Reproduce “Deep Generative Learning via Schrödinger Bridge” on OpenResearch

About this paper

Gefei Wang, Yuling Jiao, Qian Xu, Yang Wang, Can Yang

A deep generative model leverages the Schr¨odinger Bridge problem to transform a reference distribution into a target distribution over a finite time interval, providing consistency guarantees under mild smoothness assumptions. The two-stage approach generates high-fidelity images, achieving an FID of 12.32 and IS of 8.14 on CIFAR-10, and effectively captures multimodal data distributions.

Key insights

• The Schr¨odinger Bridge formulation allows establishing theoretical consistency guarantees for deep generative models under substantially milder smoothness assumptions than prior SDE-based methods.
• A two-stage process is essential for effectively bridging the gap from a simple initial distribution to complex, high-dimensional, and potentially multimodal target data.
• The second stage of the proposed algorithm demonstrates a theoretical connection to the reverse-time Variance Exploding (VE) SDE algorithm, offering deeper insight into its efficacy.

Summary by alphaXiv.