Content-Adaptive Mamba LIC (CMIC)

• The paper demonstrates that CMIC integrates content-adaptive token reorganization and global-priors prompting, reducing BD-rate by up to -21.34% on benchmarks like Tecnick.
• The methodology replaces standard blocks within a variational autoencoder compression pipeline with CAM blocks that dynamically reorder tokens to enhance global dependency modeling.
• Empirical results show CMIC achieves lower decoding latency and GPU memory usage, outperforming traditional MambaIC systems despite minimal training overhead.

Searching arXiv for the cited CMIC-related papers to ground the article in current literature. CMIC, short for Content-Adaptive Mamba for Learned Image Compression, is a learned image compression (LIC) model built around Content-Adaptive Mamba (CAM), a dynamic state-space model (SSM) designed to address two limitations of vanilla Mamba in image compression: fixed and predefined selective scans, and the restricted ability to exploit content dependencies under strict causality and long-range decay. In CMIC, CAM introduces content-aware token reorganization and global priors via a prompt dictionary, with the stated goal of better capturing global dependencies while preserving computational efficiency. Within the reported evaluation, CMIC surpasses VTM-21.0 by $-15.91\%$, $-21.34\%$, and $-17.58\%$ BD-rate on Kodak, Tecnick, and CLIC benchmarks, respectively (Chen et al., 4 Aug 2025).

1. Variational codec structure and placement of CAM

CMIC follows the standard variational-autoencoder-based learned image compression pipeline. The analysis transform is

$\mathbf{y} = g_a(\mathbf{x};\theta_a),$

followed by scalar quantization

$$\hat{\mathbf y} = Q(\mathbf{y} - \boldsymbol\mu) + \boldsymbol\mu.$$

The synthesis transform reconstructs the image by

$\hat{\mathbf x} = g_s(\hat{\mathbf y};\theta_s).$

The hyperprior path computes

$\mathbf{z} = h_a(\mathbf{y};\phi_a),$

quantizes $\mathbf z$ to $\hat{\mathbf z}$, and decodes the side information through

$$(\boldsymbol\mu,\boldsymbol\sigma) = h_s(\hat{\mathbf z};\phi_s).$$

For entropy modeling, CMIC uses a Gaussian likelihood,

$-21.34\%$0

with spatial context $-21.34\%$1. The rate–distortion objective is

$-21.34\%$2

The architectural intervention is explicit: each stage of the nonlinear transform (six stages overall) replaces standard blocks with a CAM block. A CAM block wraps a 1-D Mamba state-space model with two added components: Content-Adaptive Token Aggregation and Global-Priors Prompting. Vanilla Mamba is described as processing a fixed, causal 1-D scan of the flattened 2-D feature map, which leads to long-range decay and ignores content similarity; CAM extends that design by dynamically reordering tokens so that similar tokens become neighbors in the scan and by injecting a small, learnable prompt per token conditioned on global latent priors (Chen et al., 4 Aug 2025).

2. Content-aware token reorganization

A central mechanism in CAM is content-aware token reorganization, which clusters and reorders tokens based on content similarity so that proximity in feature space is prioritized over Euclidean space. Let the flattened tokens be $-21.34\%$3, and let $-21.34\%$4 be learnable, normalized cluster centers. In each training step, CAM runs $-21.34\%$5 iterations of cosine-based K-Means: $-21.34\%$6 then assigns each token by

$-21.34\%$7

recomputes provisional centers

$-21.34\%$8

and updates the centers through EMA: $-21.34\%$9

After training-phase updates, and also at inference, CAM computes $-17.58\%$0 and builds a permutation

$-17.58\%$1

that groups tokens by cluster index in ascending order. The reordered sequence is

$-17.58\%$2

and its inverse $-17.58\%$3 is cached to restore spatial layout after scanning.

The intended effect is twofold. First, by scanning tokens in cluster order, CAM privileges content proximity over Euclidean adjacency, which strengthens semantic interactions. Second, the original strict causal ordering within each cluster is preserved, even though clusters may be revisited out of spatial order. A common misunderstanding is to read the reorganization as abandoning causality altogether; the formulation instead remains 1-D causal within each cluster while attempting to capture non-local dependencies with linear scan complexity (Chen et al., 4 Aug 2025).

