Reproduce “ComboStoc: Combinatorial Stochasticity for Diffusion Generative Models” on OpenResearch

About this paper

Rui Xu, Jiepeng Wang, Hao Pan, Yang Liu, Xin Tong, Shiqing Xin, Changhe Tu, Taku Komura and 1 more

ComboStoc completes partial chairs by letting supplied and generated components progress at different rates, producing diverse designs that preserve the base parts. It assigns elements different progress levels from noise toward data during training, rather than synchronizing them. On class-conditional ImageNet generation, it reports FID 15.69 at 400K steps, versus 17.2 for SiT-XL, a same-step comparison, not lower inference cost.

Key insights

• Explicitly sampling the combinatorial complexity of data dimensions and attributes during diffusion training leads to faster convergence rates and systematically higher generation quality.
• A tensorized, asynchronous interpolation schedule, even with varying timesteps across dimensions, can still define a proper generative flow model, satisfying the continuity equation.
• The asynchronous nature of the stochastic process enables new inference modes, providing graded control over different dimensions or attributes for flexible partial preservation, inpainting, and assembly tasks.

Summary by alphaXiv.