Parti
By Google Research
Parti (Pathways Autoregressive Text-to-Image model) is a text-to-image generation model from Google Research, introduced in 2022, that treats image generation as a sequence-to-sequence translation problem, autoregressively predicting…
Definition
Parti (Pathways Autoregressive Text-to-Image model) is a text-to-image generation model from Google Research, introduced in 2022, that treats image generation as a sequence-to-sequence translation problem, autoregressively predicting discrete image tokens conditioned on a text prompt rather than using the iterative denoising process found in diffusion models.
Overview
Most well-known text-to-image systems of the early 2020s, including Stable Diffusion and DALL-E 2, generate images through a denoising diffusion process. Parti took a different, autoregressive approach: it first uses a discrete image tokenizer (based on a VQGAN-style encoder) to represent an image as a sequence of tokens from a fixed visual vocabulary, then trains a large Transformer to predict that sequence of image tokens one at a time, conditioned on the text prompt — treating text-to-image generation analogously to machine translation, where the "target language" is a sequence of image tokens rather than words. Google Research demonstrated Parti at scale, with the largest version reaching 20 billion parameters, and reported strong performance on content-rich, complex prompts involving multiple objects, unusual combinations, and specific text rendering within images — areas that were historically difficult for earlier text-to-image systems. Parti's results were presented as evidence that scaling autoregressive Transformer approaches, which had driven progress in language models like PaLM, could similarly benefit image generation. Parti was primarily a research system rather than a public product, and Google Research's later multimodal and image-generation efforts moved toward diffusion-based and hybrid approaches (informing later systems such as Imagen), so Parti is best understood today as an influential research milestone demonstrating a viable autoregressive alternative to diffusion for large-scale text-to-image generation, rather than a tool in active production use.
Key Concepts
- Autoregressive text-to-image generation, unlike diffusion-based competitors
- Represents images as sequences of discrete tokens via a VQGAN-style tokenizer
- Scaled up to 20 billion parameters in its largest configuration
- Frames text-to-image generation as a sequence-to-sequence translation task
- Strong reported performance on complex, multi-object, and text-in-image prompts
- Research system from Google Research, not released as a public product