thu-ml/unidiffuser
A unified diffusion framework that performs image generation, text generation, text-to-image, and image-to-text synthesis in a single transformer model.
Not currently ranked — collecting fresh signals.
star history
This repository implements a multi-modal diffusion model that unifies marginal, conditional, and joint distributions for image-text data. The approach perturbs data across all modalities simultaneously and uses a transformer backbone to predict noise for each modality with individual timesteps. The model is trained on large-scale paired image-text data and can perform diverse generation tasks by setting appropriate timesteps without architectural changes.
Frequently asked
- What is thu-ml/unidiffuser?
- A unified diffusion framework that performs image generation, text generation, text-to-image, and image-to-text synthesis in a single transformer model.
- Is unidiffuser open source?
- Yes — thu-ml/unidiffuser is open source, released under the AGPL-3.0 license.
- What language is unidiffuser written in?
- thu-ml/unidiffuser is primarily written in Python.
- How popular is unidiffuser?
- thu-ml/unidiffuser has 1.5k stars on GitHub.
- Where can I find unidiffuser?
- thu-ml/unidiffuser is on GitHub at https://github.com/thu-ml/unidiffuser.