FoundationVision/VAR

Project overview

VAR implements coarse-to-fine next-scale prediction for visual autoregressive image generation, demonstrating power-law scaling behavior and surpassing diffusion model benchmarks in image generation quality. The codebase provides training scripts and pre-trained model weights ranging from VAR-d16 to VAR-d36.

Project type

Model Development · Model Runtime · Image & Vision

Deployment

Refer to project documentation

License

MIT

Best for

• Researchers exploring autoregressive visual generation and scaling laws.
• Users needing pre-trained VAR weights for multi-resolution image generation.

Key capabilities

• Defines autoregressive learning on images as a coarse-to-fine next-scale prediction mechanism rather than standard raster-scan next-token prediction.
• Includes shell scripts that use torchrun to train VAR models on the ImageNet dataset at multiple scales.

Limitations and risks

• The project requires the user to manually download VQVAE weights and VAR model weights.
• GPU requirements, minimum hardware, telemetry policies, and cost dependencies are not documented.
• Training models requires the user to obtain the external ImageNet dataset.

Getting started

• Setup involves a medium difficulty process requiring the installation of torch>=2.0.0 and requirements.txt, followed by manually downloading VAR model weights and VAE weights.

Evidence and sources

• GitHub project description: [NeurIPS 2024 Best Paper Award][GPT beats diffusion🔥] [scaling laws in visual generation📈] Official impl. of "Visual Autoregressive Modeling: Scalable Image Generation via Next-Sc…
• README: 🔥 For the first time, GPT-style autoregressive models surpass diffusion models🚀:
• README: Visual Autoregressive Modeling (VAR) redefines the autoregressive learning on images as coarse-to-fine "next-scale prediction" or "next-resolution prediction", diverging from the…
• README: 2024-12: 🏆 VAR received **NeurIPS 2024 Best Paper Award**.
• README: We provide VAR models for you to play with, which are on <a href='https://huggingface.co/FoundationVision/var'><img src='https://img.shields.io/badge/%F0%9F%A4%97%20Huggingface-Fo…

Reviewed at: 2026-07-18T04:25:10Z