VAR is a new visual generation framework that makes GPT-style models surpass diffusion models for the first time🚀, and exhibits clear power-law Scaling Laws📈 like large language models (LLMs).

VAR redefines the autoregressive learning on images as coarse-to-fine "next-scale prediction" or "next-resolution prediction", diverging from the standard raster-scan "next-token prediction".

This repo is used for hosting VAR's checkpoints.

For more details or tutorials see https://github.com/FoundationVision/VAR.