Abstract:A wide variety of deep generative models has been developed in the past decade. Yet, these models often struggle with simultaneously addressing three key requirements including: high sample quality, mode coverage, and fast sampling. We call the challenge imposed by these requirements the generative learning trilemma, as the existing models often trade some of them for others. Particularly, denoising diffusion models have shown impressive sample quality and diversity, but their expensive sampling does not yet allow them to be applied in many real-world applications. In this paper, we argue that slow sampling in these models is fundamentally attributed to the Gaussian assumption in the denoising step which is justified only for small step sizes. To enable denoising with large steps, and hence, to reduce the total number of denoising steps, we propose to model the denoising distribution using a complex multimodal distribution. We introduce denoising diffusion generative adversarial networks (denoising diffusion GANs) that model each denoising step using a multimodal conditional GAN. Through extensive evaluations, we show that denoising diffusion GANs obtain sample quality and diversity competitive with original diffusion models while being 2000$\times$ faster on the CIFAR-10 dataset. Compared to traditional GANs, our model exhibits better mode coverage and sample diversity. To the best of our knowledge, denoising diffusion GAN is the first model that reduces sampling cost in diffusion models to an extent that allows them to be applied to real-world applications inexpensively. Project page and code can be found at https://nvlabs.github.io/denoising-diffusion-gan

A solvable generative model with a linear, one-step denoiser

Towards Faster Non-Asymptotic Convergence for Diffusion-Based Generative Models

Diffusion Model for Generative Image Denoising

RISSOLE: Parameter-efficient Diffusion Models via Block-wise Generation and Retrieval-Guidance

From Denoising Diffusions to Denoising Markov Models

Edge-preserving noise for diffusion models

What's the score? Automated Denoising Score Matching for Nonlinear Diffusions

Convergence of denoising diffusion models under the manifold hypothesis

Tackling the Generative Learning Trilemma with Denoising Diffusion GANs

Free Hunch: Denoiser Covariance Estimation for Diffusion Models Without Extra Costs

Statistical guarantees for denoising reflected diffusion models

Observation-Guided Diffusion Probabilistic Models

On the Relation Between Linear Diffusion and Power Iteration

On Memorization in Diffusion Models

A Note on the Convergence of Denoising Diffusion Probabilistic Models

Evaluating the design space of diffusion-based generative models

Denoising Diffusion Samplers

Diffusion with Forward Models: Solving Stochastic Inverse Problems Without Direct Supervision