What is dit?

DiT is a research project from Facebook AI Research (Meta) that explores a new architecture for AI image generation. The standard approach to generating images with diffusion models used a design called U-Net as the core processing component. This project replaces U-Net with a transformer, the same type of architecture that powers large language models, and shows that it works well for images too. The key finding in the accompanying research paper is that bigger transformers produce better images in a predictable way. As the model gets deeper, wider, or processes more image patches, image quality improves consistently. The best model in the paper, called DiT-XL/2, achieved state-of-the-art results on standard image generation benchmarks at both 256x256 and 512x512 pixel resolutions. The models are class-conditional, meaning you tell them what category of image to generate (such as a dog or a mushroom from the ImageNet categories) and the model produces an image matching that class. They were trained on ImageNet, a large standard dataset used in computer vision research. The repository includes the model code, pre-trained weights that download automatically, and scripts for both sampling (generating images) and training new models. There is also a runnable demo hosted on Hugging Face Spaces and a Colab notebook, so you can try generating images in a browser without installing anything locally. Training from scratch requires multiple high-end GPUs and is aimed at researchers. The project is written in Python using PyTorch and is intended as a research codebase, released with its pre-trained models so others can reproduce and build on the results.