Structure Transfer

• Structure Transfer is a method that preserves invariant internal relationships within a source during transformations across domains.
• It employs auxiliary carriers such as semantic masks, anchors, and learned embeddings to mediate constrained mappings and prevent undesired deformations.
• The approach is applied in diverse fields including vision, representation learning, and physics, showcasing both asymmetric and hierarchical transfer dynamics.

Across the cited literature, structure transfer denotes a family of procedures in which relations internal to a source—such as object-part organization, manifold neighborhoods, class geometry, task hierarchy, causal dependencies, network modes, or semantic equivalence—are carried into a target representation or domain under explicit preservation constraints. In computer vision, the target is often an image, video, or 3D scene whose appearance changes while object structure is preserved; in representation learning, the target is a feature space or classifier whose geometry is regularized by source-side structure; in formal settings, the target is a new representation or algebraic object required to satisfy a specified relation to the source (Tao et al., 2022, Zhang et al., 2018, Raggi et al., 3 Sep 2025). This breadth suggests that “structure” is not a single object but a role: it is the component of a source that remains invariant, or is deliberately constrained, while other degrees of freedom are transformed.

1. Conceptual scope and recurring abstractions

A recurring pattern is the separation of what is transferred from what is allowed to vary. In structure-aware motion transfer, motion is learned from a driving video while appearance is preserved from a source image (Tao et al., 2022). In reference-based fashion design, appearance is transferred from a reference image while the clothing structure of the input is preserved (Cao et al., 2023). In representation learning, manifold structure is transferred from data space to feature space so that the learned representation satisfies intrinsic geometric relations from the data (Zhang et al., 2018). In a representational-system agnostic formulation, the target representation is generated so that it satisfies any specified relation to the source, such as semantic equivalence, by exploiting schemas over construction spaces (Raggi et al., 3 Sep 2025).

A second recurring pattern is the use of an auxiliary structural carrier. This carrier may be explicit, as with semantic masks, classifier constellations, task-transfer graphs, intermediate finite-element models, or coalgebra-indexed coverings; or latent, as with root anchors, learned structure embeddings, or community-induced normal modes (Cao et al., 2023, Liu et al., 2021, Aristimunha et al., 2023, Dardeno et al., 23 Mar 2026, Lauve et al., 2018, Tao et al., 2022, Zhang et al., 2021, Hahto et al., 2024). A plausible implication is that many structure-transfer systems are best understood not as direct source-to-target mappings, but as constrained constructions mediated by a third object that encodes invariants.

Domain	Structural object	Transfer mechanism
Vision and graphics	Anchors, masks, instances, parts	Regularization, mask guidance, cross-view alignment, correspondence
Representation learning and tasks	Manifolds, class geometry, task relations, structure embeddings	Manifold loss, shared displacements, transfer maps, separable self-attention
Engineering, physics, and formal systems	Intermediate structures, normal modes, coverings, schemas	Interpolation chains, resonant routing, bicategorical transfer, schema-guided construction

This classification is synthetic. The cited works themselves differ in their objectives, but they consistently operationalize structure transfer through correspondence constraints, factorized representations, or transfer rules that suppress unconstrained deformation.

2. Structure-aware transfer in vision and graphics

In vision, structure transfer is typically posed as preservation of geometry or part organization under appearance, motion, or style change. The deformable anchor model (DAM) for motion transfer introduces motion anchors and a latent root anchor, and regularizes anchor flows so that structure is “well captured and preserved” without prior structure information (Tao et al., 2022). Its basic regularizer is

$$\mathcal{L}_{k\leftarrow r} = \left\|\mathcal{T}_k (z^d_k)-\mathcal{T}_r\left (z^d_{k}\right)\right\|_2,$$

while the hierarchical variant (HDAM) adds intermediate latent anchors for more complicated structures. DAM and HDAM are trained in an unsupervised manner with perceptual loss, equivariance loss, and structure regularization, and the reported evaluations on TaiChiHD, FashionVideo, MGIF, and VoxCeleb1 show lower $\mathcal{L}_1$, AKD, MKR, and AED than Monkey-Net, FOMM, and RegionMM on the presented benchmarks (Tao et al., 2022).

