Learning to Be Guided: Optimizing Latent Geometry for Classifier-Free Guidance
In brief
- Classifier-free guidance depends on class geometry in the latent space.
- Prototype regularization improves class separability under noise.
- 31% reduction in guided gFID on ImageNet with negligible overhead.
This study examines how class geometry in the latent space impacts classifier-free guidance (CFG) performance in diffusion models. The authors introduce prototype regularization to enhance class separability under noise, significantly reducing guided generation errors.
This summary gives you the essentials. The full article, its figures and quotes are on cs.AI updates on arXiv.org.
Read the original articleAI-generated summary of the source, checked by rules. Report an error