Invariant Salient Channel Space in LIC

• ISCS is a parameter-guided organization of latent channels that distinguishes Salient-Core (SC) channels and Salient-Auxiliary (SA) channels for efficient image reconstruction.
• It leverages weight variance, pairwise cosine similarity, and bias magnitude to deterministically rank channels, avoiding costly, data-dependent ablation tests.
• Empirical results indicate that ISCS maintains competitive rate–distortion performance while significantly reducing context-model runtime through slice-parallel decoding.

Invariant Salient Channel Space (ISCS) is a term introduced for channel importance in learned image compression (LIC), specifically in pretrained VAE-based LIC models. In that setting, ISCS denotes “the latent structure discovered by analyzing the encoder’s final projection weights, where each Salient-Core (SC) channel acts as an anchor surrounded by Salient-Auxiliary (SA) channels that carry correlated or complementary information, e.g., textures, colors, fine details” (Wang et al., 21 Sep 2025). The defining claim is that latent channels are not equally important, and that important channels can be identified and organized from intrinsic pretrained model parameters rather than from costly, data-dependent per-channel ablation tests. The term “Invariant” refers to the fact that the discovery process depends only on intrinsic pretrained model parameters, not on a particular dataset or input image; the paper repeatedly states that these statistics are input-invariant and dataset-invariant (Wang et al., 21 Sep 2025).

1. Conceptual definition and motivation

The starting observation behind ISCS is that, across prior pruning, latent prediction, and decoder channel-drop studies, only a small subset of channels is consistently crucial for high-quality reconstruction, while many others contribute relatively little or are strongly redundant (Wang et al., 21 Sep 2025). This is presented as a practical opportunity for improving both rate–distortion efficiency and runtime efficiency.

The paper contrasts this view with the usual empirical approach to channel importance, namely per-channel ablation tests, in which a channel is removed and the degradation in reconstruction quality is measured. That procedure is described as computationally costly, data-dependent and effectively dataset-specific, and limited by treating channels independently while ignoring interdependencies (Wang et al., 21 Sep 2025). The paper’s Figure 1 is summarized as showing that most channels follow a roughly logarithmic relation between per-channel bitrate and reconstruction loss, but that a small set of outliers do not: they carry low bits yet matter a lot.

Within this framework, ISCS is not merely a ranking of channels. It is a structured organization of latent channels composed of Salient-Core (SC) channels, Salient-Auxiliary (SA) channels, and bias-dominated important channels (Wang et al., 21 Sep 2025). SC channels are the highly important anchors; SA channels are channels strongly related to a given SC through parameter correlation; and the bias-based category is used to explain low-bitrate, high-impact outliers.

A plausible implication is that ISCS should be understood as a parameter-discovered latent organization rather than as an activation-space manifold. The paper’s claims are tied specifically to pretrained encoder parameters and their induced channel ordering and grouping, not to a learned geometric latent space in the broader representation-learning sense (Wang et al., 21 Sep 2025).

2. Parameter-side identification of ISCS

ISCS is identified by analyzing the encoder’s final projection weights, that is, the final linear transformation in the analysis transform $g_a$ that produces latent channels $z_c$ (Wang et al., 21 Sep 2025). For each output channel $c$, the corresponding parameters are a kernel tensor

$W_c \in \mathbb{R}^{C_{\mathrm{in} \times K \times K},$

and optionally a scalar bias

$b_c.$

The paper uses three parameter-side statistics. The first is weight variance, which is the main saliency score for identifying SC channels: $\operatorname{Var}(W_c) = \frac{1}{K^2 C_{\mathrm{in} \sum_{i,j,k}\bigl(W_c[i,j,k]-\mu_c\bigr)^2,$ with channel-wise kernel mean

$\mu_c= \frac{1}{K^2 C_{\mathrm{in} \sum_{i,j,k} W_c[i,j,k].$

The interpretation given is that larger weight variance implies broader linear response and larger activation dynamic range, which tends to align with higher entropy and stronger rate–distortion contribution (Wang et al., 21 Sep 2025). Channels are first ranked by $\operatorname{Var}(W_c)$, and the top ones are selected as Salient-Core channels.

