3 articles tagged with #normalizing-flows. AI-curated summaries with sentiment analysis and key takeaways from 50+ sources.
A comprehensive academic resource presenting the unified mathematical foundations of diffusion models, explaining how three complementary perspectives—variational, score-based, and flow-based—emerge from shared principles. The work bridges theoretical understanding with practical applications including controllable generation and efficient sampling methods.
Researchers propose a novel framework for detecting anomalies in dynamic graphs using limited labeled data, combining residual representation encoding with a bi-boundary optimization strategy to balance discrimination and generalization. The model-agnostic approach addresses the gap between unsupervised methods (which produce ambiguous boundaries) and semi-supervised methods (which overfit to limited anomalies).
Researchers introduce STARFlow-V, a normalizing flow-based generative model for video that challenges the dominance of diffusion models in the space. The approach offers end-to-end likelihood estimation, causal prediction capabilities, and computational efficiency advantages for video generation tasks.
