1. Workflow Overview

This workflow leverages ByteDance's InfiniteYou model for high-fidelity character migration, featuring:

• Identity Preservation: Transfers facial/attire features from reference images

• Multi-stage Control: Combines ControlNet + custom UNet for precision

• Flux Stack: Uses optimized flux1-dev-fp8 model and flux_t5xxl text encoder

2. Core Models

3. Key Nodes

4. Pipeline Stages

Stage 1: Model Loading

• Nodes: UNETLoader → DualCLIPLoader → ControlNetLoader

• Critical Models:

  • FP8-quantized UNet

  • Dual-text encoder (T5 XXL + CLIP-L)

Stage 2: Identity Transfer

• Core Node: InfiniteYouApply

• Inputs:

  • Reference image (via LoadImage)

  • ControlNet structure (weight=1.0)

  • Target scene prompt (e.g., "police officer in dark room")

Stage 3: Guided Sampling

• Tech Stack:

  • FluxGuidance (CFG=2.5)

  • BasicScheduler (30 steps, simple)

  • SamplerCustomAdvanced (Euler sampler)

5. Inputs & Outputs

• Required Inputs:

  • Reference character image (PNG/JPG)

  • Prompt describing target scene

• Outputs:

  • 1152x864 migrated image

  • Real-time preview via ailab_ImagePreview

6. Critical Notes

1. Model Dependencies:

   • Download sim_stage1_img_proj.bin to models/InfiniteYou

   • Flux models require ByteDance internal access

2. Hardware:

   • Minimum: 12GB VRAM (FP8 mode)

   • Recommended: RTX 3090+

3. Troubleshooting:

   • Tensor dtype mismatch → Verify all models are FP8

   • Lost features → Adjust ref_image weight in InfiniteYouApply (0.8-1.2)