Creative AI Briefing: Friday, September 25, 2026
Hand a flat JPEG of a poster to Ming-Image-0.1-Design-Layer, tell it how many pieces the design has, and it gives you back separate transparent PNGs: the headline on its own, the badge behind the headline on its own, the hero product on its own, the background absorbing the props and shadows. MIT licensed. Hold that next to NVIDIA's diarization model, which takes a recording apart by who was talking, and Alibaba's SparkWan checkpoints, which take the compute apart and run Wan2.2 in four steps. Three different teams shipped the same idea this week from three directions. The interesting work is no longer only about making a new file. It is about getting the parts back out of a file you already have.
New models
Ming-Image-0.1-Design and its sibling Ming-Image-0.1-Design-Layer both went up on September 17 (HF createdAt 2026-09-17T07:17:31Z and 2026-09-17T11:36:00Z) and got their card and asset refresh on September 23. They are two 6B models from inclusionAI aimed squarely at design rather than at pictures in general: UI screens, infographics, posters, anything with a lot of type in it. The first generates the design. The second pulls an existing flattened design back into editable RGBA layers.
The licence is the headline. Both repos carry MIT, stated in the card metadata and in a LICENSE file inside each repo. That matters this week specifically, because the transparent-PNG conversation for the last ten days has been about Qwen-Image-2.1, whose packaged ComfyUI files carry a qwen-research restriction. Here you get alpha-channel output and layer separation with no evaluation-only clause hanging over the result.
What it costs you is hardware. The README for both models names the same validated deployment: one CUDA GPU with at least 80 GiB of memory, running in BF16, 12 sampling steps, CFG of 1.0 for generation and 2.0 for decomposition. That is a rented card, not your desk. Weights are roughly 71 GB of storage for the generator and 65 GB for the layer model. Two community MLX conversions appeared on September 23 (4-bit, 8-bit, both createdAt 2026-09-23) for Apple silicon, but neither carries a quality note or a benchmark, so treat them as an experiment rather than a shortcut. The Ming-Image code repo sits at 93 stars, which tells you how early this is.
Free hosted demos exist for both and are the right first ten minutes: the design generator and the layer decomposer. The repo also ships a UI/UX design leaderboard image credited to Artificial Analysis; that is a chart the authors put in their own card, so read it as their claim until you check the source.
Image
The transparency trick in Ming-Image is unusually explicit, and worth copying whatever model you use. The code repo publishes ten exact phrases and says to place exactly one of them at the start of your prompt, never two. Five are English: RGBA, 4-channel, transparent background, isolated subject, alpha matte, no background, cutout PNG, alpha=0 outside the object, transparent canvas, not white, not checkerboard, production RGBA layer for compositing. Five more are the Chinese equivalents. Nothing in the README explains why combining two breaks it, only that you should not.
The second thing in that repo is a prompt rewriter, and it is the part most people will skip and should not. The published system prompt turns a short caption into a Figma-style JSON description: layers ordered back to front, normalized coordinates in one fixed string format, hex colors per layer, and every piece of visible copy quoted character for character and owned by exactly one layer. You run that rewrite outside the model, then feed the result in. It is the clearest published example yet of treating a prompt as a layout document rather than a sentence.
Video
SparkWan checkpoints from Alibaba landed yesterday, September 24 (HF createdAt 2026-09-24T08:18:40Z), across both Wan2.1 and Wan2.2 at 480P and 720P. Licence is Apache 2.0, base model Wan-AI/Wan2.2-T2V-A14B. The pitch is four inference steps with 90% to 95% of attention discarded, using a learned sparse attention method the authors call RoLA plus a distillation method they call CrossDistill. For a creator the consequence is iteration speed: four steps means you see a draft while you still remember what you asked for.
Read the headline number carefully. The cited paper promises "up to 265x single-GPU acceleration," which is the authors' own framing in their own abstract, and it does not describe this checkpoint. That figure is for three-step inference on Wan2.1 at 720P on an RTX 5090. The repo you would actually download is Wan2.2 at 480P at four steps. The card gives you the checkpoints and the sparsity range and nothing about wall-clock time on any card. Project page, GitHub at 139 stars, paper. No hosted demo surfaced.
On the research side, WanPE was published September 24 and submitted to HF papers this morning. It is a 397B prompt-enhancement model whose whole job is turning what you typed into a shot-level cinematic plan before the video model ever runs. The authors report human preference gains of 10.66 to 18.84 points at 5 to 15 seconds and 50.86 points at 30 seconds, all self-reported, all against their own Wan3.0 generator. No weights, no demo, and a 397B prompt writer is not something you will run. The idea is the takeaway: at 30 seconds the prompt is doing more work than the model is, which is the same lesson the Ming rewriter teaches in a different medium.