The second statistic is pairwise cosine similarity between channel kernels, used to identify correlated auxiliaries. If $\bar W_c$ denotes the flattened form of $W_c$, then similarity between channels $z_c$0 and $z_c$1 is

$z_c$2

For each SC channel $z_c$3, the top $z_c$4 non-SC channels with highest similarity are selected as its SA channels (Wang et al., 21 Sep 2025). This is the step that makes ISCS dependency-aware rather than a simple independent ranking.

The third statistic is bias magnitude: $z_c$5 This score is used to capture important channels that may not have high variance. The paper argues that large biases often encode input-invariant global priors, such as luminance or overall tone, and links such channels to the low-bitrate/high-impact outliers found in per-channel ablations (Wang et al., 21 Sep 2025).

The ranking and aggregation procedure is described as a one-time weight analysis:

1. Rank channels by variance score $z_c$6;
2. Select top channels as SCs;
3. For each SC channel $z_c$7, rank non-SC channels by cosine similarity $z_c$8;
4. Select top $z_c$9 correlated channels as its SAs;
5. Add further important channels using the bias score $c$0 (Wang et al., 21 Sep 2025).

3. Internal organization: SC channels, SA channels, and invariance

The paper reports a consistent pattern in pretrained LIC encoders: highly informative channels appear as compact correlated groups (Wang et al., 21 Sep 2025). This discovered organization is the substantive content of ISCS. Each such pattern contains one Salient-Core channel and several Salient-Auxiliary channels associated with it.

SC channels are identified mainly by high variance score and are described as the channels judged most important for reconstruction. Functionally, the paper states that SC channels “capture critical structural information,” “capture dominant structures,” and in visualizations clearly encode “prominent structural patterns” (Wang et al., 21 Sep 2025).

SA channels are defined relative to a specific SC. They are the channels with high cosine similarity to that SC’s kernel and are therefore its strongly correlated companions. Functionally, the paper states that SAs provide “correlated or complementary information,” often capturing “textures, colors, fine details,” and act as supporting channels around the SC anchor (Wang et al., 21 Sep 2025). The distinction is therefore not only about importance level but also about role: SCs are dominant structural anchors, whereas SAs provide complementary detail channels.

The strongest direct support for these roles comes from Figure 2, described as showing that the top-ranked SC channels “clearly capture prominent structural patterns,” their most correlated SA channels “provide complementary cues, such as fine visual details,” and bias-dominated channels “exhibit near-constant activations across different image inputs” (Wang et al., 21 Sep 2025). The paper further argues that these bias-dominated channels might be coded more cheaply in the future, potentially via a per-channel scalar rather than a full feature map.

The invariance claim is specific. ISCS is said to be input-invariant and dataset-invariant because it is discovered from pretrained parameters only, so the discovery itself is independent of the current input image or dataset (Wang et al., 21 Sep 2025). Cross-dataset consistency is supported by comparison between parameter-based channel importance and ablation-based importance on JPEG-AI and Kodak, where most channels with large reconstruction degradation also have large variance scores on both datasets. The highlighted outlier, channel #292, has only moderate variance, but very large bias score $c$1, and large degradation when removed, which the paper uses to support the claim that variance captures most salient channels while bias captures an additional invariant class of important channels (Wang et al., 21 Sep 2025).

This suggests that the word “Invariant” in ISCS refers to invariance of the discovery procedure and the statistics used for channel organization, not to invariance of latent activations under geometric transformations in the usual representation-learning sense.

4. Deterministic channel ordering, grouping, and slice scheduling

ISCS is operationalized through a parameter-guided and deterministic channel ordering and grouping strategy (Wang et al., 21 Sep 2025). It is parameter-guided because all decisions are derived from pretrained model parameters: $c$2 It is deterministic because, once these statistics are computed once, the latent channels are permuted in a fixed way.

