S-NODE: Explicit and Reversible Image Translation Encoding With Neural ODEs

Abstract

Score-based diffusion models achieve high-quality data generation through an iterative denoising process. However, the stochastic term in the diffusion process prevents them from accomplishing reversible generative modeling. To tackle this problem, we present S-NODE, a novel generative model capable of reversible and conditional data generation. Unlike score-based models, S-NODE is an entirely deterministic generative method bridging the score function and neural ordinary differential equations (ODEs). First, we propose and prove an ODE utilizing a score-related difference as the drift term to model transformations between two certain data distributions. Second, we suggest a path-constrained loss to reduce truncation errors, enhancing the model’s capabilities in generating high-quality samples. Third, S-NODE can use a single conditional model to generate and translate cross-class images in all stages without additional training. Extensive experiments on various tasks demonstrate the effectiveness and reversibility of our method. Compared with other ODE-based and score-based methods, S-NODE achieves superior performance (FID of 2.29 & IS of 9.96) on CIFAR-10 and facilitates reversible image translation and image interpolation on CelebA, MetFace, and AFHQ datasets.