Easing Color Shifts in Score-Based Diffusion Models

Generated images of score-based models can suffer from errors in their spatial means, an effect, referred to as a color shift, which grows for larger images. This paper introduces a computationally inexpensive solution to mitigate color shifts in score-based diffusion models. We propose a simple nonlinear bypass connection in the score network, designed to process the spatial mean of the input and to predict the mean of the score function. This network architecture substantially improves the resulting spatial means of the generated images, and we show that the improvement is approximately independent of the size of the generated images. As a result, our solution offers a comparatively inexpensive solution for the color shift problem across image sizes. Lastly, we discuss the origin of color shifts in an idealized setting in order to motivate our approach.