How eth-ait/gaussianhaircut works
Main Components
a) Input Processing Pipeline:
• Video Input Component
• Preprocessing Module (from src/preprocessing/)
• Image Processing
• Face Alignment
• Mask Generation
• FLAME Fitting Integration
b) Core Processing Systems:
• Gaussian Model System
• Base Gaussian Model
• Strand-Aligned Gaussian Model
• Latent Strand Model
• CUDA Rasterization Engine (from ext/diff_gaussian_rasterization_hair/)
• Training Pipeline
c) Rendering System:
• Gaussian Renderer
• Post-processing Pipeline
• Blender Integration
Component Relationships
• Data flow from raw video input through preprocessing
• Training pipeline connecting Gaussian models with the rasterization engine
• Rendering pipeline showing both real-time visualization and final output generation
Specific Implementation Details
[Input Video] → [Preprocessing Pipeline]
• ↓
[Training System] ← → [CUDA Rasterization Engine]
• ↓
[Gaussian Models] → [Rendering Pipeline] → [Output Visualization]
Important Architectural Elements to Highlight
a) External Dependencies:
• CUDA 11.8 integration
• Blender 3.6 integration
• Neural Haircut components
• FLAME fitting pipeline
b) Core Processing Pipeline:
• Strand-aligned 3D Gaussian processing
• Differential rasterization system
• Training loop for Gaussian models
Technical Details
• CUDA-accelerated components
• Model architecture showing:
• Gaussian model variants (base, strands, latent)
• Rasterization pipeline
• Training systems