Temporal Difference Learning for Diffusion Models

In three linesNovel training approach for diffusion models using temporal difference (TD) objective to enforce multi-step consistency along the denoising trajectory. Reformulates diffusion as a Markov reward process and denoising as policy evaluation in reinforcement learning. Shows significant FID improvements, especially with few sampling steps.

Summary generated by Claude — human-verified