Timezone: »

One Transformer Fits All Distributions in Multi-Modal Diffusion at Scale
Fan Bao · Shen Nie · Kaiwen Xue · Chongxuan Li · Shi Pu · Yaole Wang · Gang Yue · Yue Cao · Hang Su · Jun Zhu

Tue Jul 25 05:00 PM -- 06:30 PM (PDT) @ Exhibit Hall 1 #415
Event URL: https://github.com/thu-ml/unidiffuser »

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).

Author Information

Fan Bao (Tsinghua University)
Shen Nie (Renmin University of China)
Kaiwen Xue (University of Electronic Science and Technology of China)
Chongxuan Li (Tsinghua University)
Shi Pu (Shengshu)
Yaole Wang (Qiyuanlab )
Gang Yue (qiyuanlab)
Yue Cao (Beijing Academy of Artificial Intelligence)
Hang Su (Tsinghua University)
Jun Zhu (Tsinghua University)

More from the Same Authors