Research map / Architectures and efficiency / Attention and transformers
MambaIC: State Space Models for High-Performance Learned Image Compression
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MambaIC, a learned image compression model that uses visual state space (SSM) blocks in the nonlinear transform and channel-spatial context entropy model, plus window-based local attention, to improve rate-distortion and efficiency especially at high resolution.
Problem
Learned image compression faces a trade-off between rate-distortion performance and computational efficiency; Transformer-based methods have quadratic complexity, and prior SSM-based compression (MambaVC) merely swaps in Mamba blocks without adaptation.
Approach
Integrates visual state space (SS2D) blocks into both the nonlinear transform and the context entropy model, combining channel-wise and checkerboard spatial autoregression with window-based local attention after parameter aggregation to recover local redundancy that global SSM scans miss.
Key results
- BD-rate of -12.52% vs VVC on Kodak, versus -9.65% for Mixed and -7.31% for MambaVC.
- Decoding latency 39.42 ms vs 107.24 ms for Mixed, about 30% of Mixed's decoding time, with 202.04 GFLOPs.
- Window-based local attention with 8x8 windows gives 5.95% bitstream saving over the global-only variant on Kodak at PSNR about 34.2 dB.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use MambaIC instead of VVC | image compression · Kodak | BD-rate (PSNR): 0.00 -> -12.52% | Helped |
| add channel-wise autoregressive modeling (ablation) | MambaIC · Kodak | BD-rate: -6.73% -> -12.52% | Helped |
| add spatial context modeling (ablation) | MambaIC · Kodak | BD-rate: -8.54 -> -12.52% | Helped |
| add window-based local attention (ablation) | MambaIC · Kodak | BD-rate: -9.17% -> -12.52% | Helped |
| replace CNN blocks with SSM blocks (ablation) | MambaIC · Kodak | BD-rate: -3.81% -> -12.52% | Helped |
| replace Transformer blocks with SSM blocks (ablation) | MambaIC · Kodak | BD-rate: -7.19% -> -12.52% | Helped |
| use an 8×8 local-attention window rather than 6×6 (ablation) | MambaIC · Kodak | bpp at approximately 34.2 dB PSNR: 0.4428 -> 0.4404 | Helped |
| use an 8×8 local-attention window rather than 10×10 (ablation) | MambaIC · Kodak | bpp at approximately 34.2 dB PSNR: 0.4417 -> 0.4404 | Helped |
Surprising findings
- The CNN-based foundation block variant achieves lower decoding latency (35.53 ms) than the SSM variant, though with much worse BD-rate (-3.81%).
Limitations
- Encoding latency (60.73 ms) is higher than ELIC (40.76 ms) and the SSM block is slower than CNN blocks.
- Window size and design choices are tuned on Kodak-style evaluation; the high-resolution gains are shown mainly through BD-rate curves.
- Training only on Flickr30k at a limited set of lambdas.
Experimental setup
Datasets: Flickr30k
Benchmarks: Kodak Tecnick CLIC Professional Valid
Metrics: PSNR MS-SSIM bpp BD-rate (vs VVC) encoding/decoding latency FLOPs
Baselines: ELIC Mixed MLIC Entroformer Contextformer MambaVC BPG VTM
Builds on: Mamba VMamba (visual state space block) ELIC Minnen hyperprior checkerboard context model
Closest papers by meaning
- Linear Attention Modeling for Learned Image Compression (CVPR 2025 · similarity 0.86)
- Demystify Mamba in Vision: A Linear Attention Perspective (NeurIPS 2024 · similarity 0.86)
- Mamba-Adaptor: State Space Model Adaptor for Visual Recognition (CVPR 2025 · similarity 0.86)
- Spatial-Mamba: Effective Visual State Space Models via Structure-Aware State Fusion (ICLR 2025 · similarity 0.86)
- LongMamba: Enhancing Mamba's Long-Context Capabilities via Training-Free Receptive Field Enlargement (ICLR 2025 · similarity 0.85)
- EfficientViM: Efficient Vision Mamba with Hidden State Mixer based State Space Duality (CVPR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.