Reproduce “Diffusion Normalizing Flow” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Qinsheng Zhang, Yongxin Chen
DiffFlow integrates Normalizing Flows and Diffusion Models through learnable Stochastic Differential Equations to enable both exact likelihood estimation and efficient sample generation. The approach, developed by Qinsheng Zhang and Yongxin Chen at Georgia Institute of Technology, enhances expressivity for complex data distributions and achieves up to a 20x speedup in sampling over traditional diffusion models while maintaining competitive performance on density estimation and image synthesis tasks.
Key insights
- DiffFlow unifies Normalizing Flows and Diffusion Models, demonstrating that it can reduce to an NF as stochasticity vanishes or approximate a DM with a fixed forward drift.
- Introducing learnable stochasticity into Normalizing Flows enhances their expressivity, enabling them to model distributions with complex topologies and sharp boundaries.
- A learnable forward process in diffusion models significantly improves sampling efficiency while maintaining or improving sample quality.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.