Diffusion-based structure transfer appears in several forms. DiffFashion decouples the foreground clothing with automatically generated semantic masks by conditioned labels, uses the mask in the denoising process, and introduces DINO-ViT-based appearance and structural guidance (Cao et al., 2023). Its mask-guided mixing rule is

$$\tilde{x}^{A}_t = M \cdot x^{A}_t + (1 - M)\cdot[\omega_{mix} \cdot x^{S}_{qt} + (1 - \omega_{mix}) \cdot x^{A}_t].$$

The method is unsupervised, and the reported user study gives DiffFashion an overall user rating of $75.04$, structure similarity of $91.07 \pm 3.82$, appearance similarity of $52.89 \pm 7.92$, and realism of $81.15 \pm 4.76$ (Cao et al., 2023). In text-driven style transfer, Adaptive Style Incorporation (ASI) argues that direct concatenation of content and style prompts causes “unavoidable structure distortions,” and instead decouples content and style with Siamese Cross-Attention and recombines them with Adaptive Content-Style Blending (Ge et al., 2024). The method is tuning-free and uses attention-head and spatial masks to prevent style overwrite in structure-critical regions (Ge et al., 2024).

Three-dimensional structure transfer introduces an additional consistency constraint across viewpoints. SSGaussian stylizes key views with cross-view attention inserted into the last upsampling block of the UNet, then transfers the stylized key views onto a 3D Gaussian Splatting representation via instance-level style transfer (Xu et al., 4 Sep 2025). The cross-view attention is written as

$$\mathrm{Attn}(z, z^{1:K}) = \mathrm{Softmax}\left(\frac{Q(z)K(z^{1:K})^\top}{\sqrt{d}}\right)V(z^{1:K}),$$

and the method uses identity encodings from Gaussian Grouping to enforce instance-level consistency across stylized views (Xu et al., 4 Sep 2025). Reported consistency metrics show Short LPIPS $0.031$, Short RMSE $0.028$, Long LPIPS $\mathcal{L}_1$0, and Long RMSE $\mathcal{L}_1$1, all lower than ARF, StyleGaussian, and G-Style in the table presented in the paper (Xu et al., 4 Sep 2025).

Related work on artistic and material transfer emphasizes correspondence construction rather than direct style injection. “Coarse-to-Fine Structure-Aware Artistic Style Transfer” reconstructs coarse stylized features at low resolution with a Coarse Network, then synthesizes high-resolution outputs with a Fine Network containing three structural selective fusion modules (Liu et al., 8 Feb 2025). “Photo-to-Shape Material Transfer for Diverse Structures” uses an image translation network to translate color from the exemplar to a projection of the 3D shape and segmentation from the projection to the exemplar, then a material prediction network assigns realistic materials using translated images, perceptual similarity, and a consistency loss across the two translation paths (Hu et al., 2022). Both works treat structural preservation as an explicit design requirement rather than an emergent by-product of appearance transfer.

3. Structure transfer in learned representations and sequence generation

A major line of work formulates structure transfer as a regularization principle for feature learning. Structure Transfer Machine (STM) transfers manifold structure from the data space to the feature space by incorporating a manifold loss into a standard deep-learning objective (Zhang et al., 2018). For LLE-based regularization, the paper gives

$\mathcal{L}_1$2

and for Laplacian-based regularization,

$\mathcal{L}_1$3

The resulting architecture adds a manifold-loss module near the high-level feature layer, and the paper reports improvements on MNIST, CIFAR-10/100, ImageNet, and OTB-50, including a MNIST error rate of $\mathcal{L}_1$4 versus $\mathcal{L}_1$5 for LeNet++ in the cited comparison (Zhang et al., 2018).

In long-tailed recognition, GistNet transfers class geometry rather than local neighborhoods (Liu et al., 2021). Each class is represented by a center $\mathcal{L}_1$6 plus shared displacements $\mathcal{L}_1$7, producing a class-specific constellation

$\mathcal{L}_1$8

The classifier then uses the maximal response over the constellation, and training separates updates for class centers and shared structure parameters through a hybrid objective

$\mathcal{L}_1$9

The key claim is that overfitting to head classes should be exploited rather than eliminated, so that reliable geometry can be transferred to few-shot classes (Liu et al., 2021). On ImageNet-LT, Places-LT, and iNaturalist 2018, the reported tables show GistNet outperforming the plain model, OLTR, and Decoupling in overall accuracy and in many/medium/few-shot partitions (Liu et al., 2021).

