Creative AI Briefing: Monday, September 21, 2026
Type "a cute cartoon dragon sticker" into a free browser demo and get back a PNG with a real alpha channel, cut out, no background-removal step, no magic wand, ready to drop into a layout. That shipped Sunday. Then something more useful happened underneath it: two people, working separately, one on images and one on video, spent the weekend measuring where the memory in a generation pipeline actually goes, and both published the same answer. It is not the model that draws the picture. It is the model that reads your prompt. The image people found a text encoder eating 16.3 GB of a 21.5 GB total. The video people found one eating 22.7 GB before it had drawn a single frame. Everyone has been compressing the wrong file.
New models
Qwen-Image-2.1 landed on Sunday, September 20, and the release note is dated that day. The weights had been sitting in the repository since a quiet initial commit on September 14, but the usable checkpoint went in on September 19 and the model card describing it went up on the 20th, which is when the rest of the ecosystem moved.
What you can make with it: transparent images, natively. The autoencoder has 64 channels and handles an alpha channel directly, so the model generates RGBA rather than generating a square and asking you to key it out afterwards. Stickers, logos, overlays, UI assets, product cutouts. The prompt format is specific and the team publishes it: start with "This is an RGBA image with transparency," describe the thing, and end with "The image has alpha channel and the background is transparent." It also takes up to ten reference images in one call, which is how the group-photo and outfit-assembly examples on the model card work, and it accepts local edits marked by drawing a circle or painting an annotation straight onto the input rather than building a clean mask. Native output is 2048 x 2048 at 40 steps.
The generation half is 7B parameters across 32 layers. The licence is the Qwen Research License Agreement, and section 2(a) is unambiguous: non-commercial purposes only, where non-commercial means research or evaluation. Selling anything made with these weights requires a separate licence from Hangzhou Tongyi Laboratory. Every community conversion below inherits that, and the honest ones say so on the front of their model card.
Free hosted demo, no install: huggingface.co/spaces/Qwen/Qwen-Image-2.1. Day-zero support landed in Diffusers (PR #14804), ComfyUI, vLLM-Omni, SGLang (PR #39983) and LightX2V. The GitHub repository sits at 770 stars.
Image
Three ports of Qwen-Image-2.1 appeared within 24 hours of the model card, and the interesting thing is what each one measured.
ModelsLab's four-bit builds went up early Monday morning in two flavours: an int4 build at 4.34 GB that runs on Ada cards and newer, and an nvfp4 build at 4.54 GB that is Blackwell-only and will not load on a 4090. Both compress the same layers that take about 10 GB in full precision. On an RTX 4090 the int4 build finishes 40 steps at 1024 x 1024 in 11.6 seconds, 1.84 times faster than full precision. On an RTX PRO 6000 Max-Q the same int4 build runs at 0.35 times the speed, which is to say considerably slower than not compressing it at all, and the author prints that number in the table rather than hiding it. The nvfp4 build on that same card does 7.7 seconds.
The quality check on the nvfp4 build is the part worth copying, and note that it is the nvfp4 build specifically. Rather than only reporting distance-from-full-precision, which defines full precision as perfect by construction, the author ran eight seeds across two text-heavy prompts and read all 48 renders by eye, scoring whether the sign actually said what the prompt asked. Full precision scored 14 of 16. The nvfp4 build scored 15 of 16. The write-up refuses the obvious headline: "A single prompt where the quantized build beats bf16 is trajectory variance, not evidence of superiority." The int4 card is equally careful in the other direction, saying outright that the int4 build was never put through that sweep and should be read as "the same recipe at slightly lower fidelity, on the far wider set of cards it runs on."
And then the line that connects to everything else today. Only the generation model was compressed. The prompt-reading model stays at full precision and accounts for roughly 16.3 GB of a 21.5 GB total, "so it, not this file, decides how many reference images fit on a card."
Devin Lai's Core ML conversion arrived Monday for Mac owners: 40 steps at 1024 x 1024 in 3.7 to 4.2 minutes on an M5 MacBook Pro with 32 GB, which is 2.4 to 2.6 times faster than the same model through PyTorch. Six packages, 14.74 GB total, macOS 15 or newer. The buried finding is that Apple's Neural Engine is the wrong place to run this: measured on two transformer blocks, CPU plus GPU beat CPU plus Neural Engine by 7.01 times. The catch list is long and stated plainly: text-to-image only, 1024 x 1024 only, no image editing, no classifier-free guidance, prompts capped at 64 tokens, and the text encoder is not included in the download.
There are also GGUF builds for ComfyUI carrying transparency through.
Video
FastWan2.2-TI2V-5B-MLX-q8 shipped early Monday, Apache 2.0, and it is the video half of the same discovery. One change from the conversion it builds on: the text encoder dropped to 8 bit. On disk that takes it from 11.4 GB to 6.0 GB. Resident it takes it from 22.7 GB, because some loaders promote it to full precision on the way in, down to 6.0 GB. The meaning of the prompt survives: the embedding sits at a cosine of 0.995 against full precision, against bfloat16's 0.997.
