Dynamic Skip Connection (DSC)

• Dynamic Skip Connection (DSC) comprises a family of adaptive skip pathways that condition feature transmission based on input, learned feature statistics, or external conditions.
• DSC mechanisms enhance network performance in applications such as image restoration, sequence modeling, and segmentation by dynamically filtering, gating, or routing information.
• Implementations of DSC range from per-pixel filtering in UDC restoration to binary gating in SkipNet and recurrent state selection in LSTMs, offering flexible design choices.

Dynamic Skip Connection (DSC) denotes a family of mechanisms in which a skip or shortcut pathway is no longer a passive identity map or fixed encoder–decoder concatenation, but an adaptive operator whose effect depends on the current input, a degradation prior, a routing policy, or learned feature statistics. The term is not standardized across the literature. In one line of work it refers to per-pixel, per-channel dynamic filtering of encoder features inside an image-restoration network; in another it denotes hard per-input routing over residual blocks; elsewhere it means a recurrent shortcut across time in an LSTM, a residual-edited skip in diffusion U-Nets, or a test-time-adaptive skip block for medical segmentation (Feng et al., 2021, Wang et al., 2017, Gui et al., 2018, Cao et al., 18 Sep 2025). A stable encyclopedic reading therefore treats DSC not as a single canonical module, but as a class of skip-path interventions that alter what information is transmitted, when it is transmitted, or how it is transformed.

1. Terminological scope and conceptual boundaries

Across papers, the skip path is engineered along several distinct axes: execution, transformation, fusion, and adaptation. In execution-oriented formulations, the model decides whether a residual branch should run at all. In transformation-oriented formulations, the skip feature is filtered or edited before fusion. In fusion-oriented formulations, multiple skip sources are reweighted or reassembled. In adaptation-oriented formulations, the skip path changes with sample content, external conditions, or test-time updates. This suggests that “dynamic” may refer either to hard routing, content-conditioned filtering, condition-dependent modulation, or normalization-mediated effective scaling rather than to a single mechanism (Xu et al., 2024).

Paper	Role of the skip pathway	Nature of the dynamicity
DISCNet (Feng et al., 2021)	Encoder feature is dynamically filtered before skip transfer	Per-pixel, per-channel, PSF-conditioned local kernels
SkipNet (Wang et al., 2017)	Residual branch may be executed or bypassed	Binary per-input routing
LSTM with DSC (Gui et al., 2018)	Recurrent predecessor is selected from a history window	RL-based temporal skip selection
SCEdit (Jiang et al., 2023)	Skip tensor is residually edited before decoder concatenation	Static tuner; condition-dependent in controllable form
rSkip+LN (Liu et al., 2021)	Skip contribution is recursively normalized and modulated	Effective scaling depends on LN statistics
TransGUNet (Nam et al., 14 Feb 2025)	Cross-scale skip is graph-processed, attention-weighted, and entropy-filtered	Input-dependent graph, node attention, channel selection
DSC block for U-like nets (Cao et al., 18 Sep 2025)	Skip feature is refined by DMSK and TTT before decoder fusion	Dynamic kernel selection and test-time adaptation

Two adjacent categories are important for delimitation. First, some methods broaden or densify skip topology without making path usage strongly input-dependent. FSCN routes all preserved encoder scales to each decoder stage and uses ACM for fusion, but its skip graph is fixed after training (Lai et al., 2022). Second, some methods add learnable long-range edges while keeping topology fixed. N2NSkip introduces sparse neuron-to-neuron skip connections in pruned networks, yet the added edges are randomly selected at construction time and are not input-conditioned (Subramaniam et al., 2022). Both are important to DSC research, but neither is a canonical DSC formulation in the strong routing sense.

2. Physics-conditioned dynamic skip transformation in DISCNet

The most explicit encoder–decoder formulation of DSC in the provided literature is the Dynamic Skip Connection inside DISCNet for under-display camera restoration. The underlying problem is not ordinary blur removal. In a UDC system, the semi-transparent OLED pixel array attenuates and diffracts incident light, producing noise, flare, haze, blur, and structured highlight artifacts. The paper models image formation as

$\hat{y} = \phi[C(x*k+n)],$

with clipping

$C(x)=\min(x, x_{\max}),$

and states that “sensor saturation breaks the shift-invariance of the single-PSF-based convolution, leading to spatially-variant degradation” (Feng et al., 2021). That statement is the architectural rationale for DSC: a static U-Net-style skip may simply forward artifact-contaminated encoder features into the decoder.

