50 min
Seminar, Hosted by Simons Institute for the Theory of Computing
Global selection, local completion: a probabilistic anatomy of diffusion U-Net
Recording
Abstract
Two studies examine how diffusion U-Nets model complex images efficiently. Coarse-to-fine conditioning makes fine detail predictable through small, local models, reducing effective dimensionality. At the bottleneck, sparse active channels encode global image information whose geometry reflects semantic similarity even without labels or text conditioning. Self-guided reconstruction probes that representation, supporting an account that combines compact global selection with locally conditioned completion.
Topics
- diffusion U-Nets, search World Wide
- complex images, broader topic: images
- Coarse-to-fine conditioning, search World Wide
- fine detail, search World Wide
- local models, search World Wide
- effective dimensionality, broader topic: dimensionality
- sparse active channels, search World Wide
- global image information, broader topic: information
- semantic similarity
- Self-guided reconstruction, broader topic: reconstruction