ScoreMix: Synthetic Data Generation by Score Composition in Diffusion Models Improves Recognition
The episode discusses the 'ScoreMix' paper, a technique that uses score composition during reverse diffusion to create synthetic data. This method blends directional guidance from two distinct classes (cA and cB) to generate challenging, composite samples. The authors report up to a seven percentage point improvement in face recognition benchmarks by focusing on distant class pairs, offering a robust way to augment data without external dependencies.