VAR (Visual Autoregressive Modeling) is the research codebase that redefined how autoregressive models generate images — predicting whole resolution scales instead of individual tokens — and won a NeurIPS 2024 Best Paper Award. Released under the MIT license, it has collected more than 8,700 stars on GitHub.
Classic autoregressive image models raster-scan token by token, which is slow and ignores the 2D structure of images. VAR replaces next-token prediction with next-scale prediction: the model generates a full token map at low resolution, then progressively refines it scale by scale up to the final image. This makes GPT-style generation dramatically faster and, for the first time, let autoregressive models beat strong diffusion models on ImageNet benchmarks while showing clean scaling laws. The repository provides the official, deliberately readable implementation with pretrained checkpoints. Its coarse-to-fine structure also mirrors how images are naturally composed — global layout first, details later — which is why VAR has become a reference point for subsequent unified autoregressive models that try to handle text and images with a single transformer.
VAR is free and open source under the MIT license, including the released model checkpoints; you supply your own GPU compute.
VAR checkpoints come in several depths, so users can trade output quality against speed, and the family generates images far faster than token-by-token autoregressive baselines. The codebase’s clarity has made it a popular starting point for later work extending next-scale prediction to higher resolutions and richer conditioning.
It is for generative-model researchers and engineers studying autoregressive image generation, and for teams prototyping fast, scalable text-to-image or class-conditional image systems.