The ordering logic is:

1. Identify SC channels from high variance;
2. For each SC, sort associated non-SC channels by descending cosine similarity;
3. Form an SC–SA sequence

$c$3

where $c$4 are SAs ordered by decreasing similarity to the SC;

1. Incorporate additional bias-dominated important channels (Wang et al., 21 Sep 2025).

After ordering, each SC–SA pattern has $c$5 channels total: one SC and $c$6 SA channels. The paper then partitions each SC–SA sequence into slices using a single-pass deterministic $c$7 policy. Let

$c$8

The sequence is partitioned into slices $c$9, each containing $W_c \in \mathbb{R}^{C_{\mathrm{in} \times K \times K},$0 channels, by

$W_c \in \mathbb{R}^{C_{\mathrm{in} \times K \times K},$1

The paper notes that the indexing notation appears slightly inconsistent at the bounds, but that the intended rule is clear from the text and Figure 3: if a channel has similarity rank $W_c \in \mathbb{R}^{C_{\mathrm{in} \times K \times K},$2, then it is assigned to slice

$W_c \in \mathbb{R}^{C_{\mathrm{in} \times K \times K},$3

The stated effect is to distribute channels of different similarity ranks across slices rather than placing the most similar ones all in one slice (Wang et al., 21 Sep 2025).

The paper argues that this grouping equalizes slice-wise conditional entropy, balances prediction difficulty, and avoids weak-context cases. It further states that, in ordinary LIC, channel order does not matter much when the context model considers all channels together, but that in the ISCS framework channels are grouped for parallel processing, which deliberately reduces context space. Under that reduced-context regime, arbitrary grouping hurts performance, whereas SC–SA grouping preserves performance much better (Wang et al., 21 Sep 2025).

A comparison of grouping strategies is reported in Table 3:

Grouping strategy	bpp	PSNR
correlation ascending	0.3758	31.82
correlation descending	0.3430	31.41
TSP-based	0.3419	31.50
$W_c \in \mathbb{R}^{C_{\mathrm{in} \times K \times K},$4	0.3114	31.63

According to the paper, the proposed deterministic ordering is therefore empirically better than straightforward similarity sorting (Wang et al., 21 Sep 2025).

5. Integration into learned image compression systems

The method is designed for pretrained VAE-based LIC pipelines with analysis transform $W_c \in \mathbb{R}^{C_{\mathrm{in} \times K \times K},$5, entropy model, and synthesis transform $W_c \in \mathbb{R}^{C_{\mathrm{in} \times K \times K},$6 (Wang et al., 21 Sep 2025). Its integration is described as lightweight: keep the encoder $W_c \in \mathbb{R}^{C_{\mathrm{in} \times K \times K},$7 architecture intact, analyze its final projection weights once, reorder and regroup the latent channels according to ISCS, keep the entropy model and decoder architectures unchanged, and retrain under the original training setup using the new latent organization.

Retraining is explicitly required. The paper states: “The entropy model and decoder $W_c \in \mathbb{R}^{C_{\mathrm{in} \times K \times K},$8 are retrained under the original training configuration ... differing only by adopting the more efficient ISCS-based slice scheduler” (Wang et al., 21 Sep 2025). Thus the method is not purely zero-shot insertion. The encoder is analyzed from pretrained weights, but the remainder of the pipeline is adapted through retraining.

The main operational consequence is a new latent slicing and scheduling mechanism. Channels are reorganized into SC–SA groups; each group is split into slices via $W_c \in \mathbb{R}^{C_{\mathrm{in} \times K \times K},$9; SC–SA groups are processed independently in parallel; and slices within each group are processed serially according to the original entropy model’s scheduling policy (Wang et al., 21 Sep 2025). This enables slice-parallel decoding and analogous parallel gains in encoding.

The paper frames the efficiency gain as a reduction of unnecessary serial dependencies in context modeling. With ISCS, highly correlated channels are grouped together, inter-group processing can run in parallel, the context model still sees meaningful reduced context, and serial burden is substantially reduced (Wang et al., 21 Sep 2025). It also presents this as redundancy reduction, because channels are grouped by parameter-inferred dependency rather than treated as a broad undifferentiated context space.

The method is evaluated in two representative pretrained VAE-based LIC models:

• MLIC+, described as CNN-based with anchor-to-non-anchor channel-wise context prediction;
• STF, described as mixed CNN/Transformer with window-based attention and channel/spatial dependency modeling (Wang et al., 21 Sep 2025).

