1. Workflow Overview

This workflow is designed to fix deformed hands in AI-generated images with low GPU memory usage. Key features:

• Hand Localization: Masks precisely mark defective hand areas.

• Local Inpainting: Combines Flux models and depth learning for hand detail repair.

• Seamless Blending: Ensures natural transitions between repaired and original areas.

Core Models:

• Base Model: 基础算法_F.1 (UNET for general image generation).

• Flux Model: Loaded via DualCLIPLoader (version t5xxl_fp8_e4m3fn).

• Auxiliary: SegmentAnythingUltra V2 (auto-segment hands).

2. Key Nodes & Installation

Dependencies:

• Flux model via DualCLIPLoader.

• Place ae.sft in models/vae.

3. Workflow Structure

Group 1: Input & Preprocessing

• Input: Original image (e.g., clipspace-mask-175661.png) and hand mask.

• Steps:

  1. Load image/mask via LoadImage.

  2. Resize with InpaintResize (e.g., 1536x1536).

  3. Remove alpha channel via ImageRemoveAlpha.

Group 2: Hand Repair

• Input: Preprocessed image, mask, prompt (e.g., "perfect hand").

• Steps:

  1. Crop hand area via InpaintCrop.

  2. Optimize conditioning with BasicGuider.

  3. Generate repaired latent image via SamplerCustomAdvanced (Euler sampler).

Group 3: Blending & Output

• Input: Repaired hand image, original, mask.

• Steps:

  1. Blend with InpaintStitch (method: bislerp).

  2. Save result via SaveImage (e.g., flux/20240513_142301_修手.png).

4. Inputs & Outputs

Inputs:

• Image: Include hands (resolution ≥800x800).

• Mask: Accurately mark defective areas (black/white).

• Prompt: Simple target description (e.g., "perfect hand").

Output:

• Repaired PNG saved in flux/date/time_修手.png.

5. Notes

1. VRAM Optimization:

   • Optimized for ≥6GB GPUs. Reduce InpaintResize output (e.g., 1024x1024) if needed.

2. Mask Accuracy:

   • Manually refine masks if auto-segmentation fails.

3. Compatibility:

   • Ensure Flux model matches base algorithm version.

4. Debugging:

   • Adjust GrowMaskWithBlur (default: 30px expansion, 0.1 blur) for better edges.