Motion Alignment: Principles & Methods

Updated 4 July 2026

Motion alignment is the enforcement of correspondence between motion-bearing representations, using temporal reparameterization, semantic grounding, and cross-modal synchronization.
It spans diverse applications such as human motion timing, text-driven generation, video editing, and sensor fusion, demonstrating flexibility across domains.
Key challenges include balancing local versus global alignment and managing non-monotonic correspondences while leveraging latent dynamics for improved synchronization.

Motion alignment denotes a family of problems in which motion signals are brought into correspondence across time, modalities, semantic descriptions, or interacting agents. In the literature surveyed here, the term covers temporal reparameterization of human motion trajectories, token- and segment-level motion-language grounding, alignment between inertial streams and skeletal motion, motion-preserving video editing and compression, and collective orientation or polarity alignment in social and biological systems. This suggests that “motion alignment” is not a single algorithmic object but a recurring structural principle: the enforcement of correspondence between motion-bearing representations under task-specific constraints (Tumpach et al., 2023, Gu et al., 3 Apr 2026, Sarker et al., 2 Jun 2025, Matsushita et al., 2019, Bhowmik et al., 21 Oct 2025).

1. Conceptual scope and definitions

Different research communities use the term in substantially different ways. In temporal motion analysis, alignment typically means estimating a time reparameterization between executions of the same action. In multimodal generation, it usually means matching motion dynamics to text, audio, or another signal. In collective-behavior studies, it can mean orientation coordination rather than velocity matching. In video systems, it often means preserving or reconstructing motion while appearance, semantics, or bitrate constraints vary.

Domain	Alignment object	Representative papers
Human-motion timing and control	Temporal correspondence or trajectory progress	(Tumpach et al., 2023, Agata et al., 26 May 2025, Cuellar et al., 19 Nov 2025)
Text-driven human motion	Global semantics, local tokens, segments, or denoising steps	(Gu et al., 3 Apr 2026, Chen et al., 29 Jan 2026, He et al., 15 Dec 2025, Weng et al., 24 Nov 2025)
Sensor and video representations	IMU-pose synchronization, motion subspaces, dynamic/state semantics	(Nguyen et al., 22 Feb 2026, Bhowmik et al., 21 Oct 2025, Su et al., 2023)
Video editing, animation, and compression	Motion preservation under concept transfer, preference tuning, or coding	(Park et al., 2024, Zhang et al., 1 Jun 2025, Liang et al., 11 Jun 2025, Zhang et al., 15 Dec 2025)
Social and biological collectives	Orientation geometry or polarity-based collective order	(Sarker et al., 2 Jun 2025, Matsushita et al., 2019)

A common misconception is that motion alignment is equivalent to direct velocity matching. The social-orientation and cell-migration literature explicitly contradicts that narrow view. Another misconception is that alignment is necessarily pairwise and monotone in time. Training-free control via Gromov-Wasserstein matching and HMM-based alignment to representative demonstrations show that softer, non-DTW-style formulations are also operative in the literature. This suggests that the object being aligned may be frames, latent tokens, body parts, orientations, or distributions rather than raw trajectories alone.

2. Temporal correspondence, reparameterization, and control

A foundational formulation appears in temporal alignment of human motion data, where a motion is modeled as a curve $M:[0,1]\to\mathbb{R}^{3N}$ and the problem is to find a monotone time change $\varphi:[0,1]\to[0,1]$ such that $M_2\circ\varphi$ is “visually as close as possible” to $M_1$ . The geometric account in "Temporal Alignment of Human Motion Data: A Geometric Point of View" formalizes temporal reparameterizations as an action of $\operatorname{Diff}^+([0,1])$ , interprets aligned motions to a template as a slice of the resulting orbit structure, and proposes a consistency check based on recovering $\varphi^{-1}$ from synthetically warped examples (Tumpach et al., 2023). The same paper argues that many alignment pipelines are too invariant: full rigid-motion invariance discards vertical information that is crucial for synchronization in actions such as tennis strokes, and anchored dynamic programming improves both accuracy and runtime by forcing the warping path through keyframe correspondences derived from arm elevation (Tumpach et al., 2023).

