1. Workflow Overview

This professional workflow redesigns print patterns through:

1. Original pattern analysis (Florence-2)

2. AI regeneration (Stable Diffusion + Flux)

3. Precision masking (Segment Anything)

4. Garment compositing

Core Models:

• F.1-dev-fp8_1: FP8-optimized main model

• Florence-2-base: Visual understanding

• sam_vit_h: Segmentation model

2. Node Breakdown

Key Nodes:

1. Florence2Run

   • Auto-generates prompts from patterns

2. SegmentAnythingUltra V2

   • Hair-accurate masking

   • Requires VITMatte algorithm

3. ImageBlendAdvance

   • 90% opacity control

   • Lanczos resampling

Dependencies:

• ComfyUI-LayerUtility plugin

• Download GroundingDINO_SwinT (694MB)

3. Workflow Stages

1. Pre-Processing

   • Upscale input pattern 4x with ESRGAN

   • Load shirt template (1024x1280px)

2. Generation

   • FluxGuidance (CFG=3.5)

   • 20-step Euler sampling

3. Compositing

   • Light/shadow matching

   • Side-by-side comparison

4. I/O Specification

Inputs:

• Source print: 印花.png

• Garment: white_shirt1.png

Outputs:

• Real-time preview

• Interactive comparer

5. Critical Notes

1. Hardware

   • FP8 requires RTX 30/40 series

   • Recommended VRAM: ≥12GB

2. Model Paths

   • Florence-2 auto-downloads to:

     ComfyUI/models/florence2/

   • Manually install SAM models:

     ComfyUI/models/sam/

3. Troubleshooting

   • Rough edges:

     Adjust mask_threshold (default:0.3)

   • Poor garment fit:

     Change blend mode to "multiply"