1. Workflow Overview

This is an AI-powered clothing replacement workflow featuring:

• Automatic garment segmentation with BiRefNet

• Context-aware inpainting using Flux Fill model

• High-resolution output (default 1024x1536)

Key Models:

• BiRefNet: Human parsing model (auto-downloaded)

• Flux Fill fp8: Quantized model for clothing redraw

• ControlNet LoRA: Style control (F.1_Fashion_Model_Sportswear_1.0)

2. Critical Nodes

Dependencies:

1. Model Files:

   • Flux Fill dev - fp8_黑森林官方重绘模型_fp8_e4m3fn → Save to models/flux/

   • flux1-redux-dev.style → Save to models/style_models/

3. Workflow Structure

Phase 1: Input Processing

• Loads source image (yellow-dress.jpg)

• Generates garment mask (expanded by 50px via GrowMask)

Phase 2: Local Redraw

• Crops garment area (forced size 1024x1024)

• Applies LoRA for style control (weight=0.8)

• Prompt: "A hyper-realistic photograph, a young lady in a vibrant yellow short-sleeved dress..."

Phase 3: Composition

• Seamless blending with ImageCompositeMasked

• Post-processed with bislerp stitching

4. Inputs & Outputs

Required Inputs:

• Source person image (transparent background supported)

• Optional clothing reference (magazine.png)

Outputs:

1. Redrawn clothing only (ComfyUI/temp_...png)

2. Final composited image (ComfyUI/output_...jpg)

5. Important Notes

⚠️ Hardware:

• Minimum 8GB VRAM (FP8 model reduces usage)

• Launch with --xformers for acceleration

🔧 Troubleshooting:

1. Rough garment edges:

   • Adjust GrowMask expansion value (default=50)

   • Change InpaintCrop resize mode to bicubic

2. Style mismatch:

   • Switch LoRA model (multiple fashion LoRAs pre-loaded)