Motion2Motion: Cross-Topology Motion Transfer

• Motion2Motion is a training-free framework that transfers animation between characters with different skeletal topologies using sparse bone correspondences and a retrieval-based pipeline.
• It employs temporal patch matching and iterative blending to infer missing joint motions, ensuring improved continuity and realistic motion generation.
• Designed for one-shot and few-shot scenarios, the method enables cross-species retargeting without large paired datasets or dense correspondence mappings.

Searching arXiv for the named topic and closely related motion-transfer work to ground the article in recent literature. arXiv search query: "Motion2Motion cross-topology motion transfer sparse correspondence (Chen et al., 18 Aug 2025)" Motion2Motion is a training-free, example-based framework for cross-topology motion transfer that addresses the problem of transferring animation between characters whose skeletal topologies differ substantially. In its formulation, a source motion defined on one skeleton is transferred to a target skeleton using only one or a few target example motions together with a sparse set of bone correspondences, rather than requiring dense one-to-one bone mappings or large paired motion datasets. The method targets settings in which source and target skeletons may differ in joint count, kinematic structure, and part semantics, including cross-species transfer and retargeting to complex production characters (Chen et al., 18 Aug 2025). Within a broader research landscape, the term “Motion2Motion” is also used more loosely to describe motion transfer or motion-conditioned generation in video and animation systems, but the specific framework described here concerns sparse-condition, patch-based skeletal retargeting across heterogeneous topologies (Chen et al., 18 Aug 2025).

1. Problem setting and relation to prior motion transfer

Motion2Motion studies a retargeting regime that is more difficult than conventional motion transfer. Standard retargeting methods typically assume similar source and target skeletons or rely on learned mappings from paired datasets. Motion2Motion instead considers cases in which topological inconsistency prevents straightforward dense joint correspondence, and where large-scale paired motion corpora spanning diverse topologies are unavailable (Chen et al., 18 Aug 2025). The paper characterizes this as the central obstacle to data-driven cross-topology transfer.

The framework is motivated by a practical observation: in production pipelines, a sparse semantic binding between a small subset of source and target bones is often feasible, and one or a few target motions may already exist. Motion2Motion treats these sparse constraints as sufficient anchors, while the remaining motion degrees of freedom are inferred by retrieving and blending target-side motion patches (Chen et al., 18 Aug 2025). This places the method in contrast to approaches that require model training, motion priors learned from large corpora, or dense rig correspondences.

A useful contextual comparison is the broader motion-transfer literature. In cross-domain image-driven motion transfer, “Motion and Appearance Adaptation for Cross-Domain Motion Transfer” addresses shape and texture mismatch through shape-invariant motion adaptation and structure-guided appearance consistency, but still operates in a learned image synthesis setting rather than a training-free skeletal retargeting regime (Xu et al., 2022). In video generation, “Follow-Your-Motion” formulates motion transfer as finetuning a video diffusion transformer to preserve motion while changing appearance, again assuming a generative video model rather than direct skeleton-space retargeting (Ma et al., 5 Jun 2025). Motion2Motion is therefore distinguished by being model-free, retrieval-based, and explicitly designed for heterogeneous skeletal topology (Chen et al., 18 Aug 2025).

2. Inputs, motion representation, and sparse correspondence model

The method takes three principal inputs: a source motion sequence, one or a few target motions, and a sparse bone correspondence set between source and target skeletons. Let the source motion be $\mathbf{S} \in \mathbb{R}^{F_s \times D_s}$ and the target motion set be $$\mathcal{T} = \{ \mathbf{T}^{(i)} \in \mathbb{R}^{F_i \times D_t} \mid i = 1,\dots,N \}$$ (Chen et al., 18 Aug 2025). Each frame is represented by root velocity plus 6D joint rotations, with feature dimension

$D \leftarrow 3 + 6J,$

where $J$ is the number of joints (Chen et al., 18 Aug 2025).

The sparse correspondence set is written as

$$\mathcal{M} = \{(t,s) \mid t \in \mathcal{J}_t,\; s \in \mathcal{J}_s\},$$

with $K = |\mathcal{M}|$ the number of matched source-target joint pairs (Chen et al., 18 Aug 2025). This correspondence is converted into a block-sparse feature transfer matrix $\mathbf{C}\in\mathbb{R}^{D_t\times D_s}$ using channel-index functions $\mathcal{I}_1(\cdot), \mathcal{I}_2(\cdot)$:

$$\mathbf{C}\Big[\mathcal{I}_1(t):\mathcal{I}_2(t),\; \mathcal{I}_1(s):\mathcal{I}_2(s)\Big] = \begin{cases} \mathbf{I}, & \text{if } (t, s) \in \mathcal{M},\ \mathbf{O}, & \text{otherwise}. \end{cases}$$

This construction copies only the matched joint channels and leaves unmatched target channels unspecified (Chen et al., 18 Aug 2025). A corresponding mask vector marks which target dimensions are constrained:

$\mathbf{m}[i] = \sum_{j=1}^{D_s} \mathbf{C}[i,j].$

The paper presents correspondences as either manual or automatic. The automatic mode uses fuzzy subgraph matching based on traced skeleton chains normalized by direction vectors and compared via cosine similarity (Chen et al., 18 Aug 2025). This suggests a hybrid workflow in which manual bindings remain preferable for difficult cross-species cases, while automatic bindings may be adequate when skeletons are closer.