The training set contains over $b_c.$0 images larger than $b_c.$1, drawn from ImageNet, COCO 2017, DIV2K, and Flickr2K. Reported evaluations use JPEG-AI, Kodak, and CLIC Professional Validation (Wang et al., 21 Sep 2025).

6. Empirical results, scope, and relation to other uses of “channel space”

The experimental results emphasize competitive reconstruction quality together with substantial efficiency gains. For rate–distortion behavior, the paper reports that ISCS+MLIC+ closely follows MLIC+, with slightly higher bitrates at the same PSNR, while ISCS+STF incurs a moderate RD gap, though the overall RD curve still tracks STF reasonably closely (Wang et al., 21 Sep 2025). The paper’s strongest empirical claim is therefore not uniformly superior RD, but rather competitive RD with substantial efficiency gains.

For qualitative reconstruction, Figure 4 is described as showing that ISCS+MLIC+ and MLIC+ are “nearly identical” in perceptual quality, both preserving sharp structures and fine textures, with residual maps that are very similar (Wang et al., 21 Sep 2025).

The clearest quantitative gains are in runtime and context-module runtime on a single Tesla H100 GPU. For STF vs. ISCS+STF on Kodak, encoding changes from $b_c.$2 to $b_c.$3 ms and decoding from $b_c.$4 to $b_c.$5 ms; context encoding changes from $b_c.$6 to $b_c.$7 ms and context decoding from $b_c.$8 to $b_c.$9 ms. On JPEG-AI, encoding changes from $\operatorname{Var}(W_c) = \frac{1}{K^2 C_{\mathrm{in} \sum_{i,j,k}\bigl(W_c[i,j,k]-\mu_c\bigr)^2,$0 to $\operatorname{Var}(W_c) = \frac{1}{K^2 C_{\mathrm{in} \sum_{i,j,k}\bigl(W_c[i,j,k]-\mu_c\bigr)^2,$1 ms and decoding from $\operatorname{Var}(W_c) = \frac{1}{K^2 C_{\mathrm{in} \sum_{i,j,k}\bigl(W_c[i,j,k]-\mu_c\bigr)^2,$2 to $\operatorname{Var}(W_c) = \frac{1}{K^2 C_{\mathrm{in} \sum_{i,j,k}\bigl(W_c[i,j,k]-\mu_c\bigr)^2,$3 ms; context encoding changes from $\operatorname{Var}(W_c) = \frac{1}{K^2 C_{\mathrm{in} \sum_{i,j,k}\bigl(W_c[i,j,k]-\mu_c\bigr)^2,$4 to $\operatorname{Var}(W_c) = \frac{1}{K^2 C_{\mathrm{in} \sum_{i,j,k}\bigl(W_c[i,j,k]-\mu_c\bigr)^2,$5 ms and context decoding from $\operatorname{Var}(W_c) = \frac{1}{K^2 C_{\mathrm{in} \sum_{i,j,k}\bigl(W_c[i,j,k]-\mu_c\bigr)^2,$6 to $\operatorname{Var}(W_c) = \frac{1}{K^2 C_{\mathrm{in} \sum_{i,j,k}\bigl(W_c[i,j,k]-\mu_c\bigr)^2,$7 ms. The paper summarizes this as roughly one-third reduction in STF context runtime (Wang et al., 21 Sep 2025).