Audio and music
NVIDIA Nemotron 3 Diarization got its full write-up on September 23 and its weights refreshed on September 24. The repo itself went up September 1 (createdAt 2026-09-01T13:16:57Z), so this is a documentation moment rather than a launch, and it deserves one. The model answers exactly one question: who spoke when. It takes 16 kHz mono audio and returns time ranges tagged speaker_0, speaker_1 and so on, up to eight speakers, and it handles people talking over each other by letting two channels be active in the same frame.
For anyone who cuts interviews, podcasts or panel footage, that is the labour you currently do by scrubbing. The model is 99.2M parameters, with a 107 MB q8_0 GGUF file sitting in the repo alongside the safetensors, so this is a small download, not a weekend. Licence is OpenMDW 1.1. NVIDIA reports 14.72% diarization error rate and a first-place finish on VoiceArena's initial Diarization-Bench against 19.3% for the next system, and is careful to note those initial results may change as that benchmark completes its version 1 analysis.
Two honest catches, both stated by NVIDIA. The labels are anonymous channels, not identities: it will tell you speaker_2 talked from 1.8 to 3.25 seconds, and it will not tell you that speaker_2 is Maria. And the maximum is eight, so a press conference or a crowded table will drop or misassign voices. The live demo Space takes an uploaded file, a live mic, or a prepared conversation, and is the fastest way to judge it on your own audio.
Alongside it, Audio8-ASR-Infinite (createdAt 2026-09-21, Apache 2.0, 4B, English and Chinese) is a streaming transcription model with no fixed length limit and a free demo Space. Pair it with the diarization model and you have the two halves of a speaker-labelled transcript, both open, both free to use commercially.
Open and local
The local story today is that the useful releases got smaller in two different directions at once. SparkWan shrank the compute by throwing away most of the attention and most of the steps. Nemotron shrank the model to 99M parameters with a 107 MB quantized file. Ming-Image did neither and still needs 80 GB, which is the honest counterweight: design-quality type rendering is expensive, and the community MLX conversions arrived within six days precisely because nobody wants to rent a card to try a model.
- inclusionAI/Ming-Image: the code, the ten transparency phrases, and both published prompt rewriters for design generation and layer decomposition. 93 stars, which means early (shields.io, checked today).
- Comfy-Org/Ming-Image: ComfyUI's own packaging of both Ming models including layer-specific weight files, created September 24 (createdAt 2026-09-24T07:13:14Z). The pull request wiring it into ComfyUI is open and not merged, so the files are there and the graph is not. Two community Spaces are already pulling from it.
- AlibabaResearch/SparkDiffusion: four-step, heavily sparse Wan2.1 and Wan2.2 checkpoints under Apache 2.0. 139 stars.
- NVIDIA-NeMo/Speech: the toolkit that runs the diarization model, plus the end-to-end evaluation script. 19k stars.
- vpakarinen/better-human-motion-h3-lora: an Apache 2.0 add-on for MiniMax-H3 aimed at human movement, created September 2 and still climbing, with a free demo Space. Its sibling natural-face-speech-h3-lora targets talking faces.
- JOKER141/BUNNY_H3_Conditioning_Bridge: an action-continuity adapter for MiniMax-H3 with an
Example Workflow.jsonin the repo, created September 13. Licence is the MiniMax-H3 community agreement, not an open one, so check it before commercial use. - alibabagroup on Hugging Face: the full SparkWan set, six variants across 2.1 and 2.2 at 480P and 720P, all uploaded September 24 in one batch.
Creative workflows
1. Take a flat design apart into editable layers. Repo: inclusionAI/Ming-Image. Model: inclusionAI/Ming-Image-0.1-Design-Layer. Sample files: assets/layer_samples/card_making_input.png and assets/layer_samples/card_making_prompt.txt.
The steps. Write a rough layer plan in plain words, front to back: text on top, the card behind the text, the product, the background. Run that rough plan plus the image through the published GUIDED_PROMPT rewriter using an instruction-following vision model (the authors name Ling-3.0-flash-VL or qwen3.8-27B). The rewriter returns a numbered specification ending in the exact line the tool parses. Then run infer.py --task layer-decompose --input-image your.png --prompt spec.txt --resolution 1024. You get layer_01.png, layer_02.png and so on as RGBA files.
How it works. The rewrite step is doing the real work. It forces one rule that decides whether the output is usable: a card, panel, badge or banner sitting behind text becomes its own layer, never merged into the text and never merged into the background. The background is always the last layer and it swallows tables, floors, shadows and gradients. That single rule is what makes the result look like a design file instead of a bad cutout.
Why it is good. Layer counts are specified rather than guessed, so the output slots into a deck or a mockup without a reconciliation pass. Aspect ratio is preserved. MIT licence on the model.
Where it breaks. The tool always returns one extra composite image first, and the script silently skips it, so if you call the model directly rather than through infer.py your layer numbering will be off by one. Resolution is bucketed to 512 or 1024 only, so fine type in a large poster will soften. And the validated hardware is one 80 GB card. The hosted Space is the realistic route for most people this week.
