Reproduce “MambaIC: State Space Models for High-Performance Learned Image Compression” 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
Fanhu Zeng, Hao Tang, Yihua Shao, Siyu Chen, Ling Shao, Yan Wang
MambaIC introduces a learned image compression framework that integrates State Space Models (SSMs) into both the nonlinear transform and the context model, leveraging specially designed VSS blocks and 2D selective scanning. This approach achieves state-of-the-art rate-distortion performance with a -12.52% BD-rate gain over VVC on Kodak, while offering superior computational efficiency for high-resolution images.
Key insights
- State Space Models, when carefully adapted with 2D selective scanning and integrated comprehensively, can form an efficient and effective foundational block for learned image compression.
- Combining the global receptive field and efficiency of SSMs with targeted window-based local attention provides a synergistic approach to effectively model both global and local image redundancies.
- The thoughtful design of an SSM-based context model significantly improves the accuracy of probability estimations for latent symbols, leading to substantial improvements in bit rate and rate-distortion performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.