Mask-driven inpainting via RunComfy CLI, defaulting to Z-Image Turbo Inpainting when you have a binary mask and falling back to Nano Banana 2 Edit when you need to describe the region in words. The strength parameter matters: 0.3–0.5 for retouching, 0.8–1.0 for full replacement. The prompting advice is practical, blur your mask edges by a few pixels for better blending, name what stays outside the mask, and use spatial labels even when you have a mask. Routes to LoRA-augmented inpainting if you need brand-locked style, GPT Image 2 Edit for multi-ref layouts. Honest limitation: outpainting is out of scope, different skill.
npx skills add https://github.com/agentspace-so/runcomfy-agent-skills --skill image-inpainting