Turbocharge Your Portrait Game: 3x Faster Generation with Nunchaku!

CN
ComfyUI.org
2025-05-07 01:08:16

1. Workflow Overview

mad8o857mhfvwyq7th3cfa889a931a434f2fb6fc7efa476637070defcc848bb19e1191cb799ecf6bc5.gif

This workflow optimizes commercial-grade portraits with:

  • 3x Faster Generation: Uses Nunchaku quantized models.

  • Skin Enhancement: Auto-fixes flaws via W_skin.safetensors LoRA.

  • Tile Processing: Splits images to avoid VRAM overload.

Core Models:

  • svdq-int4-flux.1-dev: Quantized model for speed.

  • W_skin.safetensors: Custom skin LoRA (weight=0.7).

  • 4x_NMKD-Siax_200k: Upscales to 4K.


2. Key Nodes

Node

Function

Installation

NunchakuFluxDiTLoader

Loads quantized model

Install ComfyUI-Nunchaku plugin

TTP_Image_Tile_Batch

Splits images into 1024x1024 tiles

Built-in

CR Upscale Image

4x upscaling (4x_NMKD-Siax_200k)

Manual model download required

Dependencies:

  • Download models:

    • W_skin.safetensors → Save to models/loras/.

    • 4x_NMKD-Siax_200k.pth → Save to models/upscale_models/.


3. Workflow Groups

  1. Input Image: Upload portrait (e.g., 1d21adbf...png).

  2. Model Load: Load Nunchaku model and skin LoRA.

  3. Tile Sampling: Process high-res images in tiles.

  4. Post-Processing: Upscale and reassemble tiles.

  5. Output & Compare: Save 4K results with A/B comparison.


4. Inputs & Outputs

  • Inputs:

    • Image: ≥1024x1024 resolution.

    • Prompts: e.g., "fair skin, 8k photo".

  • Output:

    • 4K portrait (saved to ComfyUI/output).


5. Notes

  • VRAM: ≥12GB GPU (8GB with tiling).

  • Debugging:

    • Missing models → Check loras and upscale_models paths.

  • Optimization:

    • Set FluxGuidance=3.5 for better skin details.

    • Reduce denoise=0.4 in BasicScheduler for cleaner results.

Recommend