3. Global-priors prompting inside the Mamba recurrence

CAM’s second added mechanism is Global-Priors Prompting (GPP). The model learns a dictionary

$-17.58\%$4

where $-17.58\%$5 is the SSM state dimension. For a batch of $-17.58\%$6 tokens, it constructs a one-hot membership matrix

$-17.58\%$7

Instance-specific prompts are then assembled as

$-17.58\%$8

so that $-17.58\%$9 is the prompt for token $\mathbf{y} = g_a(\mathbf{x};\theta_a),$0. These vectors are described as capturing global category priors.

The baseline discrete Mamba recurrence is

$\mathbf{y} = g_a(\mathbf{x};\theta_a),$1

where $\mathbf{y} = g_a(\mathbf{x};\theta_a),$2 and $\mathbf{y} = g_a(\mathbf{x};\theta_a),$3 derive from continuous-to-discrete conversion of $\mathbf{y} = g_a(\mathbf{x};\theta_a),$4. CAM injects $\mathbf{y} = g_a(\mathbf{x};\theta_a),$5 by modulating the read-out matrix: $\mathbf{y} = g_a(\mathbf{x};\theta_a),$6 Equivalently,

$\mathbf{y} = g_a(\mathbf{x};\theta_a),$7

The stated interpretation is that the SSM thereby “queries” hidden state via prompts, enabling information flow from unscanned (future in spatial order) tokens. This is presented as a way to relax the strict causality and long-range decay in token interactions of vanilla Mamba. A plausible implication is that CAM does not discard Mamba’s recurrent structure, but instead changes what information the recurrent state is asked to expose at read-out time (Chen et al., 4 Aug 2025).

4. Complexity, latency, and memory characteristics

The theoretical complexity of the vanilla Mamba 1-D scan is given as

$\mathbf{y} = g_a(\mathbf{x};\theta_a),$8

per block, which is linear in token length $\mathbf{y} = g_a(\mathbf{x};\theta_a),$9. CAM adds three overheads: K-Means clustering with complexity $$\hat{\mathbf y} = Q(\mathbf{y} - \boldsymbol\mu) + \boldsymbol\mu.$$0, token re-ordering with complexity $$\hat{\mathbf y} = Q(\mathbf{y} - \boldsymbol\mu) + \boldsymbol\mu.$$1, and prompt lookup with complexity $$\hat{\mathbf y} = Q(\mathbf{y} - \boldsymbol\mu) + \boldsymbol\mu.$$2. The overall complexity therefore remains $$\hat{\mathbf y} = Q(\mathbf{y} - \boldsymbol\mu) + \boldsymbol\mu.$$3 in the number of tokens. By contrast, the paper states that full 2-D multi-directional scans would be $$\hat{\mathbf y} = Q(\mathbf{y} - \boldsymbol\mu) + \boldsymbol\mu.$$4 or $$\hat{\mathbf y} = Q(\mathbf{y} - \boldsymbol\mu) + \boldsymbol\mu.$$5 (Chen et al., 4 Aug 2025).

The empirical efficiency claims are specific. Training overhead from K-Means is reported as approximately $$\hat{\mathbf y} = Q(\mathbf{y} - \boldsymbol\mu) + \boldsymbol\mu.$$6 per step. Inference overhead is reported as negligible extra latency for reordering and prompt gathering. For 2K images, the reported decoding latency and peak GPU memory are as follows:

System	Reported value
CMIC decoding latency	0.405 s
MambaIC (quad-scan) decoding latency	0.669 s
CMIC peak GPU memory	4.44 GB
MambaIC peak GPU memory	20.32 GB

These numbers are used to support the claim that the content-adaptive additions are lightweight relative to multi-directional scan alternatives. This suggests that CAM is positioned not merely as an accuracy modification, but also as a reformulation of how global context is acquired under a linear-time budget (Chen et al., 4 Aug 2025).

5. Objective function, hyperparameters, and optimization

The training objective is

$$\hat{\mathbf y} = Q(\mathbf{y} - \boldsymbol\mu) + \boldsymbol\mu.$$7

with

$$\hat{\mathbf y} = Q(\mathbf{y} - \boldsymbol\mu) + \boldsymbol\mu.$$8

and distortion

$$\hat{\mathbf y} = Q(\mathbf{y} - \boldsymbol\mu) + \boldsymbol\mu.$$9

Two implementation choices are emphasized. First, there is no extra clustering loss; the cosine K-Means is fully unsupervised and updates centers via EMA. Second, the prompt dictionary is learned end-to-end via standard backprop through the SSM.