Sequence generation supplies another interpretation: structure is encoded separately from local content, then transferred at inference time. Melody Structure Transfer Network (MSTN) assigns each training piece a learned structure embedding $$\tilde{x}^{A}_t = M \cdot x^{A}_t + (1 - M)\cdot[\omega_{mix} \cdot x^{S}_{qt} + (1 - \omega_{mix}) \cdot x^{A}_t].$$0, injects it position-wise as $$\tilde{x}^{A}_t = M \cdot x^{A}_t + (1 - M)\cdot[\omega_{mix} \cdot x^{S}_{qt} + (1 - \omega_{mix}) \cdot x^{A}_t].$$1, and uses separable self-attention so that structure-related and note-level dependencies are disentangled (Zhang et al., 2021). The model can generate music “up to 100 bars” and is evaluated by repeat statistics, pitch and duration distributions, Rhythm Structure Similarity, Interval Structure Similarity, and duplicate rates (Zhang et al., 2021). This use of “structure transfer” is narrower than the representational or algebraic uses of the term, but it shares the same formal move: factorizing a latent representation into a transferable structural component and a variable content component.

Transferability itself can be used to infer task structure. In EEG decoding, representations are pre-trained on one cognitive task and linearly probed on another, and the transferability score is rescaled as

$$\tilde{x}^{A}_t = M \cdot x^{A}_t + (1 - M)\cdot[\omega_{mix} \cdot x^{S}_{qt} + (1 - \omega_{mix}) \cdot x^{A}_t].$$2

The resulting transfer maps are “highly asymmetric,” reveal hubs and bottlenecks, and show gains of “up to 28%” relative to pure supervised training in the reported experiments (Aristimunha et al., 2023). Here, structure transfer is not merely a method for improving a target model; it becomes a measurement device for uncovering hierarchical relations between tasks.

4. Tasks, reinforcement learning, and heterogeneous engineering systems

In reinforcement learning and interactive environments, structure transfer is often framed as transport of causal or transition structure across domains. “Structure Mapping for Transferability of Causal Models” models an environment as an object-oriented MDP, learns action-conditioned DAGs over object attributes using NOTEARS-style continuous optimization, and transfers knowledge by categorizing target objects according to causal role rather than perceptual feature (Pruthi et al., 2020). The graph-learning objective is

$$\tilde{x}^{A}_t = M \cdot x^{A}_t + (1 - M)\cdot[\omega_{mix} \cdot x^{S}_{qt} + (1 - \omega_{mix}) \cdot x^{A}_t].$$3

In the gridworld “Triggers” environment, keys and locks can swap colors between source and target while preserving causal dynamics, and a hybrid approach combining the causal model with a DQN is reported to converge faster and with less variance than a pure DQN (Pruthi et al., 2020).

A different approach is to compute structural similarity directly between two finite MDPs. SS2 constructs state and action similarity measures across a pair of MDPs by recursion on reward differences, Earth Mover’s distance over transition distributions, and Hausdorff distance over action neighborhoods (Ashcraft et al., 2022). The state update is

$$\tilde{x}^{A}_t = M \cdot x^{A}_t + (1 - M)\cdot[\omega_{mix} \cdot x^{S}_{qt} + (1 - \omega_{mix}) \cdot x^{A}_t].$$4

The paper argues that $$\tilde{x}^{A}_t = M \cdot x^{A}_t + (1 - M)\cdot[\omega_{mix} \cdot x^{S}_{qt} + (1 - \omega_{mix}) \cdot x^{A}_t].$$5 defines a distance metric across the joint state space, and then uses the resulting similarities to initialize Q-values in the target MDP through variants of T-STATE, T-AVG, T-STATE-ACT, and T-AVG-ACT (Ashcraft et al., 2022). In the reported GridWorld experiments, SS2-based transfer outperforms Song et al.’s bisimulation metric and a uniform baseline, with SS2 + T-STATE-ACT reaching $$\tilde{x}^{A}_t = M \cdot x^{A}_t + (1 - M)\cdot[\omega_{mix} \cdot x^{S}_{qt} + (1 - \omega_{mix}) \cdot x^{A}_t].$$6 relative performance in the “Small, reward 100” condition (Ashcraft et al., 2022).

