DALL-E
PyTorch package for the discrete VAE used for DALL·E.
Official code release of one component from OpenAI's DALL-E image model: the discrete VAE that compresses and reconstructs images. The full text-to-image model is not included, intended for researchers studying the architecture.
This is the official code release from OpenAI for one specific component of DALL-E, the AI system that generates images from text descriptions. The piece released here is the discrete VAE, which is a part of the model responsible for compressing and reconstructing images. The main transformer that actually generates images from text is not included in this release.
The package is written in Python using PyTorch, a popular library for building AI models. Installation is a single command. The repository includes an example notebook that demonstrates how to use the released component. The README is minimal and points primarily to the original blog post, the research paper, and a model card that describes the system's intended uses and limitations.
Where it fits
- Study how DALL-E encodes images into discrete tokens and reconstructs them as part of understanding the overall architecture.
- Experiment with the discrete VAE component in your own PyTorch project using the included example notebook.
- Use the model card to understand OpenAI's stated intended uses and limitations for this system before citing it in research.