Reproduce “MOWA: Multiple-in-One Image Warping Model” 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
Kang Liao, Zongsheng Yue, Zhonghua Wu, Chen Change Loy
Researchers from Nanyang Technological University and SenseTime Research developed MOWA, a unified "Multiple-in-One" image warping model capable of performing six diverse computational photography tasks within a single framework. The model achieves competitive or superior performance compared to task-specific state-of-the-art methods while maintaining an efficient model size and demonstrating strong generalization capabilities to unseen scenarios.
Key insights
- Disentangling motion estimation into hierarchical components enables a single model to effectively approximate diverse and complex distortions while preserving image details.
- A lightweight, point-based task classifier combined with prompt learning can mitigate performance conflicts in multi-task learning, allowing the model to adapt its behavior for different warping requirements.
- A unified multi-task learning framework can learn generalized motion representations, leading to robust performance and the ability to generalize to unseen domains and even zero-shot tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.