MBMamba: Structure-Aware Image Deblurring

• The paper introduces MBMamba, a deblurring model that combines a Mamba backbone with memory buffering and an Ising-inspired regularizer to preserve historical features and maintain spatial coherence.
• The architecture employs a frozen encoder with progressive sub-decoders and a MemVSSM module that processes image chunks via feature cross-attention to reduce channel redundancy.
• Empirical results demonstrate that MBMamba-L outperforms previous methods with higher PSNR on benchmarks and up to 84.2% computational cost reduction compared to MambaIRv2.

MBMamba is a structure-aware image deblurring network that combines a Mamba backbone with a memory buffer mechanism and an Ising-inspired regularization loss in order to preserve historical features and maintain image coherence without changing the original Mamba architecture (Gao et al., 17 Aug 2025). In the most precise usage, the term refers to the model introduced in "MBMamba: When Memory Buffer Meets Mamba for Structure-Aware Image Deblurring" (Gao et al., 17 Aug 2025). In adjacent arXiv usage, however, related expressions such as "MBMamba-style" or implicit MBMamba usage are also applied to broader Mamba-based system design, including wireless intelligence and multimodal models, so the label is not uniformly standardized across subfields (Zhang et al., 1 Aug 2025).

1. Terminological scope

In exact nomenclature, MBMamba is the title of a 2025 image deblurring method that addresses a specific weakness of vision Mamba models: the tendency of flatten-and-scan processing to forget local pixel detail and to produce channel redundancy (Gao et al., 17 Aug 2025). The same string is used less rigidly elsewhere. In "Mamba for Wireless Communications and Networking: Principles and Opportunities," “MBMamba” essentially refers to using Mamba as a core building block for wireless intelligence, more precisely Mamba-enabled wireless system design (Zhang et al., 1 Aug 2025). In "BlackMamba: Mixture of Experts for State-Space Models," the architecture is described as a Mamba-MoE hybrid and can be viewed as an MBMamba-style architecture because it preserves a state-space sequence core while adding sparse conditional computation (Anthony et al., 2024).

The multimodal literature exhibits the same looseness. "FusionMamba" is described as closely related in spirit to an “MBMamba”-style idea, but the paper does not use the term MBMamba itself (Xie et al., 2024). "ML-Mamba" is presented as highly relevant to MBMamba or Mamba-based multimodal modeling because it replaces a Transformer backbone with Mamba-2 and introduces a connector for 2D visual scanning (Huang et al., 2024). "CAF-Mamba" similarly builds a multimodal depression detection pipeline around ResMamba blocks, but again under its own model name (Zhou et al., 29 Jan 2026).

This suggests that MBMamba functions both as a proper model name and as a broader descriptive label for Mamba-centered hybridization. In encyclopedia usage, the primary referent is the deblurring model unless the context clearly indicates a looser, family-level meaning.

2. Problem setting and motivation

The deblurring formulation of MBMamba starts from a critique of vanilla vision Mamba. The paper argues that image deblurring requires simultaneous access to long-range context and fine local structure, whereas the usual flatten-and-scan strategy often leads to local pixel forgetting and channel redundancy (Gao et al., 17 Aug 2025). Local pixel forgetting weakens 2D structural modeling, while channel redundancy emerges when all image features are fed into the VSSM at once. Both effects are detrimental in restoration tasks in which sharp reconstruction depends on edge continuity, neighborhood consistency, and preservation of small structures.

The paper positions MBMamba against two existing response patterns in vision Mamba research. One pattern modifies the scan strategy; another inserts additional local feature modules. According to the paper, both approaches increase computational complexity and hinder real-time performance (Gao et al., 17 Aug 2025). MBMamba is designed to avoid that trade-off. It keeps the original Mamba computation intact and adds external mechanisms for feature reuse and structural regularization.

This design choice distinguishes MBMamba from other vision-state-space adaptations that directly alter scanning behavior. For example, MVSMamba for multi-view stereo introduces a reference-centered dynamic scanning strategy and four directional skip scans to tailor Mamba to reference-source feature interaction (Jiang et al., 3 Nov 2025). MBMamba takes the opposite stance: structure-awareness should be recovered without redesigning the scan rule.

3. Progressive architecture

MBMamba uses a frozen encoder + multiple sub-decoders design (Gao et al., 17 Aug 2025). Given a blurred image $\mathbf{I}\in\mathbb{R}^{H\times W\times 3}$, the model first extracts shallow features $\mathbf{F}\in\mathbb{R}^{H\times W\times C}$ and encodes them into multi-scale features $e_i$ by a pre-trained encoder. Reconstruction is progressive rather than one-shot: the decoder contains $n$ sub-decoders, each with four decoding stages, and the output of each sub-decoder is fed into the next one. The final decoder output $\mathbf{X}_n$ is added to the degraded image to form the restored image (Gao et al., 17 Aug 2025).

