Synthetic Data Generation (Diffusion + ControlNet)
Filled rare classes with targeted synthetic images.
Problem
Rare categories suffered from data scarcity.
Approach
- Used Stable Diffusion + ControlNet to generate domain-specific images
- Curated prompts/controls to match dataset distribution
Impact
- Increased detection accuracy by 12% on rare categories
- Reduced labeling burden for edge cases