The reported training and model configuration is concrete. The dataset is Flickr2W for training, with evaluation on Kodak (768×512), Tecnick (1200×1200), and CLIC (2K). Optimization uses Adam with lr=$\hat{\mathbf x} = g_s(\hat{\mathbf y};\theta_s).$0. The $\hat{\mathbf x} = g_s(\hat{\mathbf y};\theta_s).$1-values are

$\hat{\mathbf x} = g_s(\hat{\mathbf y};\theta_s).$2

for MSE models and

$\hat{\mathbf x} = g_s(\hat{\mathbf y};\theta_s).$3

for MS-SSIM models. The transform channel dimensions are $\hat{\mathbf x} = g_s(\hat{\mathbf y};\theta_s).$4, the block depths are $\hat{\mathbf x} = g_s(\hat{\mathbf y};\theta_s).$5, and the clustering hyperparameters are $\hat{\mathbf x} = g_s(\hat{\mathbf y};\theta_s).$6, $\hat{\mathbf x} = g_s(\hat{\mathbf y};\theta_s).$7, and $\hat{\mathbf x} = g_s(\hat{\mathbf y};\theta_s).$8 (Chen et al., 4 Aug 2025).

6. Rate–distortion performance, ablations, and qualitative behavior

The primary quantitative result is the BD-rate relative to VTM-21.0:

Benchmark	BD-rate vs VTM-21.0
Kodak	$\hat{\mathbf x} = g_s(\hat{\mathbf y};\theta_s).$9
Tecnick	$\mathbf{z} = h_a(\mathbf{y};\phi_a),$0
CLIC	$\mathbf{z} = h_a(\mathbf{y};\phi_a),$1

The ablation study isolates the two CAM components. Relative to a baseline with no CTA and no GPP, the CTA-only variant yields approximately $\mathbf{z} = h_a(\mathbf{y};\phi_a),$2 on Kodak, Tecnick, and CLIC; the GPP-only variant yields $\mathbf{z} = h_a(\mathbf{y};\phi_a),$3; and the full CTA + GPP model yields $\mathbf{z} = h_a(\mathbf{y};\phi_a),$4. Within the paper’s framing, this indicates that the two mechanisms are complementary rather than redundant (Chen et al., 4 Aug 2025).

The qualitative observations are also specific. Effective Receptive Field plots show that CMIC’s analysis network aggregates context from across the full image rather than only local neighborhoods. Clustering visualizations are reported to show that semantically similar regions, such as red doors, sky, and textures, are grouped into the same cluster and scanned contiguously, reducing cross-cluster redundancy. Reconstructed examples are described as exhibiting crisper edges and fewer blocky artifacts in textured regions than both vanilla Mamba-based and CNN/Transformer baselines.

A common misconception is that the content-adaptive machinery only changes token ordering while leaving the entropy-relevant representation unchanged. The reported ablations argue against that reading: both token aggregation and prompt-conditioned read-out contribute measurable BD-rate gains, and the full model improves beyond either component alone (Chen et al., 4 Aug 2025).

7. Relation to adjacent Mamba-based LIC systems and acronym ambiguity

CMIC emerged within a broader 2025 line of Mamba-based learned image compression. CMamba introduced a hybrid convolution-and-SSM framework with a Content-Adaptive SSM (CA-SSM) module and a Context-Aware Entropy (CAE) module, motivated by the observation that SSMs model overall content well while CNNs capture local details (Wu et al., 7 Feb 2025). MambaIC emphasized refined context modeling and window-based local attention in channel-spatial entropy modeling, and reported strong efficiency-performance trade-offs for high-resolution image compression (Zeng et al., 16 Mar 2025). Relative to these systems, CMIC’s distinguishing contribution is not hybrid CNN–SSM fusion or local attention, but content-aware token reordering and prompt dictionary conditioning within the Mamba SSM itself (Chen et al., 4 Aug 2025).

There is also an acronym issue in the surrounding literature summary. The provided material for CDMamba, a remote sensing image binary change detection model, labels that system as a Content-Adaptive Mamba-based Change-Detection model (CMIC, also called CDMamba) (Zhang et al., 2024). This suggests that the acronym CMIC is not unique across domains. In the image-compression literature, however, CMIC denotes “Content-Adaptive Mamba for Learned Image Compression” (Chen et al., 4 Aug 2025).

Taken together, these comparisons place CMIC within a family of SSM-based codecs that seek global receptive fields with linear complexity, while clarifying that its method is specifically centered on reorganizing the scan order by content similarity and augmenting the Mamba recurrence with global-prior prompts.