Enhancing Sample Generation of Diffusion Models using Noise Level Correction

TL;DR

This paper introduces a noise level correction network that improves diffusion model sample quality by better aligning estimated noise levels with true distances to the data manifold, applicable across various image restoration tasks.

Contribution

We propose a novel noise level correction method that enhances diffusion model sampling and extends to multiple image restoration tasks, improving quality and compatibility with existing schedulers.

Findings

Significant improvement in sample quality across tasks

Compatibility with existing denoising schedulers like DDIM

Effective noise level alignment enhances diffusion process

Abstract

The denoising process of diffusion models can be interpreted as an approximate projection of noisy samples onto the data manifold. Moreover, the noise level in these samples approximates their distance to the underlying manifold. Building on this insight, we propose a novel method to enhance sample generation by aligning the estimated noise level with the true distance of noisy samples to the manifold. Specifically, we introduce a noise level correction network, leveraging a pre-trained denoising network, to refine noise level estimates during the denoising process. Additionally, we extend this approach to various image restoration tasks by integrating task-specific constraints, including inpainting, deblurring, super-resolution, colorization, and compressed sensing. Experimental results demonstrate that our method significantly improves sample quality in both unconstrained and constrained generation scenarios. Notably, the proposed noise level correction framework is compatible with existing denoising schedulers (e.g., DDIM), offering additional performance improvements.