var

Introduction

VAR (Visual Autoregressive) Transformers represent a novel visual generation framework that allows GPT-style models to outperform diffusion models. The framework exhibits power-law scaling laws similar to large language models.

Architecture

VAR redefines autoregressive learning on images by implementing a coarse-to-fine "next-scale prediction" or "next-resolution prediction," which differs from the conventional raster-scan "next-token prediction." This approach enables more effective handling of visual data by predicting progressively refined image resolutions.

Training

The training of VAR models involves leveraging large datasets such as ImageNet-1k. The models are designed to scale effectively with the size of the data, following patterns observed in large language models.

Guide: Running Locally

1. Clone the Repository:
   Download the VAR repository from GitHub.

2. Install Dependencies:
   Ensure all required packages and dependencies are installed. This typically involves setting up a Python environment with libraries specified in the repository.

3. Download Pre-trained Models:
   Access and download the pre-trained checkpoints hosted in the repository.

4. Run the Model:
   Execute scripts to run the VAR model locally, utilizing your local GPU for computation.

5. Cloud GPUs:
   For enhanced performance, consider using cloud-based GPU services such as AWS, Google Cloud, or Azure to run the model.

License

The VAR framework is distributed under the MIT License, allowing for extensive reuse and modification in both personal and commercial projects.