turbo-iQ4.5 landed September 15 at 00:48 UTC, Apache 2.0, and it is the most carefully documented compression job of the week. The upstream model is 32.9 GB at full precision; this pack is 5.91 GB. What makes it worth reading is the method, which the card publishes in full: activation statistics collected for all 276 transformer linear layers while the server generated 73 real prompts, twenty of them at full size and 53 at 256px and one step, because the caption path needs many prompts rather than large images. Every layer's error was then measured at 3, 4, 5, 6 and 8 bits and the widths bought by error saved per byte, which is why the transformer ends up at 3-bit for 55 layers, 4-bit for 154, 5-bit for 29 and full precision for 38. The conditioning path, the final layer and every block's modulation stay wide on purpose. The autoencoder is untouched so adapters still compose. Same author, same method: Krea-2-Turbo-iQ4.5 (September 13) and FLUX.2-klein-9B-iQ4.5 (September 13).
The catch is large and the card states it first: no released build of mlx-serve (1.3k stars, MIT) can load it. You build from the author's feat/zimage-calibrated-quant branch, which also carries a text-encoder cache fix, and without that fix every request after a server's first is conditioned on the earlier prompts. That is a silent failure: you get an image, it is just wrong, with nothing to tell you.
Video
Beyond TaoMate, the last 24 hours on the Hub were conversions and mirrors. frtertaer/LTX-2.5-uncensored-v1.1-FP8 (September 15, 03:58 UTC) declares license: unknown on a base that is not unknown at all. JoaoZaokk/LTX-2.3-22B-distilled-1.1-W4A8-ConvRot (September 14, 04:50 UTC) continues that account's run backwards into the 2.3 base. rAVEUK/LiveWan (September 14, 13:21 UTC, Apache 2.0) advertises real-time streaming distillation over Wan 2.1 1.3B and publishes no measurement yet.
Writing and research
StepAudio 3 Gen surfaced on Hugging Face's daily papers with 71 authors and a genuinely different design: a discrete autoregressive generator over residual vector quantization tokens covering speech, voice design, singing, sound effects and music in one model, deliberately not a diffusion transformer. No weights exist on the Hub under any Step-Audio-3 name as of this morning. Paper only, so treat it as a thing to watch rather than a thing to use.
Open and local
The pattern in this section today is people publishing their working notes alongside the file. Three of the four items below tell you what the thing costs, not just what it does.
- ostris/ai-toolkit (12k stars): the trainer that now grows a reference-image path for Krea 2. For a creator, this is where a new editing behaviour gets taught before it exists anywhere else. Why now: the
krea2architecture's edit flag, plus a matching ComfyUI pack, turned a text-to-image model into an editable one in a week. - ostris/ComfyUI-Krea2-Ostris-Edit (329 stars): two nodes, no dependencies. Without them, stock ComfyUI ignores reference images on Krea 2 entirely and you get a plain text-to-image render with no error.
- ddalcu/mlx-serve (1.3k stars, MIT): one local server on Apple Silicon for images, video, music, speech and 3D. Why now: it is the runtime the week's most careful image quantizations target, and the ones that matter need an unreleased branch.
- TaoLiveAIGC/TaoMate-H3 (245 stars): the streaming runtime whose adapter is being used everywhere and whose actual product is being used nowhere. Read it for the prompt-per-five-seconds design even if you cannot run it.
Creative workflows
1. Teach Krea 2 an edit it was never trained to do. From the krea2 architecture source in ai-toolkit and the ComfyUI-Krea2-Ostris-Edit README.
The steps. Build a paired dataset: source image as control, edited result as target, matching filenames. Set the architecture to krea2 and model.model_kwargs.edit: true. Leave kv_cache off. No Krea 2 example config exists in config/examples on main, so start from train_lora_qwen_image_edit_2509_32gb.yaml and change the arch. Then in ComfyUI, clone the node pack into custom_nodes and wire Load Diffusion Model (krea2) into Load LoRA into Krea 2 Ostris Edit Model Patch into your sampler, with the positive branch running through Text Encode Krea 2 Ostris Edit carrying your prompt, up to three reference images and the autoencoder.
How it works. 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 zero where they are never noised. The denoiser predicts only the target. References keep their own aspect ratio and are never resized to match the output.
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.
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 the kv_cache toggle only works on an adapter trained with kv_cache enabled, because the base was trained fully bidirectional; turning it on otherwise is a correctness bug that still produces an image.
2. Run a three-step audio-video adapter without the runtime it was built for. From the TaoMate-H3 3-step ComfyUI card.
The steps. 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 attached; a partially attached adapter still renders. Budget the base download properly: TaoMate's own instructions pull only FL2VA/* from MiniMaxAI/MiniMax-H3, and that subtree sums to roughly 144 GB at full precision (the whole repo's usedStorage is 354 GB). To try it first, the hosted Space runs the same stack free.
