Compensation Sampling for Improved Convergence in Diffusion Models
This addresses a bottleneck in diffusion models for image generation, offering faster training and improved outputs, though it is incremental as it builds on existing denoising methods.
The paper tackles the slow convergence and lower quality in diffusion models due to accumulated reconstruction errors by proposing compensation sampling with a U-Net-based term, achieving state-of-the-art image quality and accelerating training convergence by up to an order of magnitude on datasets like CIFAR-10 and CelebA.
Diffusion models achieve remarkable quality in image generation, but at a cost. Iterative denoising requires many time steps to produce high fidelity images. We argue that the denoising process is crucially limited by an accumulation of the reconstruction error due to an initial inaccurate reconstruction of the target data. This leads to lower quality outputs, and slower convergence. To address this issue, we propose compensation sampling to guide the generation towards the target domain. We introduce a compensation term, implemented as a U-Net, which adds negligible computation overhead during training and, optionally, inference. Our approach is flexible and we demonstrate its application in unconditional generation, face inpainting, and face de-occlusion using benchmark datasets CIFAR-10, CelebA, CelebA-HQ, FFHQ-256, and FSG. Our approach consistently yields state-of-the-art results in terms of image quality, while accelerating the denoising process to converge during training by up to an order of magnitude.