A more general control-oriented view appears in "MAMM: Motion Control via Metric-Aligning Motion Matching," where the source motion $X$ , control sequence $Y$ , and aligned motion $X'$ are split into temporal patches and aligned by comparing within-domain distance matrices rather than by defining a cross-domain distance directly. The method uses a soft transport plan $T\in\mathbb{R}^{L_Y\times L_X}$ and an FSUGW objective that combines a Gromov-Wasserstein term, preserving the metric structure of control patches and motion patches, with a Wasserstein term that keeps $\varphi:[0,1]\to[0,1]$ 0 close to source motion patches (Agata et al., 26 May 2025). The paper is explicit that there is no explicit monotonicity or one-to-one DTW constraint; temporal consistency emerges from patch overlap, blending, and fidelity to the source motion. This extends motion alignment beyond same-domain synchronization to sketches, labels, audio features, and other temporal controls without paired training data (Agata et al., 26 May 2025).

A third formulation appears in learning from demonstration. "An Alignment-Based Approach to Learning Motions from Demonstrations" introduces CALM, which defines progress through a task not by absolute time or purely by current state, but by the posterior $\varphi:[0,1]\to[0,1]$ 1 over alignment to a representative mean trajectory $\varphi:[0,1]\to[0,1]$ 2. The controller follows the normalized gradient of a scalar field

$\varphi:[0,1]\to[0,1]$ 3

with $\varphi:[0,1]\to[0,1]$ 4 a Gaussian likelihood around mean point $\varphi:[0,1]\to[0,1]$ 5, and updates the alignment posterior online via an HMM recursion (Cuellar et al., 19 Nov 2025). The paper proves global asymptotic stability for the induced vector field and shows that strict forward transition rules guarantee convergence to the mean-trajectory endpoint, while more permissive transitions allow perturbation-driven realignment and cluster switching (Cuellar et al., 19 Nov 2025). Motion alignment here is neither simple time warping nor state feedback; it is a latent progress variable inferred from the executed partial trajectory.

3. Motion-language and text-conditioned alignment

In text-driven human motion generation, the dominant issue is no longer only temporal correspondence but semantic grounding. "Exploring Motion-Language Alignment for Text-driven Motion Generation" distinguishes between global motion structure and fine-grained local semantics, arguing that whole-sentence embeddings are adequate for coarse intent but not for token-level grounding. MLA-Gen therefore combines memory slots as global motion priors, frame-to-token cross-attention for local motion-language alignment, and sink-aware masking and control to address the “attention sink phenomenon,” in which attention collapses onto the start token rather than informative words (Gu et al., 3 Apr 2026). Its analysis introduces the metric

$\varphi:[0,1]\to[0,1]$ 6

used with top-2 attention weights, and couples that statistic to inference-time guidance control. On HumanML3D, the paper reports consistent gains over ACMDM in R-Precision, Matching Score, and CLIP Score, alongside larger FID improvements, which it interprets as better motion-language alignment without sacrificing realism (Gu et al., 3 Apr 2026).

The retrieval literature pushes the same idea toward finer decomposition. "Beyond Global Alignment: Fine-Grained Motion-Language Retrieval via Pyramidal Shapley-Taylor Learning" argues that global sequence-sentence matching overlooks correspondences between body joints and word tokens, and between motion segments and phrases. PST therefore organizes alignment hierarchically—joint-wise, segment-wise, and holistic—and uses Shapley-Taylor Interaction to estimate pairwise token interactions: $\varphi:[0,1]\to[0,1]$ 7 These interactions are converted into token-level correspondence distributions and distilled into an estimation head, while contrastive retrieval losses are applied at all three levels (Chen et al., 29 Jan 2026). The paper reports improved retrieval on HumanML3D and KIT-ML and uses qualitative visualizations to show segment-wise mappings such as “walks forward” or “arms up and down” to corresponding motion segments (Chen et al., 29 Jan 2026).

MoLingo addresses the same problem from the latent-space side. It asks how to build a semantically aligned continuous latent space so that diffusion over motion latents becomes easier, and how to inject text so that the motion follows the description closely. Its semantic-aligned autoencoder uses frame-level BABEL labels to construct text-derived class tokens $\varphi:[0,1]\to[0,1]$ 8 for each latent token and minimizes a cosine alignment loss

$\varphi:[0,1]\to[0,1]$ 9

with repetitive-token filtering at threshold $M_2\circ\varphi$ 0 (He et al., 15 Dec 2025). The paper reports that multi-token cross-attention with frozen T5-Large outperforms single-token AdaLN conditioning, and that the semantically aligned latent space improves R-Precision and CLIP-Score even when the lowest FID is achieved by a different tokenizer variant (He et al., 15 Dec 2025).

