DisCo-Diff: Integrating Discrete Latents with Continuous Diffusion Models

Diffusion models, encoding data into Gaussian distributions, face complexity issues. DisCo-Diff introduces discrete latents to simplify this. It pairs continuous and discrete variables, trained end-to-end without pre-trained networks. This approach eases the learning curve and improves performance across various tasks, including image synthesis and molecular docking. DisCo-Diff sets new benchmarks, notably in FID scores for ImageNet datasets.

*FID: Fréchet Inception Distance, a metric measuring image quality in generative models.