Multimodal AI · Emerging · Advanced

Continuous Autoregressive Vision

Also known as: Autoregressive Image Models, Visual Autoregression

An architecture that generates images token-by-token or scale-by-scale using autoregressive transformer mechanics.

What Continuous Autoregressive Vision is

Provides unified multimodal architectures where text and image generation share identical transformer underlying logic.

How it works

Quantizes visual tokens using vector-quantized autoencoders and predicts next visual tokens via causal attention.

Why it matters

Pioneered in models like LLaMA-Vision, Chameleon, and Emu3 for seamless multimodal synthesis.

Common uses