Efficient Diffusion Training via Min-SNR Weighting Strategy
This addresses a key training bottleneck for researchers and practitioners in generative AI, offering a simple and effective improvement over existing methods.
The paper tackles the slow convergence problem in denoising diffusion models for image generation by introducing the Min-SNR weighting strategy, which adapts loss weights based on signal-to-noise ratios to balance conflicts among timesteps, resulting in 3.4x faster convergence and a new record FID score of 2.06 on ImageNet 256x256.
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-$γ$. 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.