The measured UDC PSF has very large support, over $800\times 800$, with a strong main lobe, long-tail spike-shaped sidelobes, and color shift in the tail. DISCNet therefore conditions the network on the PSF and performs dynamic processing in feature space at multiple scales rather than attempting enormous image-space filters. The restoration branch is an encoder–decoder with skip connections. The encoder has three convolutional blocks, each containing a $3\times 3$ convolution with stride $2$, a LeakyReLU, and two residual blocks, producing $E_1,E_2,E_3$. The decoder has two convolutional blocks, each with an up-convolution and two residual blocks. A parallel condition encoder takes the degraded image concatenated with a PCA-compressed PSF descriptor; following SRMD/SFTMD-style conditioning, the PSF is projected to a $b$-dimensional kernel code with $b=5$, tiled to $H\times W\times b$, and concatenated with the degraded image.

At scale $n$, a filter generator $C(x)=\min(x, x_{\max}),$0 processes condition features $C(x)=\min(x, x_{\max}),$1 and predicts

$C(x)=\min(x, x_{\max}),$2

where the default dynamic filter size is $C(x)=\min(x, x_{\max}),$3. Each spatial location and channel receives its own $C(x)=\min(x, x_{\max}),$4 kernel. The dynamic skip transformation is then defined by

$C(x)=\min(x, x_{\max}),$5

where $C(x)=\min(x, x_{\max}),$6 is reshaped from the corresponding slice of $C(x)=\min(x, x_{\max}),$7, and $C(x)=\min(x, x_{\max}),$8 is the same-channel $C(x)=\min(x, x_{\max}),$9 patch centered at $800\times 800$0. The skip path therefore transmits $800\times 800$1, not $800\times 800$2. This is the exact operational meaning of “Dynamic Skip Connection” in the paper: the encoder feature is dynamically filtered, pixel by pixel and channel by channel, before it is cast back to the restoration branch.

Several distinctions are explicit. DISCNet’s DSC is not a scalar gate, not an attention mask, and not cross-channel mixing. The tensor size $800\times 800$3, rather than $800\times 800$4, indicates depthwise-in-spirit per-channel filtering. The exact fusion operator between transformed skip features and decoder features is not specified in the text. Likewise, the network is trained with a weighted combination of $800\times 800$5 loss and VGG loss, but the coefficients are not given in the provided text. These omissions are part of the published record and materially affect reproducibility.

3. Dynamic execution and temporal routing

A second major meaning of DSC concerns dynamic execution rather than dynamic filtering. SkipNet modifies a residual network so that the identity shortcut is always available, but the residual computation may be skipped on a per-input basis. Starting from the ResNet block

$800\times 800$6

SkipNet introduces a binary gate,

$800\times 800$7

If $800\times 800$8, the expensive residual branch is bypassed and only the identity path remains. The gating problem is formulated as sequential decision making, with a hybrid learning algorithm that combines supervised learning and reinforcement learning, and a reward

$800\times 800$9

This yields an adaptive-depth CNN that reduces computation by $3\times 3$0-$3\times 3$1 while preserving the accuracy of the original model on four benchmark datasets (Wang et al., 2017).

An even broader use of the term appears in “Long Short-Term Memory with Dynamic Skip Connections,” where the skip is not a feed-forward shortcut but a learned recurrent link across time. Instead of always depending on $3\times 3$2, the model can connect to a selected earlier state $3\times 3$3 within a bounded window $3\times 3$4. The interpolated predecessor is

$3\times 3$5

with the skip index $3\times 3$6 sampled from a categorical policy over $3\times 3$7. The policy state is

$3\times 3$8

and the agent is trained by REINFORCE with entropy regularization. Here DSC denotes a hard latent dependency decision: at each time step, the recurrence chooses which earlier state should be connected to the current token (Gui et al., 2018).

These two papers define different but structurally related notions of dynamicity. SkipNet makes the skip path the vehicle for conditional execution in depth; the LSTM model makes it the vehicle for conditional predecessor selection in time. In both cases, the defining feature is not improved feature fusion, but input-dependent path choice.

4. Skip editing, scaling, and optimization in generative and residual models

A third line of work modifies skip content without hard routing. SCEdit targets diffusion U-Nets and places a lightweight SC-Tuner directly on each encoder–decoder skip tensor. The edited skip is

$3\times 3$9

with adapter-style bottleneck operation

$2$0

The decoder receives $2$1 rather than the raw encoder feature. In the controllable extension,

$2$2

with $2$3. SCEdit is therefore condition-dependent in the controllable case, but the paper is explicit that it is not dynamic in the strong sense of learned gates, sigmoids, masks, or timestep-conditioned affine modulation. Its trainable modules are localized on skip pathways so that “backward propagation only passing to the decoder blocks” reduces memory and compute (Jiang et al., 2023).

