Ant Group Open-Sources Ming-Image-0.1-Design Models for UI Generation and Layer Separation
Ant Group's Bailing team has open-sourced the Ming-Image-0.1-Design series, featuring two 6-billion-parameter models dedicated to UI design generation and automated image layer separation. The main design model achieved the #1 ranking among open-source models on the Artificial Analysis UI/UX Design benchmark with an Elo score of 1082. Standard image generators struggle with structured text layout and generating isolated transparent assets, often requiring multi-step workarounds. By introducing end-to-end design synthesis alongside native RGBA layer separation, this release streamlines automated visual workflows for UI/UX designers and presentation tools. The series incorporates a native RGBA VAE to directly generate transparent elements like icons and product cutouts, supported by 8K structured prompt processing to maintain consistent typography and visual style. However, known limitations include instability when rendering complex human hand poses, fine reflections, and precise layer boundaries under heavy occlusion.
## BACKGROUND
Standard generative vision models operate on RGB color channels, making direct generation of cutouts with transparency challenging. An RGBA Variational Autoencoder (VAE) adds a dedicated alpha channel to capture transparency, allowing neural networks to natively reconstruct and synthesize elements with transparent backgrounds. Artificial Analysis is an independent benchmarking platform that evaluates AI models using metric-driven testing and human preference Elo scoring across targeted tasks.