Ming-Image-0.1-Design-Layer Pulls a Flat Design Back Into Editable Layers
Two models shipped in this release. Almost every writeup covered the wrong one.
There are two 6-billion-parameter models in the Ming-Image 0.1 Design release, and they do opposite jobs. The first one makes a poster. The second one takes a poster you already have and hands you back the pieces: the headline as its own transparent PNG, the badge behind the headline as its own PNG, the product shot on its own, the background with the shadows and the table baked into it.
Coverage has mostly been about the first model, because a new image generator is a familiar thing to write about. The second one is the one that would actually change a Tuesday afternoon. If you have ever been handed a flat JPEG of a design and asked to "just move the headline up a bit," you know exactly how much of your job that sentence represents.
Both models are on Hugging Face, both went up on September 17, and both carry the MIT licence. That last part is the whole argument, and I will come back to it.
The thing you cannot currently buy
Take apart what a cutout tool does today. You point it at a photo, it finds the subject, it gives you the subject with a transparent background. One object, one pass, and it works because "subject" is a well-defined thing in a photograph.
A design is not a photograph. A design is a stack of decisions that got flattened. There is type sitting on a card, sitting on a gradient, with a product photograph in front of all three and a drop shadow underneath. No cutout tool separates that, because there is no single subject to find. You reopen the source file, or you rebuild it by hand.
Ming-Image-0.1-Design-Layer separates it. You give it the flat image and a written plan describing what the layers are, front to back, and it returns that many transparent PNGs. The demo the authors ship in the repository is a greeting card pulled into six layers, with the input, the layers and a restacked version all published side by side. Judge that restacked version with your own eyes before you trust it, because a gallery image is the authors showing you their best case.
I want to be careful here. The model is not reading the design and working out what the layers were. You tell it what the layers are. That distinction is the whole craft of using it, and it is also the reason the results are more reliable than a tool that guesses.
How the written plan does the work
The pipeline the authors published has two steps, and the second step is the one most people will skip.
Step one: you write a rough plan in plain words. Headline on top. Card behind the headline. Product. Background.
Step two: you run that rough plan, plus your image, through a rewriter that turns it into a precise specification. The authors publish the exact instructions for that rewriter in the code repository, and reading them is worth ten minutes even if you never run the model, because they encode what a good layer split actually is.
Three rules do most of the work.
A card, panel, badge or banner sitting behind text becomes its own layer, never merged into the text and never pushed into the background. That single rule is the difference between output you can use and output that looks like a bad selection.
Any real text in the design is quoted word for word, in its original language, and lives on a front layer.
The background is always last, and it swallows everything supporting: tables, floors, boards, plates, gradients, patterns, shadows. Those never become separate layers, which keeps you from ending up with fourteen pieces when you wanted five.
Whether you run the automated rewriter or write the specification yourself, those three rules are what you are enforcing. The tool reads the layer count straight out of your plan, so you control how many pieces come back.
The licence is the feature
For the past ten days the open transparency conversation has been about a different model, one that produces genuinely beautiful alpha-channel output and ships with a research-and-evaluation restriction sitting on top of it. Every person who made something good with it ran into the same wall: the output is lovely and the paperwork says you are evaluating.
Ming-Image-0.1-Design and Ming-Image-0.1-Design-Layer both carry MIT. It is in the card metadata and there is a LICENSE file in each repository. MIT is about as short and as permissive as software licences get, and it is the version most working designers already recognise from the tools on their machine.
When two models do roughly the same job, the licence is not a footnote at the bottom of the announcement. It is the reason one of them ends up in your process and the other one stays a screenshot you posted.
Put this into practice
Start in the browser. It costs nothing and takes five minutes.
- Open the hosted layer demo. It went up on September 23 and it is running now.
- Use one of the five supplied examples first. The set is an infographic, a milestones graphic, a PowerPoint slide, a cartoon still and a photograph of penguins, which spans most of what people will actually throw at it.