2. Build a speaker-labelled transcript from a two-model pipeline. Guide: ASR_INTEGRATION_GUIDE.md in the model repo.
The steps. Run SortformerEncLabelModel.from_pretrained("nvidia/Nemotron-3-Diarization") over your 16 kHz mono file to get speaker segments. Separately run a transcription model that returns word timestamps (nvidia/parakeet-tdt-0.6b-v3 is NVIDIA's example). Then assign each word to whichever speaker was active at the word's midpoint.
How it works. The two models never talk to each other. One produces time ranges with speaker tags, the other produces words with times, and a few lines of Python join them on the shared clock.
Why it is good. You can swap either half. If you already have transcripts from your editing suite with timecode, you only need the diarization half.
Where it breaks. NVIDIA says this plainly and you should believe them: the midpoint rule is a heuristic. When two people talk at once it marks the word overlap/ambiguous rather than guessing, and words outside detected speech come back unassigned. Cross-talk at the top of a panel discussion is exactly where you will spend your cleanup time. There are four latency presets (30.4s, 1.04s, 0.64s, 0.32s) and the offline 30.4-second one is both the most accurate and the fastest to process, so use it unless you genuinely need live output.
Worth testing
- Ming-Image-0.1-Design-Layer Space. Upload one flat design you already own and see whether the layers come back clean. Tradeoff: output is capped at the 512 or 1024 bucket, so this will not replace a source file for print.
- Nemotron Diarization Space. Upload fifteen minutes of your messiest interview audio. Tradeoff: eight speakers maximum, and anonymous channels you still have to name yourself.
- Audio8-ASR-Infinite Space. Streaming transcription with no length cap. Tradeoff: the card lists English and Chinese only, so other languages are untested here.
- better-human-motion-h3-lora Space. A free way to judge whether the community motion add-ons for MiniMax-H3 fix the limbs. Tradeoff: one author, one checkpoint at step 500, no published comparison against the base model.
- The ten transparency phrases from the Ming repo, tried verbatim in whatever image model you already pay for. Tradeoff: they were tuned for this model, so somewhere between one and all ten will do nothing anywhere else. That is a fifteen-minute experiment with a genuinely useful answer.
What actually matters from today's signal
The Qwen transparency story ran all week and ended in the same place every time: great output, research licence, no clean commercial path. Ming-Image walks in six days later with alpha-channel generation, layer decomposition, a published prompt method, hosted demos for both models, and MIT on the file. That is the whole argument. When two models do roughly the same job, the licence is not a footnote at the bottom of the post. It is the feature.
What the day is really telling you is that decomposition has become the interesting half of the field. Making a new picture is solved enough to be boring. Taking a finished picture back apart into the pieces a designer actually edits, or taking a finished recording apart into who said what, is the work that saves real hours, and it is the work that nobody had shipped openly until now. Watch which of your own finished files suddenly become inputs.
The counter-signal is hardware, and it is not small. Eighty gigabytes of memory as the validated floor puts the design models out of reach of almost everyone reading this, and the MLX conversions that appeared within a week carry no quality claim at all. The realistic move for the next fortnight is the hosted Space: prove the output is worth the pipeline change before you rent anything. The diarization model is the exception. At 99M parameters with a 107 MB quantized file, that one runs anywhere, and it is the item on this page most likely to still be in your workflow in six months.
Source access notes: Adversarial fact-check pass run by a subagent against primary sources on this draft; two corrections folded in before save: a WanPE submission date that said "yesterday" where the HF papers page says submitted today, and the SparkDiffusion "265x" figure, which the pass established describes three-step Wan2.1 at 720P on an RTX 5090 rather than the Wan2.2 480P checkpoint covered here. That mismatch is now stated inline rather than hedged. Everything else in the draft survived that pass unchanged. A later scoped pass over today's articles surfaced a third miss that has been added back here: Comfy-Org/Ming-Image, created September 24, which the first pass did not look for because the draft never claimed anything about ComfyUI packaging. All Hugging Face dates taken from API createdAt with cache busting, never listing "Updated" timestamps; the Ming-Image, Nemotron and better-human-motion repos all show lastModified values days after createdAt and are described accordingly. GitHub star counts from shields.io with cache busting, checked 2026-09-25. Sandbox shell has no outbound network, so every fetch in this run went through the sanctioned fetch tool; no curl or wget fallback was attempted. Blocked or unusable this run: blog.comfy.org (JavaScript wall, no content served), api.github.com ComfyUI releases endpoint (returned empty), midjourney.com/updates and civitai.com (not reachable, consistent with prior runs). Quiet this run with nothing in the last 48 hours: OpenAI news, Runway research, Black Forest Labs, ElevenLabs, Adobe (last creative post was the Topaz Labs close on September 23, covered yesterday), Replicate blog, fal blog, Stability, Luma, Suno. Writing section omitted: nothing in the last 48 hours affected content creation rather than general language models. The SparkDiffusion "265x" figure and the WanPE preference gains are both the respective authors' own reported numbers and are labelled as such inline; the Artificial Analysis leaderboard image in the Ming-Image repo is likewise the authors' own asset.