InclusionAI's Ming-Image Tops UI Design Leaderboard

InclusionAI has launched Ming-Image-0.1-Design, a new 6-billion-parameter text-to-image model released under the permissive MIT license. Despite its relatively compact size, the model has secured the top spot among open-weight entries on the Artificial Analysis UI/UX Design leaderboard. This benchmark uses blind preference votes and Elo ratings to evaluate how well models generate user interfaces, dashboards, and other design assets, and Ming-Image's performance places it ahead of much larger 20-billion-parameter competitors.

The model is specifically optimized for text-heavy design compositions, such as app screens, infographics, posters, and navigation layouts. While general-purpose image generators frequently struggle with rendering legible typography and maintaining alignment across grids, Ming-Image-0.1-Design is built to keep small text readable and interface elements structured. It runs at a resolution of 2048x2048 pixels, requiring 12 steps and a classifier-free guidance scale of 1.0, and can be deployed on a single 80 GB CUDA GPU.

For design practitioners, one of the most significant features is the model's native support for RGBA output. By using specific prompt trigger phrases, developers can generate icons, badges, and product cutouts with transparent backgrounds. This eliminates the need for a separate background-removal step in the production pipeline, allowing generated assets to be imported directly into active design workflows. InclusionAI has made the inference code and vLLM-Omni serving recipes publicly available on GitHub to facilitate deployment.