An Open Model That Makes Editable, Layered Designs

Most image models a design team has tried this year produce the same thing: a flat picture. It may be a good picture, but the moment a designer wants to move the headline, swap the product shot or change the background, the picture has to be rebuilt or traced. Two open-weight models from inclusionAI, published on Hugging Face on September 17, 2026 and updated on September 23, go at that problem from both ends. One generates design images with transparent backgrounds. The other takes a flat design and splits it into editable layers.

Both are released under the MIT licence, which permits commercial use, and both were validated by the authors on a single 80 GiB GPU. That last fact is the price of admission, and this post is honest about it. What follows is what the two model cards say, why editable layers change the hand-off between a model and a design tool, the settings and hardware the cards specify, a short path to trying them, and where they sit next to the closed image models we track. No quality claim in this post goes beyond the cards' own.

The two Design and Design-Layer cards share a GitHub repository for installation and inference, demo spaces (a separate one per model), and links to a Design Skill and a PPT Skill in inclusionAI's ling-cookbook. The Design card also shows a UI and UX design leaderboard image. It is the vendor's image with no numbers in the text, so we reproduce neither the ranking nor a claim about it. The Hugging Face pipeline tags are text-to-image for the first model and image-text-to-image for the second, which is the clearest one-line description of the pair.

A design tool works in layers. A headline is a text object, a product shot is an image object with its own mask, a background is a fill. A flat raster from an image model collapses all of that into pixels, and the hand-off from model to designer becomes tracing: rebuild the layout by hand so that it can be changed. The better the flat image, the more it costs to throw it away when the client wants the price moved.

Transparent output addresses half of this. A generated element with an alpha channel drops into an existing layout as its own layer, so the model can be asked for a component rather than a finished page. Layer decomposition addresses the other half: a finished flat design, whether generated or inherited, comes back as separable parts a designer can edit. Together they make the model a participant in a layered workflow instead of a source of reference pictures. That is a different job from the one closed models such as Google's Nano Banana family are sold for, which we describe in our post on Nano Banana 2 Lite.

What the cards do not tell you is how clean the layers are. A decomposition that splits a card into background, image, headline and button is useful. One that leaves halos on the cut edges or merges two text blocks into one layer is a starting point that still needs hand work. The gallery shows recomposed results on a public dataset; it reports RGB L1 and Alpha soft IoU on the Crello test set as an image only, with no edge-halo measure, and we have not measured one on client designs.

The hardware line deserves plain language. An 80 GiB accelerator is a data-centre part. A design team will not run this on a laptop or a typical workstation card, and the cards give no quantised or reduced-memory configuration. The practical options are a rented GPU by the hour for a batch of jobs, or a hosted endpoint. Two aggregator routes for the models appeared on September 22 and 23, which we observed on September 25 at 08:01 UTC; both showed zero catalog prices, and a zero in a catalog is a missing price, not a free tier. Check the serving provider's own terms before sending client work through either route.

September's image-model news has mostly been about closed models and their calendars: Google's Gemini 2.5 Flash Image retirement on October 2, its successors priced per image, and the aggregator catalog adding three image-output routes between September 22 and 23, of which these two are the open-weight pair. Our September release tracker carries the dated list.

The Ming pair is not a competitor to those models on the axis they compete on. Nano Banana 2 produces a flat image at up to 4K for a per-image price; Ming-Image-0.1-Design produces a transparent design element at 2K on hardware you rent. What the pair offers that none of the closed models on our list offers is the second model: a way back from a flat design to layers, under a licence that lets an agency build it into a client pipeline without a usage agreement. Whether the layers are good enough is the trial in section 04. Our content engine practice runs that trial on a client's own design archive before recommending any generation model.

The pair is useful for a specific job, and the sections above exist because a trial can be run against a different job and read as a verdict on this one. The illustrative trials below each measure something the cards do not promise or skip something they state. They are design examples, not studios we have worked with.

Each trial would have produced a verdict about a job the pair does not do, or skipped the check the cards cannot do for you. The record in the next section fixes the job first and keeps the evidence against it.

The trial section gives a path of ten minutes, a day and a set of components, and a record per trial holds what each step found. The job: components with transparent backgrounds, a flat design split into layers, or both, written as one sentence before any prompt. The designs: the three flat designs used in the demo and the archive sample run on rented hardware, with a written layer plan for each. The settings: resolution, steps, guidance scale and precision as the cards specify them, with any departure recorded. The hardware: the rented GPU, its hourly rate and the hours used, since the cards validate on an 80 GiB part and state no minimum below it. The results: the layer count returned against the plan, the cut edges inspected at 200%, merged text blocks, and the hand work each design still needed. The licence check: MIT on both cards, with the serving provider’s terms read where a hosted route is used. And the decision, with the GPU hours beside the designer time they replaced.

An illustrative record for a card-design archive reads: job, splitting inherited flat cards into editable layers; designs, a simple card, a dense infographic and a page with overlapping text in the demo, then fifty archive cards on a rented GPU with plans naming background, image, headline and button; settings, 1024 working resolution, twelve steps, guidance 2.0, BF16; hardware, one 80 GiB accelerator for a day; results, the planned layer count on most cards, visible halos on some cut edges, two text blocks merged on the dense infographics, and a minority of designs still needing hand work; decision, continue for the simple and medium cards, keep tracing for the dense ones.

The record is also what makes the trial repeatable when the cards change. Both cards were updated within a week of being published, and a later revision is a different experiment; the record is where the settings and results of the first one stay legible beside the second.