The author also shipped the negative result. Four bits was tried and deliberately not released: cosine 0.911, a 44% mean error, which as the card puts it "is a different prompt, not a cheaper one." That sentence is the most useful thing published about compression in a week.
End to end on an M2 Max with 32 GB at 832 x 480, a one-second clip takes 43 seconds and a five-second clip takes 311 seconds, of which the decoder is most of it. Peak memory stays near the largest single component rather than the 18.8 GB total, because the pipeline loads one piece at a time.
Also out late Sunday: Wan2.2-TI2V-5B bundled in four precisions with the denoiser, text encoder and autoencoder packaged together so the pieces cannot drift apart, from an outfit calling itself MECHA Forge. Worth noting the int4 card's heading still says FP8, a copy-paste slip that tells you to trust the tags over the prose.
Audio and music
LiuBai-RWKV published its 2026-09-20 milestone, and it is an unusually honest failure report. It is a 3B song model that takes a style caption plus lyrics and returns 48 kHz stereo with vocals, trained on tokens from roughly 93,000 songs. Genre, instrumentation and language follow the caption. The vocals are fluent. And the sung words are mostly not your lyrics: word error rates against the prompt run 0.90 to 1.09 across the twelve styles in the table, where the original human recordings score 0.47 on the same transcriber. Twenty full samples sit on the model card, so you can hear the gap yourself.
The author names the cause, names the fix in progress, and states the gate it has to clear before the full run. CC BY-NC 4.0, forced by the renderer it depends on. Do not download this expecting a tool.
Open and local
The weekend's local story was not a new model. It was two people independently checking which file in their pipeline was actually full, and finding the same answer.
Comfy-Org's own repository confirms it without saying a word. Their Qwen-Image-2.1 bundle ships the generation model in two precisions and the Qwen3-VL prompt reader in three, including a four-bit build, plus two compressed prompt-rewriting models. The people who maintain the loaders already knew where the weight was.
- nunchaku-tech/nunchaku: the four-bit runtime both ModelsLab builds are packed for, and the reason the speedup exists at all rather than just the file shrinking (repo, 4k stars per shields.io).
- ModelTC/LightX2V: day-zero acceleration for Qwen-Image-2.1, aimed at consumer cards as well as datacenter ones (usage guide, 2.8k stars).
- QwenLM/Qwen-Image-2.1: the reference code, prompt-rewriting scripts and the licence you are agreeing to (repo, 770 stars).
- realrebelai/LOW_VRAM_Workflows: ComfyUI graphs stress-tested against an 8 GB card with 16 GB of system memory, Apache 2.0, updated Monday morning (repo, 52 likes).
- xiaol/LiuBai-RWKV: the music model above, with its dated experiment log (code repo, 1 like on the weights repo, brand new).
Star totals above are current shields.io figures, not daily movement.
Creative workflows
1. Generate a transparent sticker sheet with no cutting out. Open the free demo or load the ComfyUI template at templates/image_qwen_image_2_1_t2i.json.
The steps. Write the prompt in the three-part shape the team documents, because the model was trained on it: open with "This is an RGBA image with transparency," then your subject, then close with "The image has alpha channel and the background is transparent." Set 40 steps. Pick a size from the published aspect list rather than typing your own. Save as PNG.
How it works. The compressor that turns the model's internal picture back into pixels has a fourth channel for opacity, so transparency is something the model generates rather than something you subtract afterwards. That is why the edges hold on hair and glass, where background removal usually fails.
Why it is good. It removes a whole step and the artefacts that come with it. No halo, no matte line, no cleanup pass.
Where it breaks. The opening and closing sentences are load-bearing; drop them and you will get an opaque square. The licence is research and evaluation only, so a sticker pack you sell is out until someone buys a commercial licence. And the ComfyUI download is large because the prompt-reading half is large.
2. Edit by drawing on the picture instead of masking it. Template: templates/image_qwen_image_2_1_image_edit.json.
The steps. Load your image. Draw a rough circle around each region you want changed, or paint over it, straight onto the input. Write one instruction per region in plain language. Queue once. The model card's worked example changes three things in a single pass: removes a watch, changes hair colour, replaces clothing.
How it works. The prompt-reading model takes text and images through the same path, so a circle on the picture is an instruction in the same sense a word is. You are not supplying a precise selection, you are pointing.
Why it is good. Circles take seconds. Clean masks take minutes, and for three regions you would otherwise be doing three passes.
Where it breaks. Pointing is not selecting. Where two regions overlap, or where the circle covers something you wanted untouched, you get a judgement call rather than a boundary, and you will not know which until you look. Multiple reference images are also where memory goes: the prompt-reading model holds them, so ten references is a card-size question, not a prompt question.
