September 24, 2026
Edit graphics in parts: two 6 billion-parameter Ming-Image models
On Sep 23, inclusionAI released two 6-billion-parameter Ming-Image models under the MIT license. Ming-Image-0.1-Design creates UI, infographics, and posters. Design-Layer breaks a flat layout into editable PNG layers with transparency.
The standard Design model creates finished graphics, while Design-Layer takes a single composite image and saves separate `layer_01.png`, `layer_02.png`, and subsequent files. Set the number of layers with the `--num-layers N` flag or describe it in a layer plan.
Local deployment. It requires Python 3.10+, CUDA PyTorch, and a GPU with 80 GiB of BF16 memory. The developer recommends 1024 px, 12 steps, and CFG 2.0; 512 px works for a quick test. OpenRouter offers a version with one reference image and a layer plan, priced as Free.
Alongside the weights, inclusionAI released two agent workflows: Ling UI Design builds and verifies UI code from references, while Image to Editable PPT rebuilds a slide as text and simple PowerPoint shapes. The standard Design model works with 1024 and 2048 px buckets, with 2048×2048 recommended.
Both models and the two workflows are already available in the inclusionAI/Ming-Image repository.
