Rectified Noise: A Generative Model Using Positive-incentive Noise

The introduction of Rectified Noise (RN) marks a significant advancement in generative modeling, building on the established Rectified Flow (RF) framework. By incorporating Positive-incentive Noise (pi-noise) into the velocity fields of pre-trained RF models, RN enhances generative performance, as evidenced by a notable reduction in the Fréchet Inception Distance (FID) from 10.16 to 9.05 on the ImageNet-1k dataset. This improvement is achieved with minimal additional training requirements, only 0.39% more parameters, showcasing the efficiency of the new approach. Extensive experiments across various model architectures validate the effectiveness of RN, indicating its potential to influence future developments in AI generative models. The findings underscore the importance of innovative noise injection techniques in enhancing model performance, paving the way for more sophisticated applications in the field of artificial intelligence.