For highly disparate physical systems, the direct source-to-target assumption is explicitly rejected. “Transfer learning via interpolating structures” proposes that heterogeneous transfer can be achieved through intermediate structures that continuously morph one structure into another by varying material properties and geometry (Dardeno et al., 23 Mar 2026). Structures are parameterized by $$\tilde{x}^{A}_t = M \cdot x^{A}_t + (1 - M)\cdot[\omega_{mix} \cdot x^{S}_{qt} + (1 - \omega_{mix}) \cdot x^{A}_t].$$7 and interpolated as

$$\tilde{x}^{A}_t = M \cdot x^{A}_t + (1 - M)\cdot[\omega_{mix} \cdot x^{S}_{qt} + (1 - \omega_{mix}) \cdot x^{A}_t].$$8

At each hop, the method combines statistical alignment with the geodesic flow kernel,

$$\tilde{x}^{A}_t = M \cdot x^{A}_t + (1 - M)\cdot[\omega_{mix} \cdot x^{S}_{qt} + (1 - \omega_{mix}) \cdot x^{A}_t].$$9

The bridge-to-bridge case study reports that direct transfer fails, whereas a few intermediates can produce “up to 99%+ accuracy” with GFK, and the bridge-to-aeroplane case study shows that proper chain construction can enable positive transfer even between highly-disparate systems (Dardeno et al., 23 Mar 2026). This is an especially explicit formulation of structure transfer as path construction in configuration space.

5. Physical and networked dynamics

In physical systems, structure transfer may refer not to machine-learning adaptation but to transport of states through a structured medium. For Gaussian states on quantum complex networks, routing performance depends strongly on node degree, link density, and especially community structure (Hahto et al., 2024). The network Hamiltonian is

$75.04$0

and transfer is mediated by resonant coupling to normal modes. The paper compares a one-step protocol with a two-step protocol, finding that the former is approximate but more resilient in complex networks, whereas the latter can be perfect in principle but is generally less robust because stronger coupling excites unwanted modes (Hahto et al., 2024). Community structure is reported to control the appearance of higher-frequency normal modes useful for transfer, and modular networks can support multiple effective transfer channels rather than only the global center-of-mass mode (Hahto et al., 2024).

In aperiodic polymers, structural complexity governs both spectral organization and carrier transport (Mantela et al., 2019). Using a tight-binding wire model,

$75.04$1

the paper studies periodic, quasi-periodic, and fractal sequences and evaluates eigenspectra, HOMO-LUMO gaps, density of states, mean over time probabilities, frequency content, and the pure mean transfer rate

$75.04$2

The reported conclusion is that there is a correspondence between degree of structural complexity and transfer properties, that I polymers are more favorable for charge transfer than D polymers, and that randomized sequences are essentially non-conducting except in specific cases (Mantela et al., 2019).

Network structure can also determine measured transfer, rather than only realized transfer. In a linearly-coupled Gaussian model, pairwise transfer entropy between nodes depends not just on the local link weight but on source in-degree, target in-degree, and weighted motif counts (Novelli et al., 2019). The paper derives a motif expansion up to quartic order and reports that transfer entropy increases with the in-degree of the source and decreases with the in-degree of the target, while clustered motifs involving common parents or multiple walks increase transfer entropy in positively weighted networks (Novelli et al., 2019). Because transfer entropy is equivalent to Granger causality for Gaussian variables, these results also describe how network structure biases functional-connectivity inference (Novelli et al., 2019). In this literature, “transfer” is statistical rather than representational, but the dependence on embedded structure is direct.

