What it does After Stable Diffusion generates an image, this extension detects faces, crops them to 512×512 (or 1024 for SDXL), runs img2img to rebuild each face, then pastes them back and inpaints the seams. It can also retouch existing images via the img2img tab with denoising set to zero.

The interesting bit The per-face prompting system: separate prompts with || to give individual instructions to multiple faces left-to-right, and use @@ to splice in the main prompt. It’s a rare bit of granular control in a space that usually treats “group photo” as a monolithic gamble.

Key highlights

• Workflow Editor swaps out face detectors, processors, and mask generators via JSON definitions
• Tilt correction threshold auto-straightens wonky angles
• “Use minimal area” mode prevents face collapse when subjects are crowded together
• API support via alwayson_scripts with keyword arguments to reduce breakage on updates
• SD.Next compatibility via a dedicated branch

Caveats

• README warns that contour artifacts are common when faces aren’t straight-on; fixing this is manual parameter twiddling
• The workflow editor UI description cuts off mid-sentence in the docs

Verdict Worth a look if you’re generating character art or portraits and tired of the “five-fingered face” lottery. Skip it if you’re already using a pipeline with built-in face restoration or don’t run AUTOMATIC1111/SD.Next.