Overview
Ming-Image-0.1-Design is a visual design model series open-sourced by Ant Bailing on September 23, 2026. It includes two 6B-parameter models, respectively aimed at end-to-end design generation and editable layer decomposition. The series features structured prompt control, native RGBA alpha channel generation, and semantic layer output as its core highlights, and simultaneously open-sources two Agent Skills, covering the complete workflow from prompts to code validation, and from screenshots to editable PPT. The code is hosted on GitHub, model weights are available on ModelScope and Hugging Face, with additional ComfyUI adaptation, and a two-week free API call period is open on OpenRouter.
Key Features
- Dual 6B Models with Division of Labor and Collaboration: Ming-Image-0.1-Design is responsible for text-to-design, and Ming-Image-0.1-Design-Layer is responsible for layer decomposition. The two 6B-parameter models respectively cover the two stages of design generation and design asset structuring.
- Fine-Grained Structured Control: Supports input of structured prompts up to 8K tokens, enabling fine-grained control over layout, typography, color scheme, and image references. It can generate complete visual designs such as UIs, dashboards, infographics, and posters end-to-end in one pass, with unified color schemes, composition, and asset styles.
- Native RGBA Alpha Channel Generation: Introduces a native RGBA VAE, which can directly generate assets such as people, products, icons, and decorations with alpha channels, without the need for post-processing cutouts.
- Semantic Layer Decomposition: Input any flattened design image and output 2 to 9 semantic RGBA layers (such as title, card, subject, background). Each layer can be independently moved, replaced, and recolored.
- Open-Source Agent Skill Workflows: Simultaneously open-sources Ling UI Design Skill and Image-to-Editable-PPT Skill. The former implements a full-chain visual coding workflow from prompts or screenshots to design, assets, code, and then browser validation; the latter converts slide screenshots into native editable PowerPoint files.
Use Cases
- Quickly generate complete visual design drafts such as UI interfaces, dashboards, infographics, and posters
- Directly produce assets with alpha channels such as people, products, icons, and decorations, eliminating the cutout step
- Decompose flattened design images into independently editable semantic layers for easy secondary modification
- Convert slide screenshots into native PowerPoint files with editable text, shapes, colors, and layout
- Use Ling UI Design Skill to complete the full-chain workflow from prompts or screenshots to design, assets, code, and browser validation
Pros
- Ranked first among open-source models in the UI/UX Design specialized evaluation updated by Artificial Analysis on September 18, 2026, with an Elo score of 1082
- Layout win rate 67.4%, complex composition 67.0%, text rendering 66.7%, with outstanding performance in multiple sub-metrics
- Layer decomposition achieved the best results in all 12 evaluations of the Crello test set, and is 4.3 times faster than the 20B Qwen baseline
- Native RGBA VAE supports directly generating assets with alpha channels, reducing post-processing steps
- Code, model weights, and ComfyUI adaptation are simultaneously open, and a two-week free API call period is provided on OpenRouter
Pricing
The official pricing details have not been announced. Model weights are available on ModelScope and Hugging Face, the code is open-sourced on GitHub, with additional ComfyUI adaptation, and a two-week free API call period is open on OpenRouter; subsequent commercial use or API billing methods are subject to the official website.
Summary
Ming-Image-0.1-Design uses two 6B models to respectively cover design generation and layer decomposition. With structured prompt control, native RGBA alpha channel generation, and semantic layer output, it ranks first in open-source UI/UX design evaluations, and its layer decomposition achieves the best results in all 12 evaluations of the Crello test set. The official team also states that its output stability is still insufficient in scenarios such as complex hand movements, continuous operation steps, and fine shadows and reflections, and layer decomposition may produce edge residue or missing layers when there are many overlaid light effects, special transparent materials, and complex occlusions.
Version History
- Ant Bailing open-sources the Ming-Image-0.1-Design series (2026-09-23): Two 6B models: Design generates UI, infographics and posters end to end from structured prompts of up to 8K tokens with native transparency support, while Layer decomposes flattened design images into 2 to 9 independently editable RGBA layers. The Ling UI Design Skill and Image-to-Editable-PPT Skill were open-sourced alongside.