Reproduce “VARestorer: One-Step VAR Distillation for Real-World Image Super-Resolution” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Yixuan Zhu, Shilin Ma, Haolin Wang, Ao Li, Yanzhe Jing, Yansong Tang, Lei Chen, Jiwen Lu and 1 more

VARESTORER introduces a one-step visual autoregressive (VAR) distillation framework that efficiently transforms pre-trained text-to-image VAR models into single-step super-resolution models for real-world images. The framework achieves high perceptual quality with outputs that are perceptually appealing and statistically consistent, performing inference in approximately 0.23 seconds per image, significantly faster than multi-step alternatives.

Key insights

  • Direct application of multi-step VAR models for restoration leads to artifacts and blur, underscoring the necessity of a one-step approach to mitigate error accumulation.
  • Distribution matching with KL divergence loss is crucial for the student model to generate diverse, high-quality, and realistic predictions aligned with the teacher's generative capabilities.
  • Replacing the original causal attention with full attention across scales enables effective bidirectional communication between different resolution levels, which is vital for coherent high-quality reconstruction.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.