1. Workflow Overview

This workflow is designed for batch generation of multi-view character images, ideal for LoRA training data preparation. Key stages:

1. Multi-View Generation: Creates consistent character images from OpenPose skeletons + reference photos

2. Upscaling: Enhances resolution via FLUX model

3. Local Refinement: Fixes face/hand details

4. Cropping: Splits images into standardized tiles

Core Technologies:

• SDXL + PulID Flux (identity preservation)

• ControlNet OpenPose (pose control)

• StableSR (denoising & super-resolution)

2. Core Models

3. Key Nodes

4. Workflow Structure

Group 1: Multi-View Generation

• Input: OpenPose skeleton + reference image + prompts

• Process: ControlNet for pose + PulID for consistency

• Output: 1024x1024 images

Group 2: FLUX Upscaling

• Input: Raw generated images

• Process: 1.5x upscale + detail refinement

• Output: 1536x1536 HD images

Group 3: Local Repair

• Targets:

  • Faces (detected by face_yolov8m)

  • Hands (detected by hand_yolov8s)

Group 4: Batch Cropping

• Parameters: Custom crop coordinates (adjust manually)

• Output: 640x832 standardized tiles

5. Input/Output

Input Parameters:

• Required:

  • OpenPose skeleton image

  • Character reference photo (upper-body recommended)

  • Prompt (e.g., clothing description)

• Optional:

  • ControlNet strength (0.5-0.7)

  • Seed value

Output:

• Cropped character images (PNG)

• Super-resolution comparison slider

6. Notes

1. Hardware: 12GB+ VRAM recommended. Use --medvram for low-end GPUs.

2. Critical Parameters:

   • ControlNet end time: 0.4-0.6

   • Face repair steps: ≥20

3. Troubleshooting:

   • Download pulid_flux_v0.9.0.safetensors manually if missing

   • Skeleton image resolution ≥1024x1024