FoundationVision/VAR
VAR
VAR is a visual autoregressive model and image generation framework that applies large language model scaling laws to visual data. It functions as an image generator that uses a coarse-to-fine next-scale prediction approach rather than traditional raster-scan tokenization.
The system utilizes scale-based tokenization to represent images as a hierarchy of discrete tokens. It generates high-resolution content by iteratively predicting the next resolution level, refining coarse predictions into fine-grained details.
The project covers a broad range of capabilities including autoregressive image generation, visual scaling laws research, and visual content sampling. It incorporates classifier-free guidance to balance sample quality and diversity during the generation process.
The training infrastructure includes automated state management and checkpoint-based resumption to maintain progress during large-scale training runs.
Features
- Coarse-to-Fine Generation - Generates images by iteratively increasing resolution through a sequence of increasingly detailed scale predictions.
- Autoregressive Visual Token Predictors - Implements a generative model that predicts images across multiple scales using visual tokens.
- Scale-Based Tokenization - Represents images as a hierarchy of discrete tokens corresponding to different resolution levels for autoregressive processing.
- Classifier-Free Guidance - Uses classifier-free guidance to balance image sample quality and diversity during the generation process.
- Generative Image Models - Provides a comprehensive framework for training and sampling image generation models using a coarse-to-fine approach.
- Autoregressive Image Generation - Trains and uses models that predict image tokens in a sequence to create new visual content.
- LLM-Based Generators - An image generation architecture that applies large language model scaling laws and autoregressive sampling to visual data.
- Next-Scale Prediction - Uses a coarse-to-fine resolution approach that predicts the next scale instead of standard raster-scan token prediction.
- Generative Model Training Tools - Provides a training system for autoregressive image generation models with automated state management.
- Large Scale Training - Manages the training process for generative models on massive image datasets with checkpointing and recovery.
- Visual Scaling Laws - Studies how increasing model size and data affects the quality of generated images using a scale-based approach.
- Scale-Based Image Samplers - Produces high-resolution images by iteratively refining coarse predictions into fine-grained details.
