GAS: Improving Discretization of Diffusion ODEs via Generalized Adversarial Solver

TL;DR

The paper tackles the bottleneck of slow diffusion-model sampling by introducing Generalized Solver (GS), a simple, trainable parameterization for the ODE sampler, and Generalized Adversarial Solver (GAS), which adds adversarial training to preserve fine-grained details. GS leverages additive corrections to a theoretically guided base solver and learns a time schedule and timestep corrections, enabling high-quality few-step sampling. GAS further improves realism and artifact suppression by incorporating a relativistic GAN loss into the distillation framework, yielding superior FID/LPIPS performance across multiple datasets and NFEs. The approach delivers practical gains in sample quality with minimal inference-time overhead, offering a scalable path to faster diffusion-based generation under resource constraints.

Abstract

While diffusion models achieve state-of-the-art generation quality, they still suffer from computationally expensive sampling. Recent works address this issue with gradient-based optimization methods that distill a few-step ODE diffusion solver from the full sampling process, reducing the number of function evaluations from dozens to just a few. However, these approaches often rely on intricate training techniques and do not explicitly focus on preserving fine-grained details. In this paper, we introduce the Generalized Solver: a simple parameterization of the ODE sampler that does not require additional training tricks and improves quality over existing approaches. We further combine the original distillation loss with adversarial training, which mitigates artifacts and enhances detail fidelity. We call the resulting method the Generalized Adversarial Solver and demonstrate its superior performance compared to existing solver training methods under similar resource constraints. Code is available at https://github.com/3145tttt/GAS.

Figures (24)

• Figure 1: Illustration of the Generalized Adversarial Solver image generation in comparison with the training-free UniPC zhao2024unipc solver with equal number of function evaluations (NFEs). Our method shows superior results that are almost identical to teacher images in terms of generation quality.
• Figure 2: Illustration of the Generalized Adversarial Solver. Our student makes each sampling step by calculating the weighted average of all previous points and velocity directions. We train the corresponding weights and timestep schedule via distillation and adversarial loss.
• Figure 3: We compare our parametrization with S4S variant on CIFAR and FFHQ datasets in terms of FID and LPIPS scores. Both setups use batch size of 24, while training dataset consists of 49k samples. Teacher dataset has FID score of 2.03 and 2.60 for CIFAR10 and FFHQ datasets respectively.
• Figure 4: Incorporating an adversarial loss into the training process enhances generation quality reducing occurring image artifacts in low NFEs regimes. In this setup, the teacher model uses UniPC (3M) solver with NFE=10, while the student models operate with a reduced NFE=4.
• Figure 5: FID score for several checkpoints during training of GS and GAS for both pixel (FFHQ) and latent space (LSUN, ImageNet) datasets.