The central decoder block is MemVSSM (Vision State Space Model with memory buffering). At the block level, MBMamba applies normalization and FFN transformations around MemVSSM with residual addition. The important point is architectural rather than algebraic: MemVSSM is the location where historical feature preservation is introduced, while the outer decoder preserves the progressive reconstruction scheme.

The encoder is pre-trained and then frozen. The paper reports that pre-training improves performance and that freezing the encoder has minimal impact, so the encoder is frozen to save computation (Gao et al., 17 Aug 2025). This is a consequential implementation choice because MBMamba’s efficiency claims are tied not only to MemVSSM itself, but also to avoiding unnecessary encoder updates.

4. Memory buffering and feature fusion

MemVSSM operates by splitting the incoming feature map into channel chunks and processing them sequentially (Gao et al., 17 Aug 2025). If the feature tensor is denoted by $F$, then chunking yields

$F^d_1,\ldots,F^d_i,\ldots,F^d_N = Chunk(F).$

Each chunk is then processed by Mamba, rather than sending the full feature tensor through the VSSM in a single pass. The paper’s motivation is that chunk-wise processing reduces the burden associated with excessive hidden states while making it possible to preserve adjacent historical information through an explicit memory mechanism.

The model maintains a memory bank of depth $K$ that stores the most recent enhanced feature outputs in first-in, first-out (FIFO) order. For chunk $i$, the previous $K$ enhanced features are fetched and fused with the current chunk output:

$\mathbf{F}\in\mathbb{R}^{H\times W\times C}$0

After fusion, the memory bank pops the oldest feature and pushes the new one, and all features in the bank are concatenated to form the output representation (Gao et al., 17 Aug 2025). The mechanism is explicitly historical: current processing is conditioned on a short rolling window of previously enhanced features.

The fusion operator is the Feature Cross-Attention Mechanism (FCAM). The paper describes FCAM as an attention-based comparison between current and historical features. In the simplified one-slot example, queries, keys, and values are generated from the current feature $\mathbf{F}\in\mathbb{R}^{H\times W\times C}$1 and the previous enhanced feature $\mathbf{F}\in\mathbb{R}^{H\times W\times C}$2, and the final fused feature is the sum of bidirectional attention outputs together with residual contributions from both features (Gao et al., 17 Aug 2025). The intended effect is to allow the current chunk to be refined by explicit comparison against prior chunk outputs.

A recurrent misconception is that MBMamba is a scan-modified Mamba. The paper states the opposite: MBMamba remains structure-aware without altering the original Mamba architecture because Mamba still handles long-range modeling inside each chunk, while the memory buffer + FCAM re-injects local and historical cues, and the Ising loss adds a spatial coherence prior at the output level (Gao et al., 17 Aug 2025).

5. Ising-inspired regularization and optimization

The second defining component is the Ising-inspired regularization loss, introduced to counter outputs that become too smooth or fragmented under chunk-wise Mamba processing (Gao et al., 17 Aug 2025). The loss encourages neighboring pixels to be similar, analogous to local attraction in the Ising model. Its form is

$\mathbf{F}\in\mathbb{R}^{H\times W\times C}$3

Here, $\mathbf{F}\in\mathbb{R}^{H\times W\times C}$4 is the local neighborhood of pixel $\mathbf{F}\in\mathbb{R}^{H\times W\times C}$5 and $\mathbf{F}\in\mathbb{R}^{H\times W\times C}$6 (Gao et al., 17 Aug 2025). The paper interprets this as a mechanism for preserving structure and coherence by encouraging “mutual attraction” between adjacent pixels.

The full training objective combines the Ising regularizer with three standard restoration losses:

$\mathbf{F}\in\mathbb{R}^{H\times W\times C}$7

The component losses are a Charbonnier loss for pixel-level reconstruction, an edge loss via Laplacian comparison, and a frequency loss via FFT-domain matching, with $\mathbf{F}\in\mathbb{R}^{H\times W\times C}$8, $\mathbf{F}\in\mathbb{R}^{H\times W\times C}$9, and $e_i$0 (Gao et al., 17 Aug 2025).

The training setup is specified in standard restoration terms. MBMamba is trained on GoPro with 2,103 training pairs and 1,111 evaluation pairs, then directly tested on HIDE with 2,025 images and RealBlur with the RealBlur-R and RealBlur-J subsets (Gao et al., 17 Aug 2025). Optimization uses Adam with $e_i$1 and $e_i$2, an initial learning rate of $e_i$3, cosine decay to $e_i$4, patch size $e_i$5, batch size $e_i$6, and $e_i$7 iterations, with horizontal and vertical flipping for augmentation (Gao et al., 17 Aug 2025). The model family includes MBMamba-S with 1 sub-decoder, MBMamba-B with 2 sub-decoders, and MBMamba-L with 4 sub-decoders.

