MTIC: Mamba-Transformer Synergy with Shared Adaptive Context for Image Compression

Learned image compression (LIC) has emerged as a promising solution for efficient image storage and transmission. However, existing methods often struggle to simultaneously capture long-range global structural priors and preserve high-frequency local details within a unified framework. To address this inherent challenge, we introduce M amba- T ransformer Synergy with Shared Adaptive Context for I mage C ompression (MTIC). Our architecture fundamentally resolves the trade-off between representation power and computational complexity by synergistically allocating these two paradigms: deploying the highly capable but quadratic-complexity Transformer exclusively for local feature modeling, while utilizing the efficient, linear-complexity Mamba to establish global structural context. We extend this synergy to entropy modeling by proposing the Mamba-Transformer based Group-wise Autoregressive (MT-GAR) model. Distinct from recent separate-network approaches that suffer from linear parameter bloat, MT-GAR establishes a shared adaptive context by operating within a highly efficient shared-network paradigm. Within MT-GAR, the Group-wise Causal Selective Scan (GCSS) and the Group-wise Causal Attention Mixer (GCAM) are designed to model the global and local inter-/intra-group dependencies, respectively. To maintain distributional flexibility within this shared architecture, we uniquely introduce Multiple Low-Rank Experts (MLRE), which adapt to group-specific distributions with minimal parameter overhead. Furthermore, we implement a state cache mechanism during autoregressive inference to effectively avoid redundant feature re-computation. Extensive evaluations on the Kodak, CLIC, and Tecnick datasets demonstrate that MTIC achieves remarkable BD-Rate reductions of 17.85%, 20.66%, and 23.13% over VTM-17.0 Intra, surpassing state-of-the-art methods while offering a highly competitive and flexible trade-off between rate-distortion performance and architectural complexity.