3. Compress the encoder, not the model. Reference implementation: FastWan2.2-TI2V-5B-MLX-q8, file t5_encoder_q8.safetensors.
The steps. Check what your pipeline loads and how big each piece is. If a full-precision prompt-reading model is in there, swap in an 8-bit one. In the video case above that is 6.0 GB against 11.4 GB on disk and 22.7 GB resident. On the image side, Comfy-Org ship the equivalent files already. Do not go to four bits.
How it works. The encoder runs once per generation and is then released. Rounding its weights harder costs you a little precision in a step that happens once, not in the step that repeats forty times.
Why it is good. It is the single largest memory saving available in most local setups, and unlike compressing the generation model it does not change the picture's trajectory at all, only the prompt's reading of it.
Where it breaks. Four bits destroys it. Cosine 0.911 means the model is answering a different prompt. And a loader that is not expecting compressed weights will fail on shapes rather than on a missing file, which is exactly why the author renamed the file.
Worth testing
- Qwen-Image-2.1 hosted demo for transparency and multi-reference composition. Free, nothing to install. Tradeoff: research licence, so treat anything you make as a test.
- ModelsLab nvfp4 if you own a Blackwell card. 7.7 seconds for 40 steps and text that reads correctly. Tradeoff: it will not load on anything older, and you still need 21.5 GB because the encoder is untouched.
- Core ML conversion on an M-series Mac with 32 GB. Tradeoff: 14.74 GB download, text-to-image only at one fixed size, no editing, and the text encoder is a separate errand.
- FastWan2.2-TI2V-5B-MLX-q8 for local video on Apple Silicon, Apache 2.0 all the way through. Tradeoff: 311 seconds for five seconds of 832 x 480, and the decoder is most of that.
- LiuBai-RWKV samples if you want to hear what a music model sounds like when the style lands and the words do not. Tradeoff: it does not work yet, and the author says so first.
What actually matters from today's signal
The compression conversation has been pointed at the wrong object for about a year. Every model card leads with the generation model's file size, every forum thread argues about which quantization of the generation model looks better, and two people independently checked this weekend and found the same thing: on a real machine, running a real pipeline, the thing reading your prompt is the biggest resident object. ModelsLab measured 16.3 of 21.5 GB. The MLX conversion measured 22.7 GB, more than the generation model and the decoder combined. Comfy-Org, who have to make this work for everyone, ship five encoder files and two model files. The conclusion has been sitting in plain sight in the file listings.
What this buys you is concrete. If you have a 16 GB card and you have been told you cannot run a model, check whether the pipeline is loading a full-precision encoder and promoting it on the way in. That one substitution is worth more than any amount of arguing about four bits versus eight on the generation model, and unlike compressing the model it does not touch the image at all. The encoder runs once. The generation model runs forty times. Round the thing that runs once.
The counter-signal is the licence, and it is not small. Qwen-Image-2.1 is the most interesting image model released this month and its terms permit research and evaluation only. Every port above says so on the card, which is to their credit, but the pattern is worth naming: an enormous amount of unpaid community engineering went in over the weekend to make a model run on cheap hardware, and none of the people doing that work, or using it, can sell the output. That is the deal on offer. It is worth taking with clear eyes rather than discovering three months into a project, and the RTX PRO 6000 result is the reminder that "quantized" and "faster" are two different claims. Measure on your own card before you believe either.
Source access notes: Hugging Face dates verified via the API createdAt field, never listing "Updated" timestamps; the Qwen-Image-2.1 date is corroborated by the commit log (release checkpoint 2026-09-19, model card 2026-09-20) and the dated News entry in the GitHub README. Star totals from shields.io with cache busting. ComfyUI's Substack blog is a JavaScript wall and returned no post content, and the GitHub Releases API returned an empty body twice, so ComfyUI support was confirmed through the Qwen repository instead. The Hacker News Algolia endpoint returned zero hits for the creative query set. Midjourney and Civitai were not reached. OpenAI, Runway, Stability, Adobe, ElevenLabs, Black Forest Labs, fal and Replicate were checked and published nothing creative-facing in the window. One item was investigated and dropped: the SYNAPSEai1/SynapseMusicV12 cluster of fourteen repositories created 2026-09-20 is a mirror of ACE-Step v1.5 originals from January to June 2026, carrying a verbatim copy of the upstream card, and is not a release. It caught four errors, all corrected above: a stale star count for QwenLM/Qwen-Image-2.1 (767, now 770 per shields.io); the Rin247/MECHA Forge Wan2.2 bundle dated Monday when all four repositories have a createdAt of late Sunday, September 20; the 15-of-16 text-accuracy result attributed loosely after a paragraph about the int4 build when that sweep was run only on the nvfp4 build, which the int4 card itself says plainly; and a Hugging Face "likes" figure for LiuBai-RWKV labelled as GitHub stars. The pass confirmed the remaining load-bearing numbers, dates, licence wording, file counts, links and the meta-claim against primary sources.