6. Empirical results and efficiency

The paper evaluates MBMamba against representative CNN, Transformer, and Mamba-based methods including MPRNet, NAFNet, Restormer, FSNet, DeblurDiNAT, MambaIR, ALGNet, LoFormer, XYScanNet, PGDN, MDT, ACL, Omni-Deblurring, and MambaIRv2, using PSNR and SSIM as metrics (Gao et al., 17 Aug 2025). The reported quantitative results position MBMamba-L as the strongest of the MBMamba variants across the principal benchmarks.

Dataset	Representative comparison	MBMamba result
GoPro	MambaIRv2: 33.62 / 0.967	MBMamba-L: 34.68 / 0.972
HIDE	LoFormer: 31.86 / 0.949	MBMamba-L: 32.22 / 0.953
RealBlur-R	ALGNet: 41.16 / 0.981	MBMamba-L: 41.21 / 0.981
RealBlur-J	ALGNet: 32.94 / 0.946	MBMamba-L: 33.41 / 0.953

On GoPro, the reported sequence is MBMamba-S: 33.89 / 0.968, MBMamba-B: 34.33 / 0.969, and MBMamba-L: 34.68 / 0.972 (Gao et al., 17 Aug 2025). The paper highlights that MBMamba-L improves over the previous best MDT by 0.42 dB on GoPro and over MambaIRv2 by 1.06 dB. On HIDE, MBMamba-L reaches 32.22 / 0.953, which the paper notes is a 0.36 dB gain over LoFormer despite training only on GoPro. On RealBlur-J, the paper reports a 0.47 dB PSNR gain over ALGNet, whereas the gain on RealBlur-R is described as small (Gao et al., 17 Aug 2025).

The efficiency claims are also explicit. On GoPro, MBMamba-S requires 0.249 s and 54 GFLOPs, MBMamba-B requires 0.283 s and 74 GFLOPs, and MBMamba-L requires 0.442 s and 105 GFLOPs; the comparison point MambaIRv2 requires 0.743 s and 664 GFLOPs (Gao et al., 17 Aug 2025). The paper states that MBMamba-L achieves up to 84.2% computational cost reduction and nearly 1.5× faster inference versus the previous best Mamba baseline.

Ablation evidence attributes most of the improvement to the memory-buffered decoder. Using NAFNet as baseline, the paper reports Baseline: 32.83, Baseline + MemVSSM: 33.64, Baseline + Ising loss: 32.85, and Baseline + MemVSSM + Ising loss: 33.89 (Gao et al., 17 Aug 2025). This indicates that MemVSSM provides the major gain, while the Ising loss adds complementary improvement of about 0.25 dB when combined with MemVSSM. The MemVSSM design also outperforms two alternatives, with VSSM: 33.37, ASSM: 33.65, and MemVSSM: 33.89. The paper further presents channel activation statistics and argues that MemVSSM reduces inactive channels compared to standard SSMs.

7. Position within the broader Mamba literature

Within the larger Mamba ecosystem, MBMamba belongs to a class of task-specific systems that preserve Mamba’s linear-time sequence modeling while compensating for weaknesses that emerge when Mamba is ported outside its original sequence domains. In deblurring, the compensation is historical feature reuse and output regularization (Gao et al., 17 Aug 2025). In multi-view stereo, MVSMamba instead uses a Dynamic Mamba module and a reference-centered dynamic scanning strategy to achieve efficient intra-view and inter-view feature interaction, omnidirectional multi-view representations, and multi-scale global feature aggregation (Jiang et al., 3 Nov 2025). In multimodal image fusion, FusionMamba augments visual Mamba with dynamic convolution, channel attention, and a Dynamic Feature Fusion Module consisting of DFEM and CMFM (Xie et al., 2024).

The same pattern appears outside vision restoration. In multimodal depression detection, CAF-Mamba organizes the pipeline into UEM, CIME, and AAMFM, with ResMamba blocks used for unimodal encoding, explicit intermodal interaction, and implicit higher-order fusion (Zhou et al., 29 Jan 2026). In multimodal LLM design, ML-Mamba replaces the Transformer backbone with Mamba-2 and introduces the Mamba-2 Scan Connector (MSC) to convert 2D visual features into a form compatible with 1D causal sequence processing (Huang et al., 2024). In language modeling, BlackMamba combines a Mamba selective state-space backbone with routed MoE blocks to obtain linear-complexity generation together with sparse expert computation (Anthony et al., 2024). In wireless communications, Mamba-enabled wireless system design is framed through two application frameworks: replacement of traditional algorithms and enabler of novel paradigms (Zhang et al., 1 Aug 2025).

A plausible implication is that MBMamba should be understood less as a single architectural template than as a member of a broader research direction: retaining Mamba’s sequence efficiency while introducing narrowly targeted mechanisms that restore locality, cross-modal interaction, structural bias, or domain-specific inductive constraints. In that sense, the exact deblurring model MBMamba is representative precisely because it does not claim universal optimality; rather, it demonstrates one concrete route for making Mamba structure-aware in a domain where 2D coherence is central (Gao et al., 17 Aug 2025).