Mastering the Art of Chinese Illustrations with Advanced CLIP Encoders

CN
ComfyUI.org
2025-04-30 09:21:16

1. Workflow Overview

ma3q6u10glpvkel1xx8c107c3ec4c792abe18d0a1dd9440c74567396c476d8ac7fdb714901bcee4a350.png

This workflow specializes in generating Eastern-style illustrations using Flux technology stack. Key features:

  • Dual CLIP encoders (t5xxl_fp8 + clip_l)

  • High-res output (2272x1280)

  • Style enhancement via Lora

Core Models:

  • BaseAlgo_F.1: Main model (FP8 optimized)

  • ae.sft: Dedicated VAE

  • FantasyMirage.lora: Style adapter (0.9 strength)

2. Node Breakdown

Critical Nodes:

  1. DualCLIPLoader

    • Loads t5xxl_fp8_e4m3fn (Chinese-optimized) + clip_l

  2. FluxGuidance

    • Advanced conditioning control (CFG=3.5)

  3. GetNode/SetNode

    • Data routing system:

      • "1": Latent flow

      • "c": CLIP model flow

      • "l": Conditioning flow

  4. easy cleanGpuUsed

    • VRAM cleaner (requires KJNodes)

Dependencies:

  • FP8 support (RTX 30/40 series)

  • ComfyUI-Flux recommended

3. Workflow Structure

Generation Stages:

  1. Text Encoding

    • Processes complex Chinese prompts (e.g. philosophical descriptions)

  2. Latent Generation

    • Euler sampler (20 steps)

    • Empty latent: 2272x1280 (batch=4)

  3. Post-Processing

    • Auto VRAM cleanup

    • Dual output channels

4. I/O Specification

Inputs:

  • Positive prompt:

    Cascading ochre-yellow composition...(200+ words artistic description)  
  • Seed: 566815113879132 (randomizable)

Outputs:

  • Path: /ComfyUI/

  • Format: PNG (ultra-high res)

5. Critical Notes

  1. Hardware

    • FP8 support required

    • VRAM: ≥16GB (for 2272px)

  2. Model Paths

    • Main model: ComfyUI/models/unet/

    • Loras: ComfyUI/models/loras/

  3. Troubleshooting

    • NaN errors: Reduce FluxGuidance CFG

    • VRAM issues:

      Decrease batch size in EmptyLatentImage (current:4)  

Recommend