ReAlign shifts the locus of alignment from training losses to inference dynamics. It defines an ideal aligned distribution

$M_2\circ\varphi$ 1

trains a step-aware reward model by prepending a timestep token $M_2\circ\varphi$ 2 to noisy motion inputs, and adds $M_2\circ\varphi$ 3 to reverse diffusion updates (Weng et al., 24 Nov 2025). The reward combines a text-aligned cosine similarity term and a motion-aligned term based on the nearest retrieved training motion for the same text. Ablations show that the text-aligned reward is the dominant source of semantic gains, while step awareness is necessary for stable guidance on noisy intermediate motions (Weng et al., 24 Nov 2025). Together, these papers indicate that text-motion alignment now spans latent geometry, token-level correspondence, attention regulation, and inference-time reward steering.

4. Representation alignment across video, pose, and inertial signals

A different branch of work treats motion alignment as correspondence between heterogeneous sensor modalities or internal representations. "MoBind: Motion Binding for Fine-Grained IMU-Video Pose Alignment" argues that IMUs should be aligned with 2D skeletal motion rather than raw RGB because inertial sensors are localized and primarily encode motion. MoBind decomposes full-body pose into $M_2\circ\varphi$ 4 body parts, pairs each with its corresponding IMU stream, and aligns the modalities at three levels: token-level temporal segments, local body-part embeddings, and global whole-body embeddings (Nguyen et al., 22 Feb 2026). The token-level loss uses within-pair temporal negatives,

$M_2\circ\varphi$ 5

and is designed for about $M_2\circ\varphi$ 6 s resolution on 5-second windows (Nguyen et al., 22 Feb 2026). The resulting model improves cross-modal retrieval, temporal synchronization, subject localization, body-part localization, and action recognition across mRi, TotalCapture, and EgoHumans, with synchronization MAE as low as $M_2\circ\varphi$ 7– $M_2\circ\varphi$ 8 s on two of the three datasets (Nguyen et al., 22 Feb 2026).

MoAlign applies alignment not across external modalities but between a video diffusion model and a motion-only representation learned from a frozen pretrained video encoder. The core claim is that raw encoder features entangle appearance and dynamics, so the target of alignment must itself be motion-disentangled. A low-dimensional bottleneck $M_2\circ\varphi$ 9 is trained to predict RAFT-computed optical flow via $M_1$ 0, and diffusion hidden states are then projected into the same space and aligned through spatial and temporal token-relation matrices rather than direct feature matching (Bhowmik et al., 21 Oct 2025). The temporally weighted alignment loss improves physical commonsense on VideoPhy and VideoPhy2 while preserving prompt adherence better than raw-feature alignment baselines, though the paper also notes a tradeoff with reduced motion magnitude on some benchmarks (Bhowmik et al., 21 Oct 2025).

Motion-state alignment for video semantic segmentation adopts yet another representation split. MSAF separates dynamic semantics, learned by a motion alignment branch across neighboring frames, from static semantics, learned by a state alignment branch across current-frame feature stages. The motion branch uses a decoupled transformer with a Pyramid Spatial Transformer and an Aligned Temporal Transformer, including deformable alignment

$M_1$ 1

to associate current-frame pixels with corresponding positions in adjacent frames (Su et al., 2023). Dynamic semantics are converted into class-wise region descriptors, while static semantics become pixel descriptors; semantic assignment via minimum cosine distance then yields the final segmentation (Su et al., 2023). Taken together, these works suggest that motion alignment increasingly operates through carefully designed intermediate spaces rather than direct pixel, token, or frame matching alone.

5. Motion-preserving alignment in generation, editing, and compression

In video generation and editing, motion alignment often means preserving a source motion trajectory while changing some other attribute. "Spectral Motion Alignment for Video Motion Transfer using Diffusion Models" critiques consecutive-frame residuals as motion proxies because they lack global temporal context and are contaminated by spatial artifacts. SMA retains residual-based motion vectors but augments alignment with a wavelet-domain global loss over temporal motion sequences and a Fourier-domain local loss over motion-vector amplitude and phase spectra, with low-frequency weighting to suppress artifact-prone high-frequency components (Park et al., 2024). The combined objective improves motion transfer across MotionDirector, VMC, Tune-A-Video, and ControlVideo while adding little computational cost (Park et al., 2024).

"Motion-Aware Concept Alignment for Consistent Video Editing" addresses concept transfer into a moving object while preserving the original motion path. MoCA-Video first inverts the base video into the latent diffusion trajectory, tracks the target object by class-agnostic segmentation and overlap maximization, performs spatially localized latent mixing

$M_1$ 2

and then applies momentum-corrected DDIM denoising and gamma residual noise stabilization to reduce jitter and flicker (Zhang et al., 1 Jun 2025). The paper introduces CASS, alongside SSIM and LPIPS-T, to measure whether the edited video moves semantically toward the conditioned concept while remaining temporally coherent (Zhang et al., 1 Jun 2025).

