Traditional face restoration always felt off to me: it comes with its own fixed style baked in, so anime images end up with faces that get smudged into a realistic look after fixing. So I built a set of ComfyUI facial restoration nodes, and the core difference is just one thing — it’s not locked to any specific face-fixing model. It directly reuses whatever checkpoint, VAE, prompt, sampler, and scheduler you currently have hooked up, and regenerates each face in the style of the original image, so the style naturally matches.
Breaking the pipeline down, it’s these steps:
- Detection: find the faces in the image first, crop each one out and process it separately, so the face-fixing resolution isn’t limited by the full image.
- img2img redraw: the cropped face goes through img2img using your current checkpoint, with the same prompt, sampler, and scheduler already wired into your workflow.
- Semantic mask: only keep the face region, to avoid the redraw dragging hair edges or the background out of shape too.
- Color matching: align the redrawn face’s skin tone and lighting with the original image — skip this step and you’ll get a noticeable color mismatch when it’s pasted back.
- Seamless blend: paste the processed face back into its original spot, with the boundary blended out.
Multiple faces get processed one by one. It works with realistic, anime, illustration styles, as well as families like SDXL, Pony, and FLUX — since it was never tied to any specific model to begin with, it just uses whatever’s hooked up in your workflow.
Anyone who needs it, go try it. If you hit an issue, drop a reply and I’ll fix it when I see it.