One Transformer Fits All Distributions in Multi-Modal Diffusion at Scale
This addresses the need for efficient and versatile multi-modal generative models, offering a unified approach that reduces computational overhead compared to bespoke models, though it builds incrementally on existing diffusion methods.
The paper introduces UniDiffuser, a unified diffusion framework that simultaneously models marginal, conditional, and joint distributions for multi-modal data, enabling tasks like image, text, and cross-modal generation without extra overhead. It achieves superior or comparable quantitative results, such as FID and CLIP scores, to specialized models like Stable Diffusion and DALL-E 2 in tasks like text-to-image generation.
This paper proposes a unified diffusion framework (dubbed UniDiffuser) to fit all distributions relevant to a set of multi-modal data in one model. Our key insight is -- learning diffusion models for marginal, conditional, and joint distributions can be unified as predicting the noise in the perturbed data, where the perturbation levels (i.e. timesteps) can be different for different modalities. Inspired by the unified view, UniDiffuser learns all distributions simultaneously with a minimal modification to the original diffusion model -- perturbs data in all modalities instead of a single modality, inputs individual timesteps in different modalities, and predicts the noise of all modalities instead of a single modality. UniDiffuser is parameterized by a transformer for diffusion models to handle input types of different modalities. Implemented on large-scale paired image-text data, UniDiffuser is able to perform image, text, text-to-image, image-to-text, and image-text pair generation by setting proper timesteps without additional overhead. In particular, UniDiffuser is able to produce perceptually realistic samples in all tasks and its quantitative results (e.g., the FID and CLIP score) are not only superior to existing general-purpose models but also comparable to the bespoken models (e.g., Stable Diffusion and DALL-E 2) in representative tasks (e.g., text-to-image generation).