1. Workflow Overview

This workflow is designed for fashion photography and e-commerce model generation, combining:

• Pose Control: Via reference images using ControlNet.

• FLUX.1 Model: High-resolution photorealistic outputs.

• Multi-LoRA Fusion: Sportswear and Asian portrait enhancements.

• Auto-Translation: Chinese-to-English prompts (Baidu Translate API).

2. Core Models

• FLUX.1 (黑暗森林工作室-FLUX.1-dev-fp8): Optimized for professional photography.

• ControlNet Pose (FLUX.1-dev-ControlNet-Union-Pro-InstantX): Precise body keypoints.

• LoRAs:

  • F.1 Sport Model (weight=0.7): Enhances sportswear texture.

  • 极氪白白酱F.1 Portrait (weight=0.4): Improves Asian skin details.

• Depth Preprocessor (depth_anything_v2_vitl.pth): Extracts scene depth.

3. Node Details

Key Nodes

• DWPreprocessor: Detects pose keypoints (uses YOLOX and DW models).

• ControlNetApplySD3: Applies pose control (strength=0.8).

• FluxGuidance (strength=3.5): FLUX-specific detail enhancement.

• BaiduTranslateNode: Translates Chinese prompts to English.

Installation

• FLUX Models: Manual download to ComfyUI/models.

• DWPreprocessor: Install comfyui_controlnet_aux via Manager.

• Baidu Translate: Requires API key setup.

Dependencies

• ControlNet Models: Download from HuggingFace.

• Pose Models: Auto-downloaded.

4. Workflow Structure

1. Input Group: LoadImage, BaiduTranslateNode.

2. Model Loading: UNETLoader, LoraLoader, VAELoader.

3. Pose Processing: DWPreprocessor, ControlNetApplySD3.

4. Generation: KSampler (DEIS, 30 steps), VAEDecode.

5. Inputs & Outputs

• Inputs: Pose image (1024x1536), Chinese prompt (e.g., "white sportswear").

• Output: Preview with solid-color background (e.g., blue sky).

6. Notes

1. Errors:

   • Missing FLUX files break UNETLoader.

   • Blurry pose images may fail DWPreprocessor.

2. Optimization:

   • Use FP8 precision (fp8_e4m3fn) for lower VRAM.

3. Compatibility: Requires SD3-supported ComfyUI version.