Elucidating the Design Space of Diffusion-Based Generative Models

Explore the intricacies of diffusion-based generative models in this comprehensive conference talk presented at NeurIPS 2022. Delve into a newly proposed design space that simplifies the theory and practice of these models by clearly separating concrete design choices. Discover improvements to sampling and training processes, as well as score network preconditioning, leading to state-of-the-art results in both class-conditional and unconditional settings for CIFAR-10. Learn how these advancements significantly enhance efficiency and quality of pre-trained score networks, including a notable improvement in ImageNet-64 model performance. Gain insights from Miika Aittala, a Senior Research Scientist at NVIDIA Research, as he shares his expertise in neural generative modeling, image processing, and realistic image synthesis in computer graphics.

Miika Aittala: Elucidating the Design Space of Diffusion-Based Generative Models