1. Workflow Overview

• Purpose: Generate stylized images from input photos (e.g., model poses) using ControlNet for structure preservation and LoRAs for style (e.g., ethnic costumes, autumn forest themes).

• Core Models:

  • 基础算法_F.1: Base text-to-image model (likely SDXL variant).

  • FLUX.1-dev-ControlNet-Union-Pro-InstantX: Multi-ControlNet for pose/structure control.

  • Meta-Llama-3.1-8B-bnb-4bit: Image captioning model (prompt reverse engineering).

2. Key Nodes

3. Workflow Groups

1. Reference Image Group

   • Input: User-uploaded photo (e.g., lQDPKGyzHiAGKAfNB9DNBQOwJksZaqj6fsIH2j_m_4e8AA_1283_2000.jpg).

   • Process: Generates depth map via DepthAnythingV2 for ControlNet.

2. LoRA Group

   • Loads two LoRAs: 少数民族服饰_V1.0 (weight=0.2) and 秋日森林_秋天女孩_V1.0 (weight=0.7).

3. Generation Group

   • Output: 1280x2000 image after latent upscaling and VAE decoding.

4. Inputs & Outputs

• Inputs:

  • Reference image (required).

  • Resolution: Default 768x1024 (adjustable via EmptyLatentImage).

  • Negative prompt: "Imperfect, non-standard, poor quality".

• Output: Stylized model image (e.g., in ethnic costume).

5. Tips & Warnings

• ⚠️ Errors:

  • Missing ControlNet/LoRA files trigger "Missing model" errors.

  • Llama-3 requires ≥8GB VRAM; disable Joy_caption on low-end devices.

• ✅ Optimization:

  • Use fp8_e4m3fn precision to save VRAM.

  • Adjust ControlNet weight (default: 0.75) in ControlNetApplyAdvanced.