CharDiff: Improving Sampling Convergence via Characteristic Function Consistency in Diffusion Models

Diffusion models have demonstrated extensive capabilities for generative modelling in both conditional and conditional image synthesis tasks. Reverse sampling has been the centre of interest to improve the overall image quality without retraining the model from scratch. In this work, we propose a plug-and-play module by utilizing the charac-teristic function of the distributions to minimize sampling drift. We experiment with existing diffusion solvers with our module during the denoising step to provide additional per-formance gain in image synthesis, linear inverse problem tasks and text-conditioned image synthesis. Moreover, We theoretically establish the method's effectiveness in terms of improved Fréchet Inception Distance (FID) and second-order Tweedie moment for reduced trajectory deviation.