3. Retrieval-based transfer pipeline

Motion2Motion is explicitly training-free and model-free. There is no neural pretraining, no learned retargeting network, and no test-time parameter optimization. Instead, the algorithm uses patch-based retrieval and iterative blending over target example motions (Chen et al., 18 Aug 2025). The paper describes the pipeline in the following stages.

First, rest poses between source and target skeleton files are pre-aligned. Second, source and target motions are patchified temporally with a sliding window. Third, the source motion is projected into target feature space by applying the sparse correspondence matrix and initializing unmatched channels with Gaussian noise:

$$\mathcal{T} = \{ \mathbf{T}^{(i)} \in \mathbb{R}^{F_i \times D_t} \mid i = 1,\dots,N \}$$0

The resulting projected motion, denoted $$\mathcal{T} = \{ \mathbf{T}^{(i)} \in \mathbb{R}^{F_i \times D_t} \mid i = 1,\dots,N \}$$1, mixes deterministic source-driven channels with stochastic placeholders on target-specific dimensions (Chen et al., 18 Aug 2025). This is the key mechanism by which sparse semantic anchors are combined with target-side completion.

Fourth, each projected motion patch is matched against a database of target patches using a masked mean squared error objective:

$$\mathcal{T} = \{ \mathbf{T}^{(i)} \in \mathbb{R}^{F_i \times D_t} \mid i = 1,\dots,N \}$$2

where $$\mathcal{T} = \{ \mathbf{T}^{(i)} \in \mathbb{R}^{F_i \times D_t} \mid i = 1,\dots,N \}$$3 is MSE and $$\mathcal{T} = \{ \mathbf{T}^{(i)} \in \mathbb{R}^{F_i \times D_t} \mid i = 1,\dots,N \}$$4 weights the importance of matched versus unmatched dimensions (Chen et al., 18 Aug 2025). Fifth, the retrieved patches are averaged or blended to reconstruct an updated motion estimate. Sixth, the retrieval-and-blending cycle is repeated for $$\mathcal{T} = \{ \mathbf{T}^{(i)} \in \mathbb{R}^{F_i \times D_t} \mid i = 1,\dots,N \}$$5 iterations to improve temporal coherence (Chen et al., 18 Aug 2025).

The default settings reported are patch size $$\mathcal{T} = \{ \mathbf{T}^{(i)} \in \mathbb{R}^{F_i \times D_t} \mid i = 1,\dots,N \}$$6 frames, blending weight $$\mathcal{T} = \{ \mathbf{T}^{(i)} \in \mathbb{R}^{F_i \times D_t} \mid i = 1,\dots,N \}$$7, and iterative refinement steps $$\mathcal{T} = \{ \mathbf{T}^{(i)} \in \mathbb{R}^{F_i \times D_t} \mid i = 1,\dots,N \}$$8 (Chen et al., 18 Aug 2025). Because the method is retrieval-based rather than optimization-based, it runs efficiently on CPU; the paper reports inference on a MacBook M1 without GPU (Chen et al., 18 Aug 2025).

4. Temporal patching, iterative refinement, and transfer dynamics

The temporal patch representation is central to the method. Motion2Motion patchifies along the temporal axis because motion dynamics are locally structured in time (Chen et al., 18 Aug 2025). This stands in contrast to methods that learn global sequence mappings. A patch-based view allows the algorithm to match short temporal fragments that are semantically compatible under the sparse correspondence constraints and statistically plausible under the target skeleton.

The iterative match-and-blend procedure is presented as more than a heuristic post-process. After initializing $$\mathcal{T} = \{ \mathbf{T}^{(i)} \in \mathbb{R}^{F_i \times D_t} \mid i = 1,\dots,N \}$$9, the current estimate is repeatedly patchified, matched, and reassembled (Chen et al., 18 Aug 2025). Each iteration suppresses noise in unmatched channels by replacing them with realizable target-side local motion patterns while preserving the mapped source-driven cues on matched channels. The authors also describe the process as a kind of frequency interpolator, since blending patches drawn from target examples can better adapt timing and periodicity than fixed learned transfer functions with limited coverage (Chen et al., 18 Aug 2025).

The parameter $D \leftarrow 3 + 6J,$0 governs a tradeoff between semantic adherence and freedom in unmatched channels. Larger $D \leftarrow 3 + 6J,$1 emphasizes strict faithfulness to sparse correspondences and generally reduces diversity; smaller $D \leftarrow 3 + 6J,$2 gives more influence to target-side completion but can destabilize motion (Chen et al., 18 Aug 2025). The ablation indicates that too small a value makes motion erratic and degrades contact consistency, whereas too large a value over-constrains the result and reduces diversity (Chen et al., 18 Aug 2025).

