Parti is an autoregressive text-to-image generation model designed for high-fidelity photorealistic image synthesis, supporting complex compositions and world knowledge. It targets researchers and engineers exploring generative AI, offering an alternative to diffusion models by leveraging advances in large language models through scaling. Parti enables sophisticated image generation by treating the process as a sequence-to-sequence problem, similar to machine translation.

How It Works

Parti models text-to-image generation as a sequence-to-sequence task, analogous to machine translation, allowing it to benefit from large language model advancements. It encodes images into sequences of discrete tokens using the ViT-VQGAN tokenizer, which are then decoded back into high-quality, visually diverse images. This autoregressive approach, particularly when scaled in model size and data, unlocks significant capabilities in generating complex and nuanced visual content.

Quick Start & Requirements

No installation, setup, or specific requirements are detailed in the provided text.

Highlighted Details

• Achieves consistent quality improvements by scaling up to 20 billion parameters.
• Reports state-of-the-art zero-shot FID score of 7.23 and finetuned FID score of 3.22 on the MS-COCO benchmark.
• Demonstrates effectiveness across diverse categories and challenges, evaluated using the newly released PartiPrompts benchmark, a set of over 1600 English prompts.

Maintenance & Community

Parti is a collaborative effort involving multiple Google Research teams, with numerous authors and acknowledgments indicating substantial internal development. No external community channels or specific maintenance details are provided.

Licensing & Compatibility

The provided text does not specify a license or any compatibility notes for commercial or closed-source use.

Limitations & Caveats

Parti is explicitly stated as "not an officially supported Google product." The significant parameter count (up to 20 billion) implies substantial computational resources are necessary for training and potentially inference.