OpenViSTAI · Agents · Blockchain · Crypto
Submit a project
Menu
Image & Video AI

Var

Var is an open-source implementation of visual autoregressive image generation that predicts images from coarse to fine through next-scale prediction, with model weights and an interactive text-to-image demo linked from its repository.

Explore the practical guide ↓

Independent OpenViST project profile, based on linked sources.

Var — OpenViST project profile in Image & Video AI.
OpenViST project illustration. See the linked sources for product details.
Understand the project

Practical overview

When it fits

Use the demo to try interactive text-to-image generation; use the repository when you need to inspect the implementation, work with model weights, or follow its documented training and evaluation workflow.

Who it is for

  • Researchers studying autoregressive visual generation and its scaling behavior.
  • Developers exploring image-generation implementations and model weights.
  • People who want to try the linked text-to-image demo.

Documented uses

  • Generate images interactively through the linked text-to-image demo.
  • Explore a next-scale prediction approach to autoregressive image generation.
  • Train Var models on ImageNet and evaluate generated samples using image-generation metrics.
Before you choose

Requirements & limits

The repository is public and declares an MIT license. It links to model weights and a separate interactive demo.

  • The central method is coarse-to-fine next-scale prediction, rather than conventional raster-scan next-token prediction.
  • The documented training workflow requires PyTorch 2.0 or later, the repository's Python packages, and an ImageNet dataset.
  • Training instructions cover distributed runs at 256×256 and 512×512; FlashAttention and xformers are optional speed optimizations.
  • The repository describes sampling and evaluation steps, but says a sampling script will be provided later.

Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.

Source-based practical guide

Prepare VAR for ImageNet training

Install the listed Python dependencies and organize ImageNet data for VAR training.

Before you start

  • Install torch version 2.0.0 or later.
  • Have the ImageNet dataset available, organized into train and val directories with class subdirectories.

Steps

  1. Install torch version 2.0.0 or later.

  2. Install the other Python packages with the command below.

    Documented command
    pip3 install -r requirements.txt

  3. Arrange ImageNet data under a dataset directory with train and val folders, each containing class subdirectories.

  4. Pass the dataset location to the training script using --data_path=/path/to/imagenet.

Expected outcome

VAR's listed dependencies are installed and the ImageNet data is arranged for use by a training script.

Guide limitations

  • This guide covers preparation only; it does not document a complete training invocation or execution.
  • Installing and compiling flash-attn and xformers is optional; the repository says they can speed up attention computation when installed.
Documentation 1 ↗
Evidence and transparency

Sources & review

Official repository ↗
Source record details
Source availability
Active
Sources checked
17 October 2026
Original website title
FoundationVision/VAR · Official homepage title · retained in its source language
How we review sourcesSuggest a correction