Patch size likewise shapes the balance between semantics and variability. Smaller patches can miss meaningful temporal dynamics, while larger patches reduce diversity and may over-constrain retrieval (Chen et al., 18 Aug 2025). The chosen default of 11 frames is reported as an effective compromise. Similarly, increasing the number of iterations improves temporal consistency but yields diminishing returns beyond the default $D \leftarrow 3 + 6J,$3 (Chen et al., 18 Aug 2025).

5. One-shot and few-shot operation across similar and cross-species skeletons

Motion2Motion is designed for one-shot and few-shot operation. The target example motions are not paired with the source motion; they serve only as a motion prior and a patch database for the target skeleton (Chen et al., 18 Aug 2025). The source motion contributes semantic structure through the sparse bindings, while the target examples contribute realizable topology-specific dynamics. The paper reports a “test-time scaling” effect in which increasing the number of target examples improves FID, frequency alignment, contact consistency, and diversity (Chen et al., 18 Aug 2025). This indicates that the framework benefits from additional target motion coverage without requiring retraining.

The method is evaluated in both similar-skeleton and cross-species regimes. In similar-skeleton transfer, direct copying of bound features can already be effective, and the gap between automatic and manual binding is reported to be small (Chen et al., 18 Aug 2025). In cross-species transfer, sparse correspondences become more critical because the unmatched target parts must be inferred indirectly from anchors and exemplar patches. The paper gives examples including bear to dog, flamingo to monkey, anaconda to king cobra or raptor, and SMPL human to a complex character with extra joints (Chen et al., 18 Aug 2025).

This division clarifies the intended scope of the framework. It is not limited to small rig variations; it is specifically aimed at topological mismatch, including cases with different limb counts or auxiliary articulated structures. The reported transfer of SMPL motion to a character with 331 joints illustrates this production-oriented use case (Chen et al., 18 Aug 2025). A plausible implication is that the method is especially valuable when fully re-rigging or authoring new target-specific motion data would be costly.

6. Evaluation, results, applications, and relation to adjacent “Motion2Motion” formulations

The evaluation benchmark consists of 14 character animations totaling 1,167 frames, spanning running, walking, jumping, and attacking motions, using Truebones-Zoo for animals and LAFAN for human motion (Chen et al., 18 Aug 2025). The paper divides cases into similar skeletons and cross-species skeletons, and evaluates with FID, frequency alignment, contact consistency, diversity, and FPS (Chen et al., 18 Aug 2025). The binding rate is also reported as

$D \leftarrow 3 + 6J,$4

The main quantitative results are summarized below.

Setting	FID	Freq. align	Contact consistency
Similar-skeleton	0.033	96.2%	93.5%
Cross-species	0.492	90.3%	79.7%

Motion2Motion also reports diversity and speed: diversity of 3.20 and FPS of 778 in similar-skeleton transfer, and diversity of 1.90 and FPS of 752 in cross-species transfer (Chen et al., 18 Aug 2025). The method outperforms WalkTheDog and Pose-to-Motion on all reported metrics in both settings (Chen et al., 18 Aug 2025).

The paper identifies several failure modes. The principal limitation arises when the target examples are semantically too different from the source motion, as in transferring kungfu motion to dancing examples (Chen et al., 18 Aug 2025). Quality also depends on the representativeness of the target motion set, and sparse binding quality matters more as topology differences increase (Chen et al., 18 Aug 2025). These are inherent constraints of retrieval-based completion under sparse conditioning.

The practical emphasis of the work is reinforced by a Blender add-on supporting source-motion loading, target example selection, source and target armature configuration, $D \leftarrow 3 + 6J,$5 tuning, and automatic or manual binding (Chen et al., 18 Aug 2025). This UI integration indicates that the method is presented not only as an algorithmic contribution but as a usable animation tool.

In a broader encyclopedia sense, “Motion2Motion” also names or overlaps conceptually with several adjacent research directions. In video diffusion, motion transfer methods such as “Follow-Your-Motion” transfer object trajectories, human poses, and camera movement between videos by finetuning video diffusion transformers (Ma et al., 5 Jun 2025). “SynMotion” frames motion customization as learning transferable motion concepts from a few exemplar videos for new subjects in T2V and I2V generation (Tan et al., 30 Jun 2025). “MotionShot” performs training-free motion transfer across arbitrary objects for text-to-video generation through semantic feature matching and TPS-based morphological retargeting (Liu et al., 22 Jul 2025). In 3D motion generation, “DiMo” uses the label M2M for text-free motion completion and prediction via masked discrete token refinement (Zhang et al., 4 Feb 2026). These usages are related but not identical: they concern video-level appearance-changing motion transfer, motion-conditioned generation, or token-level motion reconstruction rather than cross-topology skeletal retargeting.

Taken in this specific sense, Motion2Motion designates a sparse-correspondence framework that reframes cross-topology retargeting as motion patch matching under partial semantic constraints. Its distinctive contribution is the claim that reliable cross-species and cross-rig transfer can be achieved without training, using only one or a few target examples and a sparse set of correspondences, while remaining efficient enough for interactive or industrial workflows (Chen et al., 18 Aug 2025).