Efficient Diffusion Training via Min-SNR Weighting Strategy

TL;DR

This paper introduces Min-SNR-$\gamma$ weighting for diffusion model training, significantly accelerating convergence and improving image generation quality on ImageNet benchmarks.

Contribution

It proposes a novel Min-SNR-$\gamma$ weighting strategy that balances timestep conflicts, leading to faster and more effective diffusion model training.

Findings

3.4× faster convergence than previous methods

Achieved a new FID score of 2.06 on ImageNet 256×256

More effective with smaller architectures

Abstract

Denoising diffusion models have been a mainstream approach for image generation, however, training these models often suffers from slow convergence. In this paper, we discovered that the slow convergence is partly due to conflicting optimization directions between timesteps. To address this issue, we treat the diffusion training as a multi-task learning problem, and introduce a simple yet effective approach referred to as Min-SNR-$\gamma$. This method adapts loss weights of timesteps based on clamped signal-to-noise ratios, which effectively balances the conflicts among timesteps. Our results demonstrate a significant improvement in converging speed, 3.4$\times$ faster than previous weighting strategies. It is also more effective, achieving a new record FID score of 2.06 on the ImageNet $256\times256$ benchmark using smaller architectures than that employed in previous state-of-the-art. The code is available at https://github.com/TiankaiHang/Min-SNR-Diffusion-Training.