1. Workflow Overview

Purpose:
Generates production-ready chibi figure designs with facial feature control (PulidFlux) and 4x super-resolution, optimized for 3D printing/packaging.

Core Models:

• ReVAnimated_v168 (Anime-style base model)

• PulidFlux_v0.9.1 (Precise facial feature binding)

• 4x_NMKD-Siax_200k (Upscaling)

2. Key Nodes

Dependencies:

• PulidFlux model file (pulid_flux_v0.9.1.safetensors) in ComfyUI/models/pulid_flux

• insightface library with CUDA support

3. Workflow Structure

Group 1: Model Parameters

• Inputs: Prompt (e.g., "programmer CODELIE"), negative prompt

• Key Nodes:

  • DualCLIPLoader: Dual text encoder (CLIP + T5)

  • EmptyLatentImage: 768x1024 canvas

Group 2: Pulid Control

• Input: Reference face image (e.g., Einstein.png)

• Process:

  1. FaceBoundingBox detects face area

  2. ApplyPulidFlux merges facial features

Group 3: Upscaling

• Flow: Base image → 4x upscale → 50% downscale → refine

4. Inputs & Outputs

Required Inputs:

• Positive prompt (detailed figure description)

• Reference face image (.png/.jpg)

• Seed (default -1 for auto)

Output:

• Final PNG with metadata

• Resolution: 1536x2048 → downscaled to 768x1024

5. Notes

⚠️ Requirements:

• GPU VRAM ≥8GB (PulidFlux needs CUDA)

• Python 3.10+ recommended

💡 Optimization:

• Adjust strength in ApplyPulidFlux (default 0.95) if face fusion is imperfect

• Replace upscale model with 4x-UltraSharp for sharper details

🔧 Troubleshooting:

• PulidFlux errors: Verify model file path

• Face detection fails: Ensure clear frontal face in input image