Rectified Noise: A Generative Model Using Positive-incentive Noise

TL;DR

This paper introduces Rectified Noise, a novel method that enhances existing Rectified Flow generative models by injecting positive-incentive noise, leading to improved image generation quality with minimal additional training.

Contribution

The paper proposes Rectified Noise, a new algorithm to transform pre-trained RF models into pi-noise generators, improving performance with minimal extra parameters.

Findings

RF models with Rectified Noise reduce FID from 10.16 to 9.05 on ImageNet-1k.

Pi-noise generators achieve better performance with only 0.39% more parameters.

Extensive experiments validate the effectiveness across various architectures and datasets.

Abstract

Rectified Flow (RF) has been widely used as an effective generative model. Although RF is primarily based on probability flow Ordinary Differential Equations (ODE), recent studies have shown that injecting noise through reverse-time Stochastic Differential Equations (SDE) for sampling can achieve superior generative performance. Inspired by Positive-incentive Noise (pi-noise), we propose an innovative generative algorithm to train pi-noise generators, namely Rectified Noise (RN), which improves the generative performance by injecting pi-noise into the velocity field of pre-trained RF models. After introducing the Rectified Noise pipeline, pre-trained RF models can be efficiently transformed into pi-noise generators. We validate Rectified Noise by conducting extensive experiments across various model architectures on different datasets. Notably, we find that: (1) RF models using Rectified Noise reduce FID from 10.16 to 9.05 on ImageNet-1k. (2) The models of pi-noise generators achieve improved performance with only 0.39% additional training parameters.