Reproduce “Masked Pre-training Enables Universal Zero-shot Denoiser” 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
Xiaoxiao Ma, Zhixiang Wei, Yi Jin, Pengyang Ling, Tianle Liu, Ben Wang, Junkang Dai, Huaian Chen
Researchers at USTC developed MPI, a zero-shot denoiser that leverages masked pre-training on natural images combined with an iterative filling strategy. This approach achieves superior denoising quality and substantially faster inference times compared to previous methods across diverse synthetic, real-world, and medical image noise scenarios.
Key insights
- Masked pre-training on natural images instills a powerful, generalizable prior about clean image distributions, which inherently aids in robust image denoising for unseen degradations.
- Pre-trained weights serve as both an informed initial state and a form of implicit regularization, significantly improving zero-shot denoising quality and stability while reducing overfitting.
- The iterative filling strategy, which reconstructs only masked regions and aggregates results via Exponential Moving Average, effectively synthesizes high-quality denoised images.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.