1. Workflow Overview

This is a FLUX-based text-to-image workflow specialized in skin texture enhancement and detail refinement. Key features:

• Uses FP8-quantized FLUX model for efficient generation

• Integrates Ultralytics detector + SAM for face/body inpainting

• Applies LUT color grading and super-resolution

Core Models:

• FLUX (flux1-dev-fp8.safetensors): Base FP8-quantized model

• SD3 CLIP (clip_l + t5xxl_fp8): Dual-text encoder for prompt understanding

• LoRA (FLUX太强悍lora): Skin detail fine-tuner

• ControlNet (Eyeful_v2-Paired.pt): Facial feature control

2. Critical Nodes

Dependencies:

• Model Files:

  • Place flux_ae.sft in models/vae

  • Download x1_ITF_SkinDiffDetail_Lite_v1.pth to models/upscale_models

• Plugins: Must install Impact Pack and ComfyUI-Manager

3. Workflow Structure

Group Logic:

1. Text Encoding (Top-Left):

   • Input: Fantasy-style prompt + negative prompt

   • Output: CLIP conditioning

2. FLUX Generation (Center):

   • Input: 1024x1024 latent + Euler sampler

   • Output: Raw image

3. Detail Refinement (Bottom-Right):

   • Pipeline: Face detect → Inpaint → 2x upscale → LUT grading

   • Key params: Denoise 0.3, Steps 20

4. Inputs & Outputs

Required Inputs:

• Positive prompt: Must include skin descriptors (e.g., "porcelain skin")

• Resolution: Fixed at 1024x1024 (FLUX constraint)

• Seed: Randomizable

Final Output:

• Format: PNG with metadata

• Processing chain: Raw → Sharpen → 2x upscale → LUT

5. Notes

• VRAM: ≥12GB required (even with FP8)

• Common Errors:

  • LUT not found: Verify PB_Boulder.CUBE path

  • SAM load fail: Download sam_vit_b_01ec64.pth

• Optimization:

  • Enable TAESD decoder for faster preview

  • Reduce CR Upscale tile size to 512 for lower VRAM