Reproduce “Learning When to Denoise: Optimizing Asynchronous Schedules for Latent Diffusion” on OpenResearch

About this paper

Bingshuo Qian, Xiang Cheng

The method learns a schedule for semantic and texture denoising, replacing a fixed offset with a curve and correcting texture-loss weighting. A short probe learns the curve, which is then fixed for final training. On class-conditional ImageNet at 256×256, Ours-XL reaches FID 1.05 after 200 epochs versus SFD-XL’s 1.06 after 800; both use 250 solver evaluations, and the comparison includes corrected loss.

Key insights

• Learning asynchronous schedules dynamically adapts the denoising process, leveraging the distinct informational roles of semantic and texture representations more effectively than fixed schedules.
• A kinetic energy regularizer is essential during schedule learning to prevent the schedule from collapsing to undesirable extremal curves and to ensure smooth, stable denoising trajectories.
• The two-stage schedule probe efficiently identifies optimal schedules by decoupling schedule discovery from the primary denoiser training, enabling practical integration into large-scale model development.

Summary by alphaXiv.