inclusionAI/Ming-Image-0.1-Design-Layer

This open-source model decomposes flattened design images into separate RGBA layers. Given a visual input and a specific plan, the system outputs standalone transparent graphic files. The included configuration supports a default request for six distinct layers. Performance metrics from the Crello test set suggest effective separation of visual elements. Users install dependencies via the companion repository and run a Python script for local execution. A working resolution of 1024 pixels is recommended, or 512 for faster processing. Sampling steps should be set to 12 with a CFG scale of 2.0. The system requires one CUDA GPU with 80 GiB of VRAM according to the validated configuration. Independent verification of the reported layer quality metrics is currently unavailable. Precision runs in BF16 mode, and the output preserves the input image aspect ratio. The MIT License permits broad use of the released assets. Details on prompt rewriting utilize specific language model options for enhancement.