Ming Image 0.1 turns flat designs into editable layers
inclusionAI just open-sourced Ming-Image-0.1, a 6B model series purpose-built for visual design like UIs and posters. It natively outputs up to 2K graphics with transparent alpha channels, and a companion model can break flattened designs back into editable layers. Why it matters: It directly connects generative AI with real graphic design pipelines.
Ming-Image-0.1 is an open-source 6B visual design model series from inclusionAI, built specifically for UI screens, infographics, posters, and text-heavy graphic compositions.
Most generative image models output a single flattened canvas where text and graphics are fused together forever. Ming-Image-0.1 splits the workflow into two dedicated 6B models: Ming-Image-0.1-Design handles direct generation, while Ming-Image-0.1-Design-Layer decomposes existing flat graphics back into separate, editable transparent layers.
What it does
The primary generator supports up to 2K resolution across native aspect-ratio buckets like 16:9, 4:3, and 1:1, accepting custom ratios between 1:4 and 4:1.
Crucially, it natively renders 4-channel RGBA images with clean alpha transparency. Prefixing a prompt with standard phrases like "RGBA, 4-channel, transparent background" or "cutout PNG, alpha=0 outside the object" yields isolated design elements and assets ready for immediate compositing without third-party background removers.
The companion model, Ming-Image-0.1-Design-Layer, takes a flattened graphic and slices it into distinct RGBA PNG layers based on layer count and specifications. It puts text verbatim on the topmost layer, cards and panels directly behind text, the hero subject in the middle, and environmental textures at the bottom.
inclusionAI also built downstream agent workflows around it: Ling UI Design helps code UI components from generated mockups, and Image to Editable PPT translates flat slide images into native PowerPoint shapes and text boxes.
How it runs and specs
To keep layouts sharp, the pipeline relies on Figma-style structured JSON captions. Users run an instruction-following VLM like Ling-3.0-flash-VL or qwen3.8-27B to expand natural-language requests into normalized bounding box coordinates, layer hierarchies, and exact hex colors before hitting the diffusion transformer.
On the Artificial Analysis UI/UX open-weights leaderboard, Ming Image Design currently holds the top spot. Everything is released under the permissive MIT License.
The base Design model weighs around 12GB, while the Design Layer model sits around 24GB. The official validated baseline requires an 80 GiB GPU running in BF16, with optional FlashAttention 2 support for faster inference.
Code, weights, and demos are available on GitHub, Hugging Face, and ModelScope.
Why it matters
Generative design has long suffered from the "flattened JPEG" dead end: you get an impressive picture, but you cannot fix a typo or nudge a button. Ming Image bridges that gap by treating design as layers and structured geometry rather than pixel soup, giving creators transparent assets and editable slices that slot straight into Figma or presentation software.
The catch is hardware overhead. An 80 GiB GPU requirement rules out casual local tinkering on consumer laptops. Still, for developers and design-tool builders assembling automated UI or marketing pipelines, it is easily one of the most practical open-weight image releases available.
Sources (1)
- github.com(opens in a new tab)Originalgithub.com