A novel approach, Visual AutoRegressive (VAR) modeling, has been introduced to the realm of autoregressive learning on images, shifting the method from the traditional raster-scan next-token prediction to inherently scalable image generation via next-scale prediction. VAR’s application marks the first instance where AR models excel over diffusion transformers in this domain. Read more

Key results of VAR include demonstrable improvements such as:

• A reduction in Frechet inception distance (FID) from 18.65 to a remarkable 1.80.
• A boost in inception score (IS) from 80.4 to a staggering 356.4.
• An impressive 20-fold increase in inference speed.
• Competency in zero-shot generalization, extending to tasks like image in-painting.

Given these impactful results, VAR provides a solid foundation for future AR model exploration, particularly for visual generation and unified learning.