FervorCreative AI
Live Latest 26.09.26 · morning 86 tools tracked 240 workflows indexed 211 topics Hot: MiniMax H3, Qwen-Image-2.1, ComfyUI

Creative AI Briefing September 26, 2026 morning edition 11 min read

The version of last week's model you can actually run this weekend was not made by the lab that trained it, it was made by somebody else three days later, and it is between five and eleven times smaller or faster.

Qwen-Image-2.1-viggle-turboMing-Image-0.1-DesignZ-Image-TurboGoogle Flow ToolsSparkWanFLUX 3 Actionimage-genlocal-creative-aicomfyuiopen-weightslicensing-provenancedesign-toolscreative-workflowsaudio-gen

Creative AI Briefing: Saturday, September 26, 2026

Type a prompt into a free web page this morning, wait about six seconds instead of thirty, and get an image that most people cannot tell apart from the forty-step original. That page is a distilled student of Qwen-Image-2.1 built by Viggle, and the version behind it is two days old. Hold it next to what else landed this week: a 71 GB design model repackaged into a single 6.3 GB file for ComfyUI yesterday morning, and Alibaba's Z-Image running on a machine with no graphics card at all. Nobody launched a new frontier model in the last 48 hours. What happened instead is that three separate people took models released in the last ten days and cut them down until they fit on hardware normal people own. The compression is the release now.

New models

Qwen-Image-2.1-viggle-turbo from Viggle went up September 22 (HF createdAt 2026-09-22T04:14:56Z) and hit the version worth using, v0.2.1, on September 24 by the card's own dating. It is a distilled student of Qwen-Image-2.1 that does text-to-image and instruction editing with 1 to 3 reference images in six passes through the model instead of forty, with no guidance scale. The authors measure it at roughly 5x faster end to end and claim it is hard to tell apart from the base model on most prompts. They also say exactly where it loses: small, dense text, where forty steps still wins and eight steps narrows the gap.

That last part is why this release is worth your attention over the dozen other speed adapters this month. The card publishes its own losses. It names a held-out set of 96 user requests, reports that this version drifts the composition on 0% of them against 4% for the version that launched a day before it, and gives you the exact six sampling values to use (sigmas=[1.0, 0.9375, 0.875, 0.75, 0.5, 0.25]) plus a rule for what happens if you change them: add or remove steps only at the noisy end, because the layout is decided in the first stretch and one big jump there ghosts the figures. The known-limits section admits that multi-reference composition, face swaps and identity-preserving edits still fall short of the base model.

The catch is the licence, and it is a hard one. The weights carry the Qwen RESEARCH LICENSE: research or evaluation only, no commercial use, commercial terms by separate agreement. So this is a tool for deciding whether the speed is real, not a tool for client work.

Free hosted demo, and it is the right first ten minutes: the official Space has a Comparison tab with 32 of Qwen's own example prompts rendered turbo-in-6-steps against base-in-40, same prompt, same inputs, same seed, in a slider you drag. Judge it yourself rather than taking the table's word. Two community mirrors of the Space went up on the 24th and 25th (one, another) if the original is queued.

Image

Yesterday's story was Ming-Image-0.1-Design needing a rented card with 80 GiB of memory. This morning at 08:55 UTC, a repackaged INT4 version went up (HF createdAt 2026-09-26T08:55:01Z) as one file for ComfyUI, 6,296,983,672 bytes, against the original repository's 71,503,501,573 bytes of storage for a 6.15B parameter model. The licence stays MIT, and a copy ships in the repo.

Read the fine print before you clear the disk space. The file is the diffusion transformer only. The author states plainly that the text model, connector, MLP and VAE are not included, so you still pull those from the original repository or another ComfyUI package. It also needs a recent ComfyUI build with both Ming-Image and native INT4 ConvRot support, 102 feed-forward layers were converted while the attention and normalization layers stayed at full precision, and the author calls it "a test conversion" and tells you to check output quality on your own prompts first. Same author shipped Apple silicon builds on September 24 (4-bit, 8-bit), also MIT.

The gap nobody has closed: every one of these conversions is of the generator. Ming-Image-0.1-Design-Layer, the model that takes a flat poster apart into separate transparent layers, still has no quantized build anywhere, only the original and one straight mirror. The half that changes a working day is the half still stuck on rented hardware. Free demo for it is here and remains the only practical way in.

Third data point on the same trend, and the most extreme: POCKET-Zimage-CPU went up September 25 and runs Alibaba's Z-Image-Turbo with no GPU, no CUDA and no Python inference stack, on stable-diffusion.cpp. Reference numbers from the author, on two Intel Xeon Gold 6526Y chips with 48 threads and zero GPUs: 512x512 in 46.4 seconds, peak RAM 6.42 GB. The Space itself sits on shared CPU hardware and the README warns it will be slower than that. Base model is Apache 2.0, runtime is MIT, so unlike the Qwen student this one has no licence cloud over the output.

Video

SparkWan from Alibaba filled in its Wan2.1 line on September 24 (HF createdAt 2026-09-24T07:55:00Z through 08:18:40Z across the batch), joining the Wan2.2 checkpoints from the same morning: 1.3B and 14B at 480P, 14B at 720P, and a three-step distilled 720P build. Apache 2.0 on all of them, which for video work matters more than the speed claim. The pitch is three or four passes with 90% to 97% of the attention thrown away.

Two more sparse-attention builds landed in the same window, both labelled previews by their own authors: Veda-Sparse put up an eight-pass MiniMax-H3 on September 25 (createdAt 2026-09-25T10:56:07Z), and memset0 has been iterating Wan2.1 builds across September 23 to 25. Treat them as previews.