How it works. 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.
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.
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 own, so the 11.45x belongs to an eight-GPU node. And no quality comparison exists anywhere against the full-precision original, the older turbo adapters, or plain H3.
3. Catch a compressed model that renders fine and is wrong anyway. Carried forward as method, prompted by the Z-Image iQ4.5 card's own warning.
The steps. 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 text-encoder cache is not being cleared.
How it works. Servers cache prompt embeddings to avoid re-encoding. If the cache key is wrong or missing, later requests reuse earlier conditioning. Nothing errors.
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.
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.
Worth testing
- akhaliq/taomate-h3-3step-demo. Free, running, no install, five seconds of synchronized audio and video. Tradeoff: shared hardware, short clips, and the licence on the underlying weights excludes four major markets.
- ComfyUI-Krea2-Ostris-Edit plus krea2_lineart_edit. The fastest way to see whether reference-image editing on Krea 2 is worth your time. Tradeoff: the adapter's card is four lines long, with no training details, no evaluation and no sample settings.
- WarmBloodAban/Krea-Edit-lora. Four-view character turnarounds from text or a reference, workflow JSON included. Tradeoff: the licence field on the card is wrong; the Krea 2 Community License governs whatever you make.
- justintime47/Z-Image-Turbo-iQ4.5. 5.91 GB instead of 32.9 GB on a Mac, with the calibration method published. Tradeoff: needs a build from an unreleased branch, and no released version of the server can load it.
What actually matters from today's signal
The labs published nothing yesterday. OpenAI, Google DeepMind, Adobe, Runway, Black Forest Labs, Stability, Luma, Midjourney, Kling, ElevenLabs and Suno were all quiet, and fal and Replicate's most recent posts are weeks old. What filled the gap was five or six people with conversion scripts, and the work they did has a consistent shape worth naming: they took systems and extracted parts.
That is a good deal more often than it is a bad one. A 1.24 GB file that makes a three-step sampler work on one card is worth more to most people than an eight-GPU streaming runtime they will never provision. But the extraction has a cost the extracted file never states, and TaoMate shows it cleanly. Alibaba's product is continuity: prompt switches every five seconds, identity and voice held across the seam, generation that keeps going. The community's product is the speed alone. Downloads say the market prefers the fragment, 1,596 against 175. That does not make the fragment the same thing. Worth knowing before you call it new: lightx2v/Minimax-h3-Turbo has had 1,452,743 downloads since August 7, and larryvrh/MiniMax-H3-Turbo-Lora 223,215 since August 5. Three steps is fewer than four, and that is the whole margin.
The second cost is legal and it is getting sloppier. Three cards read this morning declare a licence their base model does not have: Apache 2.0 on a Krea 2 adapter, unknown on an LTX 2.5 derivative, and yesterday's MLX build of a non-commercial music model. A converted file inherits the upstream terms whatever its metadata says. The tempting read is that the Apache-tagged turbo adapters route around the MiniMax territory clause; they do not, because an adapter is useless without the base, and the base is what the clause covers.
What I would do with today: the hosted TaoMate demo costs nothing and settles the speed question in five minutes. And if you work in images, the Krea 2 edit path is the week's real new capability, with the honest caveat that both available adapters are hobby work with no documentation. That gap takes a paired dataset and an afternoon to fill.
Source access notes: An adversarial fact-check pass ran against this draft before publication. It falsified nothing: every ship date, byte count, download and like figure, star count, timing, bit-width table, licence clause and code-level claim above was re-derived from primary sources and matched, including the MiniMax H3 Excluded Territories wording, the ai-toolkit source constants, and the 71-author count on the StepAudio paper. Two things it could not settle, now reflected above: the exact node name of the lineart preprocessor the Krea 2 adapter recommends, which is why that sentence now quotes the card rather than naming a node; and the live status of the two Spaces, which is true as of this morning's read and can change at any time. It also confirmed that justinluque/mlx-serve, the repository carrying the branch the Z-Image pack needs, is a fork of ddalcu/mlx-serve rather than an unrelated project. 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 15. Download and like counts are single reads of the same morning and move. The api.github.com endpoint returned empty bodies through the fetch tool all run, so repository file listings were read through a browser session and file contents through raw.githubusercontent.com with cache busting; GitHub's code search requires a sign-in and was not used. blog.comfy.org and civitai.com were not checked in usable form this run. No creative-model release was published by any frontier lab inside the 24-hour window: fal's newest post is dated August 27 and Replicate's August 4, both read this morning. StepAudio 3 Gen is paper-only; a Hub search for Step-Audio-3 returned no matching weights. The claim that ai-toolkit ships no Krea 2 edit example config is based on a read of config/examples on main this morning. The Image section omits nothing; the Audio and music section is omitted because nothing inside the window met the bar.