parti-pytorch

Why we included this project

Parti is Google's answer to diffusion-based text-to-image models: instead of denoising pixels, it treats image generation as an autoregressive sequence problem over discrete visual tokens. This repository reimplements that architecture from scratch in PyTorch, and it pairs the transformer with a ViT-based VQ-GAN VAE that comes with working training code, so you can trace the whole pipeline from image encoding to text-conditioned generation. The API is refreshingly direct: train the VAE, plug it into the Parti module, and call generate() with a list of captions to get PIL images, with classifier-free guidance exposed through a conditioning scale. Just know this is a research-oriented implementation, not a turnkey service; you'll need to bring your own dataset and training budget. It's a good fit if you want to compare autoregressive and diffusion approaches side by side, or poke at the architecture before committing to something bigger.