In audio-driven human animation, "AlignHuman" reframes alignment as post-training preference optimization over two competing objectives: motion naturalness and visual fidelity. Its central observation is timestep-specific: early denoising timesteps mainly control motion dynamics, whereas later timesteps govern fidelity and human structure. Timestep-Segment Preference Optimization therefore trains two LoRA experts, one active in the early motion interval and one in the late fidelity interval, with a switch best placed at $M_1$ 3 (Liang et al., 11 Jun 2025). Relative to the base model, TPO improves HKV from $M_1$ 4 to $M_1$ 5, HKC from $M_1$ 6 to $M_1$ 7, and FVD from $M_1$ 8 to $M_1$ 9, while also enabling a reported 3.3× speedup from 100 to 30 NFEs (Liang et al., 11 Jun 2025). Here, motion alignment is operationalized as alignment to human preference for temporally coherent, natural movement rather than alignment to an explicit kinematic target.

Video compression uses the term more literally. CAMA inserts motion alignment between motion estimation and temporal prediction. It first warps reference features by reconstructed flow, then refines them through flow-guided deformable warping with coarse and fine offsets,

$\operatorname{Diff}^+([0,1])$ 0

and complements this with multi-reference quality-aware training and a training-free smooth-motion-estimation module (Zhang et al., 15 Dec 2025). The paper reports that the Two-Stage Motion Compensation module alone yields a $\operatorname{Diff}^+([0,1])$ 1 BD-rate reduction, while the full framework achieves $\operatorname{Diff}^+([0,1])$ 2 BD-rate savings over DCVC-TCM (Zhang et al., 15 Dec 2025). Across these works, motion alignment functions as a preservation constraint: maintain the original dynamics while the surrounding objective—semantics, preference, or bitrate—changes.

Outside generative modeling, motion alignment has a distinct meaning in collective behavior. "Alignment Phase Transition in Socially Driven Motion" studies low-speed preschool interactions and defines alignment through body orientation relative to the dyadic axis, using

$\operatorname{Diff}^+([0,1])$ 3

A Fourier decomposition of empirical $\operatorname{Diff}^+([0,1])$ 4 reveals three distance-dependent mechanisms—parallelization, opposition, and reciprocation—captured by the pseudo-potential

$\operatorname{Diff}^+([0,1])$ 5

The control parameter $\operatorname{Diff}^+([0,1])$ 6 changes sign near $\operatorname{Diff}^+([0,1])$ 7, separating a short-range side-by-side regime from a longer-range face-to-face regime (Sarker et al., 2 Jun 2025). Motion alignment here is not velocity consensus but socially mediated orientation geometry.

"Cell Motion Alignment as Polarity Memory Effect" proposes yet another mechanism. In a cellular Potts model with excluded-volume interactions and no explicit polarity-polarity coupling, polarity vectors evolve toward past velocity directions, yielding a memory effect that can align cell motion collectively (Matsushita et al., 2019). The central criterion is not a Vicsek-style alignment rule but a geometric relation between the persistent length $\operatorname{Diff}^+([0,1])$ 8 and average cell-cell distance $\operatorname{Diff}^+([0,1])$ 9: $\varphi^{-1}$ 0 This threshold explains why collective alignment emerges only when polarity memory is long enough relative to intercellular spacing (Matsushita et al., 2019). The result broadens the concept of motion alignment from explicit alignment interactions to emergent order induced by memory and collisions.

Across the surveyed literature, several recurring limitations and controversies appear. Global alignment often underuses local semantics, motivating token-wise, frame-wise, or body-part-wise mechanisms (Gu et al., 3 Apr 2026, Chen et al., 29 Jan 2026, Nguyen et al., 22 Feb 2026). Alignment targets can be entangled with appearance, prompting motion-only subspaces or motion-aware conditioning (Bhowmik et al., 21 Oct 2025, Park et al., 2024). Some methods benefit from soft, non-monotone correspondences, but this flexibility can weaken guarantees about temporal order (Agata et al., 26 May 2025). Others offer convergence or stability results, but only under restrictive transition rules (Cuellar et al., 19 Nov 2025). Dataset annotation granularity also remains a bottleneck: motion-language work repeatedly notes that holistic captions are weaker supervision for local alignment than frame- or body-part-level labels (He et al., 15 Dec 2025, Chen et al., 29 Jan 2026). A plausible implication is that future work will continue to move away from monolithic “motion alignment” toward explicit decomposition into the object being aligned—time, tokens, regions, body parts, orientations, or latent dynamics—and the invariances that each task can legitimately assume.