Audio and music

Google shipped six new Google Flow Tools on September 23, each built by a working creative rather than by Google, and the sound one is the standout. Mondo Sónico, by creative director Ricardo Villavicencio and sound designer Sebastián Carvallo, generates background ambience, foley and contextual effects and delivers them synchronized across separate, editable tracks, with individual stems you can export for post. Generated audio that arrives already split is a different product from generated audio that arrives as one mixed file, and almost every music and sound tool this year has shipped the latter.

The other five automate jobs you currently do by hand: CaptionCast transcribes, styles and animates multilingual captions in one pass; ThumbnailForge formats cover art across platforms from one image and a headline; Surface generates architectural materials and maps them onto 3D walls and floors in real time; CollageMotion Pro animates mixed-media collages from text; SwissFlow Studio turns scripts into Swiss-style motion graphics. Every link opens the actual tool, and Google says you can duplicate and remix any of them. No pricing detail appears in the post, so the answer is whatever your Flow plan already is.

Open and local

Three of the four items above are somebody taking a model they did not train and making it fit: 40 steps to 6, 71 GB to 6.3 GB, a graphics card to no graphics card. All three published their own failure cases, which is what makes them usable rather than just impressive.

Creative workflows

1. Run the six-step Qwen student in ComfyUI without losing the speedup to a bad loader. Files and workflows are in comfyui/: viggle_turbo.py goes in ComfyUI/custom_nodes/, then drag in Qwen-Image-2.1-viggle-turbo-t2i.json or Qwen-Image-2.1-viggle-turbo-edit.json.

The steps. 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.

How it works. 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.

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.

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 older than 580 the prompt enhancer throws a CUDA error; turn it off or update. And the edit workflow sizes output from the first reference image while the reference pipeline uses the last, so the same inputs give you a different aspect ratio in the two tools.

2. Get Ming-Image's design-and-type generation onto a consumer card. Pull ming_image_0.1_design_int4_convrot.safetensors into ComfyUI/models/diffusion_models/.

The steps. 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.

How it works. 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.

Why it is good. MIT on both halves. You can ship what comes out.

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 hurt.

3. Generate foley and ambience as separate stems. 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 Google's plan structure and there is nothing to run locally.

Worth testing

  1. The Viggle comparison slider, Comparison tab. Free, no install, 32 matched pairs. Tradeoff: the weights behind it are research-licence, so a good result here does not become client work without a separate agreement.
  2. POCKET-Zimage-CPU. Free. Tradeoff: the 46-second number came from a 48-thread server, and the Space itself is on shared hardware, so budget minutes per image.
  3. The Ming-Image layer decomposer. Free, and still the only way to try it. Tradeoff: no quantized build exists, so liking it leads nowhere you can run locally today.
  4. Surface in Google Flow, for material studies before you commit to a render. Tradeoff: real-time mapping onto simple geometry is not the same as a render, and Google shows GIFs rather than stills you can inspect.
  5. The three-step SparkWan 720P checkpoint. Apache 2.0, so commercially clear. Tradeoff: 97% of the attention discarded is a lot to discard, and the authors publish no side-by-side you can drag.

What actually matters from today's signal

The useful release calendar has decoupled from the announcement calendar. Ming-Image-0.1-Design went up September 17 and was unusable to almost everyone for nine days; it became usable yesterday morning, from a stranger, for free. Qwen-Image-2.1 went up September 14 at forty steps; it became a usable six steps on September 24, from a company that did not train it. Z-Image has been out for weeks; it stopped needing a graphics card on September 25. If you plan your week around vendor blog posts you are consistently a week to ten days early, watching a model you cannot run.

So stop treating a launch as a date and treat it as the opening of a window. When something big drops, note it and wait. The version that fits your machine arrives within a week or two, usually from one person, usually with a README more honest than the original. This week's conversions state their own failure modes in plain language. The labs' own cards rarely do.

The counter-signal is the licence, and it is getting worse rather than better. The most impressive thing this week, the six-step Qwen student, is research-only, and so is its base. The MIT and Apache items (Ming-Image, Z-Image, SparkWan) are shippable but individually less exciting. The people doing the compressing cannot fix this: a derivative inherits its parent's terms, so every clever adapter built on a research-licensed model is a demo forever. Watch which base models the compressors pick. When they start choosing Apache-licensed parents over better-performing restricted ones, the ecosystem will have decided something. Right now they are still chasing the best model and inheriting its handcuffs.

Source access notes: Hugging Face's sort=createdAt filter feeds served badly stale pages this run (the text-to-video feed topped out at September 5, text-to-image at July 28), so discovery ran through sort=lastModified with every date confirmed against the per-repo API createdAt field with cache-busting. blog.comfy.org is a Substack JavaScript wall as usual. GitHub's API and release-atom endpoints for ComfyUI returned empty bodies this run, so no ComfyUI release notes here. replicate.com/blog's index appears frozen at April 2026. elevenlabs.io/blog and stability/luma/midjourney showed nothing inside the window. civitai.com skipped per standing note. An adversarial fact-check pass ran on this draft and caught three date errors, all corrected above: the Viggle repository's API createdAt is 2026-09-22 rather than the 09-24 its own version label carries (both are now stated); the bfl.ai FLUX 3 Action post carries a published timestamp of 09-22 against the Hugging Face write-up's 09-23 (both now stated); and Qwen-Image-2.1's createdAt is 2026-09-14T03:47:26Z, not the "around September 20" the draft claimed, which understated how long the base model sat before anyone made it fast. The pass confirmed every figure in the Viggle, Ming-Image, POCKET-Zimage, SparkWan, Veda-Sparse and Google Flow items, and confirmed that no quantized build of the Ming-Image layer decomposer exists.