SRC-Flow: Compact Semantic Representations Enable Normalizing Flows for Image Generation
Researchers have developed SRC-Flow, a novel method for image generation using normalizing flows. This approach addresses the challenge of high-dimensional representations in visual data by first compressing features into a lower-dimensional semantic space. The method achieves state-of-the-art results among normalizing flow techniques on ImageNet datasets, maintaining exact likelihood computation and deterministic sampling. AI
IMPACT Introduces a new method for likelihood-based image generation that rivals diffusion models in quality while retaining flow-based advantages.