Visual AutoRegressive modeling (VAR) introduces a fresh take on autoregressive image generation, moving away from the traditional ‘next-token prediction’ to a ‘next-scale prediction’ method. This approach has allowed autoregressive models to surpass their diffusion transformer counterparts, with impressive results:

• Achieved on ImageNet 256x256, VAR’s method greatly improved the Frechet inception distance (FID) from 18.65 to 1.80, and inception score (IS) from 80.4 to 356.4.
• VAR boasts an approximated 20x faster inference speed over traditional methods.
• The models demonstrate strong performance in zero-shot generalization for image in-painting, out-painting, and editing.
• Evidence of power-law scaling laws observed in LLMs, with linear correlation coefficients near -0.998.

In my assessment, the VAR method represents a substantial advancement in visual distribution learning and visual generation capabilities. Its ability to emulate the crucial properties of LLMs, such as scaling laws and zero-shot task generalization, opens up possibilities for further research in areas like zero-shot learning, model scaling, and the application of autoregressive models across different AI tasks. Discover more about VAR.