Research Notes

Parti and Autoregressive Image Generation

By Satwik · March 4, 2026

Parti, also from Google, approached text-to-image generation as an autoregressive sequence problem rather than diffusion. It used an image tokenizer to turn pictures into discrete tokens, then trained a transformer to generate those image tokens conditioned on text, essentially treating image synthesis like language modeling over a visual vocabulary. Scaling the model up to billions of parameters produced steady gains in fidelity and, notably, in handling complex, compositional prompts with multiple objects and relationships.

Parti mattered as a proof that the autoregressive, transformer-based recipe familiar from language models transfers to high-quality image generation, running in parallel to the diffusion approaches. It showed two quite different generative paradigms converging on similar photorealistic capability, which is a strong signal that the capability was a function of scale and data rather than any single architectural trick.

The compositionality gains are the security-relevant part. As models get better at rendering exactly the scene a prompt describes - specific objects, arrangements, and text within images - the precision of synthetic media rises. Controllable, prompt-faithful generation is what turns image models from novelty into tools for targeted fabrication.

Like Imagen, Parti was not openly released, and Google framed it within the same responsible-disclosure posture. For our tracking, Parti reinforces the theme that this capability was arriving from multiple directions at once. When two independent architectures both cross the photorealism threshold in the same year, containment via any single non-release becomes clearly insufficient, and the governance conversation has to shift to the capability class as a whole.