ComboStoc: Enhancing Diffusion Models via Combinatorial Stochasticity
A new method called ComboStoc addresses under-explored combinatorial complexity in diffusion generative models. The approach constructs stochastic processes that fully exploit combinatorial structures, improving training efficiency across images and 3D shapes. It also enables asynchronous time steps during generation.
Key facts
- ComboStoc targets combinatorial complexity in diffusion models.
- Existing training schemes may insufficiently cover dimension-attribute combinations.
- The method accelerates network training for images and 3D shapes.
- It allows asynchronous time steps for different dimensions at test time.
- The paper is available on arXiv with ID 2405.13729.
Entities
Institutions
- arXiv