1. Workflow Overview

This is an advanced Stable Diffusion 3 (Flux architecture) workflow specialized for anime-style (Re-style) generation, featuring:

• Flux-DiT model replacing traditional UNet

• ControlNet Upscaler for detail control

• T5-XXL + CLIP-L dual-text encoder system

Key Models:

• NunchakuFluxDiTLoader: Flux-architecture SD3 model

• Flux-ControlNet-Upscaler: Super-resolution control

• T5-XXL + CLIP-L: Enhanced text understanding

2. Critical Nodes

Dependencies:

1. Model Files:

   • svdq-int4-flux.1-dev → Save to models/fluxdit/

   • Flux.1-dev-Controlnet-Upscaler.safetensors → Save to models/controlnet/

   • t5xxl_fp16.safetensors → Save to models/clip/

3. Workflow Structure

Group 1: Model Loading

• Nodes: NunchakuFluxDiTLoader, DualCLIPLoader, VAELoader

Group 2: ControlNet Processing

• Input: Reference image (ComfyUI_temp_pecyg_00001_...png)

• Params: ControlNet strength=0.6, start=0%, end=54.56%

Group 3: Generation Core

• Key Params: Resolution=768x1024, steps=28, seed=526841747880726

4. Inputs & Outputs

Required Inputs:

• Positive/Negative prompts (via two CLIPTextEncode nodes)

• Reference image (auto-loaded)

Output:

• Final image saved to ComfyUI folder

5. Important Notes

⚠️ Hardware:

• Minimum 12GB VRAM (T5-XXL + Flux model)

• Launch with --medvram flag

🔧 Troubleshooting:

1. If Nunchaku plugin missing:

   cd custom_nodes && git clone https://github.com/T8star/ComfyUI-nunchaku

2. ControlNet strength >0.7 may cause artifacts