VAR

Why we included this project

VAR is the official implementation of Visual Autoregressive Modeling, the approach that won the NeurIPS 2024 Best Paper award. Rather than generating an image pixel by pixel, it predicts progressively finer image scales, letting a GPT-style transformer produce 256x256 images on ImageNet at lower compute cost than comparable diffusion models. The repository ships pretrained checkpoints up to 2.3B parameters, along with training, sampling, and evaluation scripts, so you can either load a model quickly or rerun the experiments yourself. It is a strong reference for researchers and engineers exploring autoregressive image synthesis, and the README links to a live text-to-image demo for a quick hands-on check. Teams expecting a full text-to-image service might find the scope narrower, but anyone curious about next-scale prediction as an alternative to diffusion will find a clean, readable codebase grounded in a celebrated paper.