A related but distinct notion appears in “Rethinking Skip Connection with Layer Normalization in Transformers and ResNets.” The paper treats residual addition as modulation,

$2$4

and argues that the standard coefficient $2$5 is a hard-coded choice whose naive alteration causes gradient exploding or vanishing with depth. The proposed recursive formulation,

$2$6

makes the effective shortcut-to-residual ratio adaptive through layer-normalization statistics. For $2$7,

$2$8

and the general effective ratio is

$2$9

The skip path is not routed or filtered by a separate controller, but its effective contribution becomes layer-, depth-, and input-statistics-dependent (Liu et al., 2021).

A complementary optimization-centric interpretation is proposed in “Rethinking skip connection model as a learnable Markov chain.” There the residual stack is written as

$E_1,E_2,E_3$0

and efficiency is measured by the alignment of predicted directions $E_1,E_2,E_3$1 with ideal directions $E_1,E_2,E_3$2,

$E_1,E_2,E_3$3

The practical “penal connection” does not alter forward skip topology; instead it modifies the residual-branch gradient to

$E_1,E_2,E_3$4

This is not a canonical DSC mechanism, but it shows that skip-path adaptivity may be framed as optimization of residual transitions rather than explicit routing or gating (Chen et al., 2022).

5. Cross-scale adaptive skip fusion in dense prediction and medical segmentation

Encoder–decoder segmentation models motivate another class of DSC-like designs because same-scale static skips can be too restrictive. FSCN replaces conventional same-resolution skips with a full skip connection network in which each decoder stage receives resized encoder features from all preserved scales. Its core fusion equation is

$E_1,E_2,E_3$5

followed by

$E_1,E_2,E_3$6

This design is better viewed as a static full-skip architecture with adaptive post-concatenation channel attention than as a full DSC method. The scalar weights $E_1,E_2,E_3$7 are learned but fixed after training, whereas the SENet component is input-dependent (Lai et al., 2022).

TransGUNet moves closer to strong DSC semantics. It constructs an Attentional Cross-Scale Graph Neural Network over fused skip features. Encoder outputs $E_1,E_2,E_3$8 are reduced to $E_1,E_2,E_3$9, resized to $b$0, and concatenated into $b$1. After flattening, each spatial position becomes a graph node, and edges are built by dilated $b$2-nearest neighbors with $b$3. Graph convolution yields $b$4, followed by node attention,

$b$5

with $b$6. Entropy-driven feature selection then computes per-channel entropy, selects the Bottom-$b$7 channels with $b$8, and uses only those channels to generate spatial attention for the full graph-enhanced skip tensor. The result is adaptive at several levels simultaneously: graph connectivity depends on current features, node weights are input-dependent, channel selection is sample-specific, and spatial gating is computed on the fly (Nam et al., 14 Feb 2025).

The most explicit recent medical-segmentation use of the name appears in “Enhancing Feature Fusion of U-like Networks with Dynamic Skip Connections.” Its DSC block combines a Dynamic Multi-Scale Kernel module and a Test-Time Training module. DMSK begins with

$b$9

predicts kernel probabilities

$b=5$0

selects

$b=5$1

and applies cascaded depthwise convolutions,

$b=5$2

After spatial and channel attention, the module returns

$b=5$3

TTT then adapts skip-path parameters at inference time through

$b=5$4

using the self-supervised objective

$b=5$5

with $b=5$6, $b=5$7, and output

$b=5$8

This formulation directly addresses the two limitations named by the paper: inter-feature constraint, meaning static cross-layer transmission, and intra-feature constraint, meaning insufficient adaptive multi-scale modeling within the skip feature itself (Cao et al., 18 Sep 2025).

6. Empirical evidence, misconceptions, and unresolved issues

The empirical record shows that the effectiveness of DSC depends on what exactly is made dynamic. In DISCNet, the ablation table is unusually direct: a baseline restoration branch trained on variational kernels achieves $b=5$9 and $H\times W\times b$0; adding single-scale dynamic filters with image condition only yields $H\times W\times b$1; PSF condition only yields $H\times W\times b$2; both conditions yield $H\times W\times b$3; and the full multi-scale DISCNet reaches $H\times W\times b$4 and $H\times W\times b$5. On synthetic comparisons, DISCNet reports $H\times W\times b$6 for PSNR/SSIM/LPIPS, ahead of SFTMD at $H\times W\times b$7, while using $H\times W\times b$8M parameters and $H\times W\times b$9G FLOPs versus SFTMD’s $n$0M and $n$1G FLOPs (Feng et al., 2021).

