VAR

VAR is an image generation research toolkit for researchers and developers studying autoregressive models as an alternative to diffusion. It generates images through next-scale prediction, building finer detail across successive image scales.

• Pretrained image generation: Downloadable VAR models let you generate images on your own hardware, with checkpoints for 256×256 and 512×512 output.
• Sampling controls: You can adjust generation settings to balance image quality and diversity, and use the sampling notebook to explore model behavior.
• ImageNet training: The toolkit includes code for training VAR models on ImageNet and automatically resumes interrupted training from a saved checkpoint.
• Quality evaluation: Its evaluation workflow uses OpenAI's FID evaluation toolkit to measure FID, Inception Score, precision and recall against reference images.

The local implementation uses PyTorch and requires both a VAR checkpoint and the accompanying image autoencoder weights. Training saves checkpoints and logs locally. The code is open source under the MIT license.

A separate hosted browser demo provides interactive VAR text-to-image generation; that demo runs outside the local toolkit.