For MLIC+ vs. ISCS+MLIC+ on Kodak, encoding changes from $\operatorname{Var}(W_c) = \frac{1}{K^2 C_{\mathrm{in} \sum_{i,j,k}\bigl(W_c[i,j,k]-\mu_c\bigr)^2,$8 to $\operatorname{Var}(W_c) = \frac{1}{K^2 C_{\mathrm{in} \sum_{i,j,k}\bigl(W_c[i,j,k]-\mu_c\bigr)^2,$9 ms and decoding from $\mu_c= \frac{1}{K^2 C_{\mathrm{in} \sum_{i,j,k} W_c[i,j,k].$0 to $\mu_c= \frac{1}{K^2 C_{\mathrm{in} \sum_{i,j,k} W_c[i,j,k].$1 ms; context encoding changes from $\mu_c= \frac{1}{K^2 C_{\mathrm{in} \sum_{i,j,k} W_c[i,j,k].$2 to $\mu_c= \frac{1}{K^2 C_{\mathrm{in} \sum_{i,j,k} W_c[i,j,k].$3 ms and context decoding from $\mu_c= \frac{1}{K^2 C_{\mathrm{in} \sum_{i,j,k} W_c[i,j,k].$4 to $\mu_c= \frac{1}{K^2 C_{\mathrm{in} \sum_{i,j,k} W_c[i,j,k].$5 ms. On JPEG-AI, encoding changes from $\mu_c= \frac{1}{K^2 C_{\mathrm{in} \sum_{i,j,k} W_c[i,j,k].$6 to $\mu_c= \frac{1}{K^2 C_{\mathrm{in} \sum_{i,j,k} W_c[i,j,k].$7 ms and decoding from $\mu_c= \frac{1}{K^2 C_{\mathrm{in} \sum_{i,j,k} W_c[i,j,k].$8 to $\mu_c= \frac{1}{K^2 C_{\mathrm{in} \sum_{i,j,k} W_c[i,j,k].$9 ms; context encoding changes from $\operatorname{Var}(W_c)$0 to $\operatorname{Var}(W_c)$1 ms and context decoding from $\operatorname{Var}(W_c)$2 to $\operatorname{Var}(W_c)$3 ms. The paper explicitly notes that for JPEG-AI, MLIC+ context times are reduced by approximately $\operatorname{Var}(W_c)$4 (Wang et al., 21 Sep 2025).

The ablation comparing SC–SA grouping with naive grouping reports that at all tested $\operatorname{Var}(W_c)$5, SC–SA grouping gives lower bpp and higher PSNR than naive grouping. For example, at $\operatorname{Var}(W_c)$6, SC–SA gives $\operatorname{Var}(W_c)$7 bpp and $\operatorname{Var}(W_c)$8 dB, while naive grouping gives $\operatorname{Var}(W_c)$9 bpp and $\bar W_c$0 dB; at $\bar W_c$1, SC–SA gives $\bar W_c$2 bpp and $\bar W_c$3 dB, while naive grouping gives $\bar W_c$4 bpp and $\bar W_c$5 dB (Wang et al., 21 Sep 2025). Table 4 is described as showing strong agreement between channels causing large degradation in ablation tests and channels having large variance scores.

The paper also states several limitations. The method is mainly applied to VAE-based LIC models and has not been tested on pure transformer LIC frameworks or diffusion-based compression models. It is also most useful when the codec uses channel-aware context prediction; methods using solely spatial context prediction may not benefit, because channel order would not affect them. The paper further states that “A deeper theoretical connection between the two will be explored in future work,” referring to the relation between parameter-based saliency and ablation-based saliency (Wang et al., 21 Sep 2025).

Outside learned image compression, the phrase “Invariant Salient Channel Space” is not the native terminology of several superficially related works. In multiple salient object detection, the closest component is a channel-space non-local block inside a Dual-Space Non-Local Block (DSNLB), which builds global dependencies among channels through a learned channel affinity matrix, but that work does not define an invariant salient channel space and does not formally claim invariance to translation, scale, or appearance (Deng et al., 2021). In relaxed maximum-likelihood blind MIMO channel estimation for OSTBCs, the closest counterpart is a deterministic invariant space associated with residual ambiguity, but the paper does not use the term ISCS (Kalbasi et al., 2017). In reciprocal-lattice symmetry analysis of crystals, the closest object is the wave vector star channel and the corresponding wave vector star channel group, again without using ISCS as a term (Kim et al., 2016).

In its precise technical usage, therefore, ISCS refers to a stable, parameter-discovered organization of latent channels in pretrained VAE-based learned image compression models. Its central claim is that important latent channels are not isolated but form structured groups in which an SC channel carries dominant structural information, nearby SA channels carry correlated or complementary details, and some additional important channels are revealed by large bias magnitude (Wang et al., 21 Sep 2025). The practical consequence is a deterministic latent reordering and grouping mechanism that enables slice-parallel decoding, reduces context-model seriality, and yields substantial runtime gains while maintaining nearly the same reconstruction quality.