How boyuanjiang/fitdit works
Main Components Identification
a. Preprocessing Pipeline:
• DWPose Module (/preprocess/dwpose)
• Human Parsing Module (/preprocess/humanparsing)
b. Core Processing Components:
• Attention Mechanisms (src/attention_garm.py, src/attention_vton.py)
• Transformers (src/transformer_sd3_garm.py, src/transformer_sd3_vton.py)
• Pose Guider (src/pose_guider.py)
c. Pipeline Orchestrator:
• Main Pipeline (src/pipeline_stable_diffusion_3_tryon.py)
[Input Layer]
• Model Image
• Garment Image
[Preprocessing Layer]
• DWPose Processing
• Pose Estimation
• Human Parsing
• Segmentation Masks
[Core Processing Layer]
• Pose Guider
• Pose-aware Feature Extraction
• Attention Mechanisms
• GARM Attention
• VTON Attention
• Transformers
• GARM Transformer
• VTON Transformer
[Output Layer]
• Try-on Result
Important Design Elements
• Two-step process mentioned in the README:
1. Mask Generation
2. Try-on Generation
b. Processing Stages:
• Resolution handling (1152x1536 default)
• Mask adjustment capabilities
• Different precision options (bf16, fp16)
c. Memory Management:
• Offload options (CPU offload, aggressive offload)