Visual AutoRegressive (VAR) modeling has redefined image generation through ‘next-scale prediction’ in autoregressive transformers, dramatically outshining diffusion transformers.

• Superior image generation capabilities exceed diffusion transformer benchmarks.
• Instantaneous improvements include Frechet inception distance (FID), inception score (IS), and a 20x faster inference speed on ImageNet 256x256.
• Demonstrates striking power-law scaling laws with solid evidence presented.
• Showcases impressive zero-shot task generalization abilities in various image processing tasks.

The transformative qualities of VAR models position them to emulate significant traits of LLMs: Scaling Laws and Task Generalization. Their ability to perform in image in-painting, out-painting, and editing tasks, without prior training, marks a cornerstone for future unified learning explorations in visual generation. Access the study.