6. Algebraic, categorical, and representational formalisms

In algebra and representation theory, structure transfer is formulated through explicit morphisms and normalization laws. “Bialgebra Coverings and Transfer of Structure” defines a partial covering $75.04$3 as a measuring that is also a coalgebra map, and a covering as a surjective partial covering (Lauve et al., 2018). The measuring conditions are

$75.04$4

The central transfer theorem states that if $75.04$5 is primitive, then $75.04$6 is primitive in the image algebra; and if both source and image are Hopf algebras, then the antipode is transferred by

$75.04$7

The same framework also supplies a universal covering coalgebra $75.04$8 and generalizes Nichols’ result on when a bialgebra quotient inherits a Hopf algebra structure (Lauve et al., 2018).

In endoscopy, the term refers to normalization of transfer factors rather than transport of latent geometry. Shelstad analyzes two versions of endoscopic transfer factors, classical and renormalized, associated with the different normalizations of the local Langlands correspondence (Shelstad, 2014). The standard factor is written as

$75.04$9

while the renormalized version is its inverse, up to normalization:

$91.07 \pm 3.82$0

The paper emphasizes that this relation allows immediate passage between the two transfer formalisms on both geometric and spectral sides, and that the Whittaker normalization gives a canonical way to fix absolute factors compatibly (Shelstad, 2014).

A more general and explicitly cross-domain formalization appears in “Structure Transfer: an Inference-Based Calculus for the Transformation of Representations” (Raggi et al., 3 Sep 2025). There, a construction space is a triple

$91.07 \pm 3.82$1

and a multi-space system is

$91.07 \pm 3.82$2

Schemas have the form

$91.07 \pm 3.82$3

and encode preservation of information across relations between representational systems. Transfer schemas and $91.07 \pm 3.82$4-reification then permit the construction of a target representation in a target RS while keeping the source fixed and guaranteeing the desired relation, including semantic equivalence or depictive correctness (Raggi et al., 3 Sep 2025). This is the most explicit attempt in the cited corpus to make structure transfer representational-system agnostic.

7. Recurring principles, asymmetries, and limitations

A common misconception is that structure transfer is equivalent to copying a source into a target with superficial modifications. The surveyed literature instead emphasizes constrained reconstruction. DAM learns anchors rather than using predefined skeletons, DiffFashion relies on semantic-mask guidance rather than direct reference insertion, ASI rejects prompt-level concatenation because it causes structure distortions, and SSGaussian shows that direct fine-tuning of 3DGS on stylized views leads to blurring and artifacts (Tao et al., 2022, Cao et al., 2023, Ge et al., 2024, Xu et al., 4 Sep 2025). In engineering transfer, direct bridge-to-bridge or bridge-to-aeroplane transfer is ineffective unless intermediate structures are introduced (Dardeno et al., 23 Mar 2026). These examples argue against the view that structure can be preserved automatically by a powerful generator or classifier.

Another misconception is symmetry. Several works report directional or asymmetric transfer. EEG transfer matrices are “highly asymmetric,” and some tasks are good sources while others are good targets (Aristimunha et al., 2023). In Gaussian-state routing, source and receiver degree influence fidelity in different ways, and community structure rather than link density controls the availability of useful transfer channels (Hahto et al., 2024). Transfer entropy likewise increases with source in-degree and decreases with target in-degree (Novelli et al., 2019). A plausible implication is that structure transfer should generally be analyzed as an oriented relation, not an undirected similarity.

The design principles that recur across fields are correspondingly precise. First, a transferable structural surrogate is usually introduced: anchors, masks, constellations, task maps, normal modes, or schemas. Second, preservation is enforced by an explicit objective, metric, or theorem: manifold loss, equivariance loss, cosine-distance feature matching, GFK alignment, primitive-preservation, or inversion laws for transfer factors (Zhang et al., 2018, Tao et al., 2022, Xu et al., 4 Sep 2025, Dardeno et al., 23 Mar 2026, Lauve et al., 2018, Shelstad, 2014). Third, successful transfer often depends on choosing the correct granularity of structure. HDAM adds intermediate anchors for articulated bodies, SSGaussian transfers style at the instance level, GistNet shares displacements rather than whole classifier weights, and interpolating-structure methods succeed when hops are small enough in feature space (Tao et al., 2022, Xu et al., 4 Sep 2025, Liu et al., 2021, Dardeno et al., 23 Mar 2026). This suggests that the central difficulty of structure transfer is neither mere correspondence estimation nor mere invariance enforcement, but the construction of the right intermediate object on which both can operate.