Reproduce “QuantVSR: Low-Bit Post-Training Quantization for Real-World Video 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
Bowen Chai, Zheng Chen, Libo Zhu, Wenbo Li, Yong Guo, Yulun Zhang
QuantVSR is a low-bit post-training quantization framework for real-world video super-resolution, significantly reducing computational and memory requirements of diffusion models. It achieves over 80% reduction in parameters and operations while maintaining perceptual quality comparable to full-precision models across various synthetic and real-world datasets.
Key insights
- Adaptive precision allocation via the STCA mechanism, considering both spatial and temporal complexities, is effective in preserving critical information and maintaining VSR quality during low-bit quantization.
- Biased quantization errors are a significant source of accuracy loss in low-bit VSR quantization, and the LBA module effectively compensates for these errors with minimal inference overhead.
- It is feasible to achieve substantial compression (over 80% parameter and operation reduction) in diffusion-based VSR models while retaining high perceptual and structural quality comparable to full-precision models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.