- Then upload one flat design you own. A social graphic you made six months ago is the ideal test, because you remember what the layers were and can judge the result honestly.
- Ask for fewer layers than you think you need. Four or five is a good first request. Asking for twelve is how you get twelve, several of which are nothing.
- Be patient. That demo runs on free processor-only hardware, so a queue and a wait are normal. The generator demo runs on a shared graphics card and is noticeably quicker if you want to compare.
If the output earns it, run it yourself.
The code is one repository and installing it is an ordinary Python install. The command is one line:
python infer.py \
--model inclusionAI/Ming-Image-0.1-Design-Layer \
--task layer-decompose \
--input-image your-design.png \
--prompt your-layer-spec.txt \
--resolution 1024 \
--output-dir outputs/layers
Your layer files come out as layer_01.png, layer_02.png and so on, in transparent PNG, with the original aspect ratio preserved.
Now the part nobody tells you before you spend the money: the setup the authors actually tested is one graphics card with at least 80 gigabytes of memory. That is a rented machine by the hour, not the card in your desktop. Weights alone run to roughly 65 gigabytes of download. Two community conversions for Apple silicon appeared on September 23, but neither one publishes a single quality comparison, so treat them as an experiment rather than a shortcut.
The free experiment worth running today, even if you never touch this model again: the same repository publishes ten exact phrases that switch transparent output on, five of them in 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. The instruction is to put exactly one of them at the very start of your prompt and never combine two. Try them verbatim in whatever image tool you already pay for. They were tuned for this model, so somewhere between one and all ten will do nothing anywhere else. Fifteen minutes, useful answer either way.
Where it breaks
The resolution is bucketed. Layer decomposition runs at one of two working sizes, 512 or 1024, and nothing in between. Your aspect ratio survives, but a large poster gets worked on small, so fine type comes back softer than it went in. This is a tool for getting editable pieces, not for getting print-ready ones, and anyone hoping it replaces a lost source file for a billboard is going to be unhappy.
There is an off-by-one trap waiting for anyone who calls the model directly instead of through the supplied script. The model always returns one extra image first, a flattened composite of the whole canvas. The script silently drops it. If you write your own loop and do not drop it, every layer in your output is numbered one too high, and you will not notice until something is stacked wrong.
The phrase list is Chinese-first in spirit. Five of the ten transparency phrases are Chinese and the model's home documentation is a WeChat post. The English phrases are published and presumably tested, but I would not assume they are equally strong.
The clickable version is half-built. ComfyUI's own organisation put up a packaging of both models on September 24, including files specifically for the layer model, and there is a pull request open against ComfyUI to wire it in. Open, not merged. So if you live in ComfyUI, the weights are sitting there in the shape you expect and the graph to run them is still someone else's branch. The code repository sits at 93 stars, which is the honest signal about how early all of this is.
And the hardware floor is real. Eighty gigabytes is not a rounding error on a consumer card. Most people reading this will use the hosted demo or nothing, at least this month.
What I would actually do with it
The use I keep coming back to is not rescuing lost files. It is taking someone else's finished design, the one a client loves and cannot supply the source for, and getting far enough into it to make the one change they asked for.
The second use is quieter and might matter more. The published rewriter instructions describe a design as a back-to-front stack with coordinates, colours and text ownership per layer. That is a written format for a layout. It is readable, it is diffable, and a person can edit it. We have spent three years learning to write prompts as sentences. This is the first release I have seen that treats the prompt as a layered document instead, and that idea does not need an 80-gigabyte card to be useful.
The open question I do not have an answer to: if a flat export can be reliably taken apart, what is a source file for? Version history and exact type, probably. But the assumption that a JPEG is a dead end has been load-bearing in how studios hand work over, and it is starting to wobble.
If you run this on something messier than a greeting card, a dense infographic or a screenshot of a real interface, I want to know how many layers came back usable and how many came back as a smear. That is the number that decides whether this is a tool or a demo.
Medium metadata
Suggested kicker: MIT licensed, 80GB of memory, and one written plan that decides whether the output is usable.