In dynamic-routing models, the evidence emphasizes conditional computation and long-range dependency. SkipNet reports computation reductions on the order of roughly $n$2 to $n$3 while maintaining accuracy close to the original backbone. The LSTM model reports that, in number prediction, dynamic skip improves over standard LSTM by nearly $n$4 accuracy; for sequence length $n$5, accuracy rises from $n$6 to $n$7, and for sequence length $n$8, from $n$9 to $C(x)=\min(x, x_{\max}),$00 when $C(x)=\min(x, x_{\max}),$01 (Wang et al., 2017, Gui et al., 2018).

In diffusion and segmentation, the gains arise from targeted skip manipulation rather than hard execution decisions. SCEdit achieves FID $C(x)=\min(x, x_{\max}),$02 with $C(x)=\min(x, x_{\max}),$03M trainable parameters, $C(x)=\min(x, x_{\max}),$04G memory, and $C(x)=\min(x, x_{\max}),$05 training time on COCO2017 text-to-image generation, outperforming full fine-tuning at FID $C(x)=\min(x, x_{\max}),$06 and LoRA/r=64 at FID $C(x)=\min(x, x_{\max}),$07; the paper highlights a $C(x)=\min(x, x_{\max}),$08 memory reduction relative to LoRA. In controllable generation, SCEdit/k=3 reports FID $C(x)=\min(x, x_{\max}),$09 against ControlNet’s $C(x)=\min(x, x_{\max}),$10 (Jiang et al., 2023). TransGUNet’s full cross-scale GNN plus node attention yields Seen DSC $C(x)=\min(x, x_{\max}),$11 and Unseen DSC $C(x)=\min(x, x_{\max}),$12, compared with $C(x)=\min(x, x_{\max}),$13 and $C(x)=\min(x, x_{\max}),$14 for the no-GNN skip baseline, while the model reports $C(x)=\min(x, x_{\max}),$15M parameters, $C(x)=\min(x, x_{\max}),$16G FLOPs, $C(x)=\min(x, x_{\max}),$17 ms, and $C(x)=\min(x, x_{\max}),$18G GPU memory (Nam et al., 14 Feb 2025). The U-like DSC block improves multiple backbones; for example, on abdomen CT, nnU-Net moves from Dice $C(x)=\min(x, x_{\max}),$19 to $C(x)=\min(x, x_{\max}),$20, and on cell segmentation, U-Mamba_Bot moves from F1 $C(x)=\min(x, x_{\max}),$21 to $C(x)=\min(x, x_{\max}),$22 (Cao et al., 18 Sep 2025).

Several misconceptions recur. First, DSC is not synonymous with attention. DISCNet predicts full local convolution kernels; SkipNet uses binary routing; the LSTM paper samples a discrete predecessor; SCEdit primarily performs residual skip tuning; and rSkip+LN relies on normalization-mediated effective scaling (Feng et al., 2021, Wang et al., 2017, Gui et al., 2018, Jiang et al., 2023, Liu et al., 2021). Second, not every sophisticated skip design is dynamic. FSCN broadens the skip candidate set but keeps the graph fixed, and N2NSkip improves sparse-network connectivity through fixed randomly sampled long-range edges rather than conditional routing (Lai et al., 2022, Subramaniam et al., 2022). Third, the broader skip-connection survey does not organize the field around DSC; its closest explicit predecessors are Highway Networks, attention-modulated residual branches, and stochastic depth, which indicates that the terminology remains unsettled (Xu et al., 2024).

The unresolved issues are correspondingly heterogeneous. Discrete policies introduce optimization difficulty and policy-gradient variance, as seen in SkipNet and the LSTM model. Test-time adaptation improves skip specificity but increases latency, as acknowledged for the TTT-based DSC block. Several papers also leave central implementation details unspecified: DISCNet does not specify the exact fusion operator or loss weights; FSCN does not specify the exact resampling primitive for every skip path; the medical DSC block does not fully specify TTT update schedules or exact kernel sets; and SCEdit, despite condition-dependent skip fusion, does not define a learned skip-gating law (Feng et al., 2021, Lai et al., 2022, Jiang et al., 2023, Cao et al., 18 Sep 2025). These gaps do not negate the central result, but they show that DSC remains better understood as a research direction—active, varied, and only partially standardized—than as a single settled architectural primitive.