Parti

Google Research's 20-billion parameter autoregressive model designed for high-fidelity text-to-image synthesis.

Parti (Pathways Autoregressive Text-to-Image) treats image generation as a sequence-to-sequence task: converting text tokens into discrete visual codebook entries. Built on the Pathways infrastructure, the model scales from 350 million to 20 billion parameters to capture fine-grained world knowledge and complex spatial logic. It utilizes a ViT-VQGAN encoder-decoder to produce initial 256x256 patches (upscaled to 1024x1024) that excel at rendering legible text and precise object counts. Performance on the 1,600+ prompt PartiPrompts benchmark confirms its ability to handle intricate descriptions that often challenge diffusion models.

https://parti.research.google/