VARestorer: One-Step Distillation for Real-World Image Super-Resolution
A team of researchers has introduced VARestorer, a distillation framework designed to convert a pre-trained text-to-image VAR model into a streamlined one-step model for real-world image super-resolution (Real-ISR). This innovative method tackles the difficulties of adapting visual autoregressive models (VAR) for ISR, including the constraints of causal attention and the accumulation of errors during iterative predictions. By utilizing distribution matching, VARestorer removes the need for iterative refinement, thus minimizing error propagation and reducing inference time. Additionally, it employs pyramid image conditioning with cross-scale attention to boost performance, achieving leading results on benchmark datasets.
Key facts
- VARestorer is a distillation framework for one-step real-world image super-resolution.
- It transforms a pre-trained text-to-image VAR model into an ISR model.
- The method uses distribution matching to avoid iterative refinement.
- It addresses causal attention limitations and error accumulation in VAR.
- Pyramid image conditioning with cross-scale attention is introduced.
- The approach reduces inference time and error propagation.
- It achieves state-of-the-art results on benchmark datasets.
- The paper is available on arXiv under ID 2604.21450.
Entities
Institutions
- arXiv