Ming 分层 + GPT Image 2.5 局部改图工作流

WARDON @NeoORDON原推发布于 2026.10.11

在 X 上打开原推

The hardest part of AI image work has never been making something good. It has been changing one thing without destroying everything around it. You want to fix a title, swap a person, or redraw a button. You send the whole image back to the model and tell it “only touch this.” It still moves the background, drifts the layout, or quietly rewrites the parts you already liked. Then you spend three more rounds trying to put the pieces back. Ming-Image-0.1-Design-Layer (the 6B open-source model in the Ming series) flips the workflow. Instead of regenerating the image, it decomposes a finished design into separate transparent RGBA layers: background, text, person, logo, decorations, components. You pull out only the layer you care about, send that single layer to something precise like GPT Image 2.5, edit it cleanly, and composite it back. If the layer is still too crowded, you split it again. The post shows this as a practical loop: layer → extract → edit the piece → recombine. No more hoping the model leaves the rest alone. The Ming series also ranked in the top tier on AA’s open-source UI/UX design evals, but the author is clear that the interesting part is not generation. It is the ability to tear a design apart so you can edit it like a real file. OpenRouter already has the API (and it has been free). The thread also covers local options: ComfyUI pack, MLX builds for Apple Silicon, INT8/INT4 for NVIDIA, and GGUF. The video walks through the author’s own small tool that strings the whole process together, including secondary splits when one layer still holds too much. If local editing has been the thing that keeps breaking your workflow, this is the cleanest separation of concerns yet: Ming splits, another model retouches, you stay in control of what actually changes.

这条是引用着另一条写的：「用 Ming 拆图层 + GPT Image 2.5 做定点局部精修」