Unlock Flawless Skin: Advanced Facial Texture Restoration Workflow

CN
ComfyUI.org
2025-05-26 08:25:43

1. Workflow Overview

mb4to57m11i6r9evxgys0a6d230b37ed696200fe4c21fbd6a4d6fc8b46974f750f59aa8688419b8bf38b.jpg

This workflow specializes in facial skin texture restoration and super-resolution enhancement, integrating:

  • SUPIR Model: Diffusion model for detail reconstruction

  • Stable Diffusion XL Lightning: Base image generation

  • ControlNet (via ReActor): Facial feature correction

  • Upscale Models: 4xFFHQDAT.safetensors (2x) + 1x-ITF-SkinDiffDetail (4x)

2. Components Breakdown

Key Nodes:

  1. SUPIR_model_loader_v2

    • Purpose: Load SUPIR-v0Q.ckpt

    • Install: Requires comfyui-supir via ComfyUI Manager

    • Dependency: Download SUPIR-v0Q.ckpt

  2. ReActorRestoreFace

    • Purpose: Face restoration with GFPGANv1.3

    • Install: reactor-node plugin

    • Model: Auto-downloads retinaface_resnet50

  3. ColorAdjust(FaceParsing)

    • Purpose: Adjust facial color balance

    • Install: comfyui_face_parsing plugin

Special Dependencies:

  • t5xxl_fp8_e4m3fn.safetensors (text encoder)

  • 1x-ITF-SkinDiffDetail-Lite-v1.pth (skin detail upscaler)

3. Workflow Structure

Group Logic:

  • Group 1: SUPIR Restoration

    • Input: 512x512 portrait

    • Process: SUPIR_encode → SUPIR_conditioner → SUPIR_sample

    • Output: Refined latent

  • Group 2: Facial Enhancement

    • Steps: ReActorRestoreFace → ImageResize+ → ColorAdjust

    • Key Params: Skin smoothness=1.05, Color saturation=1.1

  • Group 3: Super-Resolution

    • Model Chain: 4xFFHQDAT (2x) → 1x-ITF-SkinDiffDetail (4x)

    • Output: 2048x2048 HD image

4. Inputs & Outputs

Input Parameters:

  • Required:

    • Source image path (ComfyUI_01312_.png)

    • Prompt: "high quality,detailed,photography"

  • Optional:

    • Seed (default: 123)

    • Denoising strength (default: 1.5)

Outputs:

  • Format: PNG

  • Save Path: ComfyUI/output

  • Includes: Before/after comparison (via Image Comparer)

5. Notes

  • Hardware: Minimum 12GB VRAM (16GB recommended)

  • Common Errors:

    • SUPIR_VAE not found: Verify model path under models/supir

    • Face distortion: Reduce ReActor's Fidelity Scale

  • Optimization:

    • Use PurgeVRAM to free memory

    • Disable live preview for batch processing