1. Workflow Overview

Purpose:
This workflow generates multi-view consistent AI model images by:

• Using ControlNet for pose control & Flux for detail enhancement

• Generating 4 coherent views (front/side/back) from an input pose map

• Utilizing TTP Tile technology for high-resolution generation

Core Models:

2. Key Components & Installation

Required Nodes:

• ControlNetApplySD3: Applies ControlNet constraints (requires FLUX.1-dev-Controlnet-Depth model)

• FluxGuidance: Ensures character consistency (install Flux plugin)

• TTP_Tile: Processes large images in tiles (install Tiled Diffusion via ComfyUI Manager)

• Florence2Run: Generates image captions (download HuggingFace model)

Dependencies:

• Lora: 苏-FLUX小红书极致真实_v1.0 (place in models/loras)

• ControlNet Model: FLUX.1-dev-Controlnet-Depth-InstantX.safetensors

3. Workflow Structure

Group 1: Input Control

• Nodes: LoadImage (pose map), easy positive (prompts)

• Inputs: Pose map (e.g., skeleton image), character description prompts

• Outputs: Encoded conditioning vectors

Group 2: Image Generation

• Nodes: KSampler + ControlNetApplySD3 + FluxGuidance

• Logic: Generates latent images with pose consistency via Flux

Group 3: Tiled Processing (TTP Tile)

• Nodes: TTP_Image_Tile_Batch → SamplerCustomAdvanced → TTP_Image_Assy

• Function: Splits high-res images into tiles for VRAM efficiency

Group 4: Post-Processing

• Nodes: ImageCrop+ (view cropping), Image Overlay (multi-view merge), SaveImage

• Output: Final PNG with 4 aligned model views

4. Inputs & Outputs

Input Parameters:

• Required: Pose map (e.g., POSE2.png), positive prompts

• Optional: Seed value, resolution (default 1152x896), ControlNet strength (0.6)

Output:

• A single PNG with 4 model views (left/front/back/right), saved to ComfyUI/output

5. Notes

• VRAM: ≥12GB GPU recommended (reduce tile size for lower usage)

• Troubleshooting:

  • Missing ControlNet model → Download to models/controlnet

  • Flux plugin not found → Install via ComfyUI Manager

• Optimization:

  • Lower TTP tile size (e.g., 512x512) for better performance

  • Use easy cleanGpuUsed to free VRAM manually