ComfyUI.org

Unlock Consistent Anime-Style Characters with AI-Driven Workflow

CN
ComfyUI.org
2025-06-18 13:38:04

1. Workflow Overview

mc1zyf2lprv2cc06uybComfyUI_00229_.png

This workflow is designed to test and validate character consistency in AI-generated images, ideal for social media posts or comic character design. Key features:

  • Generates anime-style characters using SDXL base model.

  • Dynamically combines traits (e.g., clothing, hairstyle) via PromptGenerator.

  • Enhances details with SDXL Detail LoRA (e.g., facial textures).

  • Supports multi-prompt testing to ensure consistency across variations.

2. Core Models

Model Name

Function

Huimeng YuanYu SDXL

Anime-fantasy base model (for posters/comics).

SDXL Detail LoRA

Enhances fine details (e.g., fabric folds).

sdxl_vae.safetensors

Improves color and decoding quality.

3. Key Nodes & Installation

Node Name

Function

Installation

PromptGenerator

Generates dynamic character prompts.

Requires pysssss custom nodes.

SDXLPromptStyler

Standardizes prompt format (positive/negative).

Via ComfyUI Manager.

CLIPSetLastLayer

Adjusts CLIP layers for text influence.

Built-in.

KSampler (Efficient)

Optimized sampler (supports DPM++ 2M).

Built-in.

Dependencies:

  • Download Huimeng YuanYu SDXL and SDXL Detail LoRA to models/ and loras/.

  • Install pysssss extension for dynamic prompt generation.

4. Workflow Structure

  • Group 1: Model Loading

    • Nodes: CheckpointLoader → VAELoader → LoraLoader → CLIPSetLastLayer

    • Output: Loaded models, VAE, LoRA, and CLIP config.

  • Group 2: Dynamic Prompt Generation

    • Nodes: PromptGenerator → ShowText (preview prompts)

    • Output: Combined traits (e.g., "burly male teacher in denim shorts").

  • Group 3: Sampling

    • Nodes: CLIPTextEncode → KSampler → EmptyLatentImage

    • Output: Generated latent image.

  • Group 4: Output

    • Nodes: VAEDecode → SaveImage

    • Output: Final PNG (default 1024x1024).

5. Inputs & Outputs

  • Inputs:

    • Required: Character traits (e.g., "ponytail hair", "white crop top").

    • Optional: Seed (e.g., 392280905866782), steps (45), CFG scale (6.5).

  • Output: A character image matching the description, with multi-variant testing support.

6. Notes

  1. Consistency Tips:

    • Fix the seed value for reproducible results.

    • Set CLIPSetLastLayer to -2 for finer control.

  2. Performance:

    • Reduce resolution to 768x768 if VRAM is limited.

  3. Troubleshooting:

    • Prompt conflicts: Avoid opposing traits (e.g., "smiling" + "angry").

    • LoRA issues: Verify model path and strength (default 1.0).

previous
Face Swap Like a Pro: Mastering the PuLID-Flux & FLUX-REDUX Workflow for Stunning Results
next
Bring New Life to Old Photos: Advanced Damage Repair and Super-Resolution

Recommend