Latent-Delta Motion Guidance

Updated 5 July 2026

Latent-Delta Motion Guidance is a framework that controls motion by leveraging structured differences in latent space rather than relying solely on pixel-level changes.
It employs various formulations—including explicit temporal differences, sparse residual edits, and structured coefficient shifts—to modulate motion effectively.
This approach enhances motion fidelity and identity preservation while enabling diverse applications in image animation, video generation, and human motion modeling.

Latent-Delta Motion Guidance (Editor's term) denotes a family of motion-control strategies in which motion is represented, supervised, or steered through changes in a latent state rather than through pixel-space motion alone. In current usage, the term covers several distinct but related constructions: explicit temporal latent differences such as $\delta z^i = z^i-z^{i-1}$ or $\Delta_\tau z(i)=z(i+\tau)-z(i)$ ; sparse residual edits to latent conditions; latent coefficient shifts inside structured motion subspaces; and delta-like compact motion codes injected into diffusion backbones. The concept is therefore not a single standardized algorithm. Some systems define motion directly by inter-frame latent differences and optimize or guide them during training or sampling (Dai et al., 2023, Jiang et al., 22 May 2026, Han et al., 4 Jun 2026), whereas others are more accurately described as latent residual control, latent trajectory editing, or latent attention guidance (Xu et al., 2024, Zhao et al., 30 Jul 2025, Chu et al., 9 Dec 2025).

1. Scope, genealogy, and conceptual boundaries

A precursor to the modern usage appears in deep motion inbetweening, where a network predicts residual corrections over a simple baseline rather than absolute motion. The Deep $\Delta$ -Interpolator operates in a $\Delta$ regime over linear interpolation for root positions and spherical linear interpolation for rotations, and also shows that the same residual formulation remains viable with a last-known-frame baseline. Its central claim is that operating in a reference frame local to the inputs is more accurate and robust than operating in the global frame (Oreshkin et al., 2022). That local-residual viewpoint anticipates later latent-space formulations in which motion control is implemented as a correction, discrepancy, or transport field rather than as a full generative specification.

The term now spans several mathematically distinct mechanisms.

Formulation	Representative latent carrier	Papers
Explicit temporal latent differences	$\delta z^i$ , $\Delta_\tau z(i)$ , $\mathcal{T}(\mathbf z)$	(Dai et al., 2023, Jiang et al., 22 May 2026, Han et al., 4 Jun 2026)
Predicted-latent or velocity deltas	$\Delta(z_0(t))$ , $\Delta u_{\theta,t}$	(Wang et al., 6 Mar 2026, Shaulov et al., 1 Jun 2025)
Structured latent coefficient edits	$z=D^\top(w_0+\alpha)$ , interpolation or local shifts	(Xu et al., 2024, Takamidoa et al., 12 Oct 2025)
Sparse residual condition editing	$\Delta_\tau z(i)=z(i+\tau)-z(i)$ 0 along trajectories	(Chu et al., 9 Dec 2025)
Compact delta-like motion code	1D motion latent plus relative translation/scale	(Zhao et al., 30 Jul 2025)
Trust-region transition discrepancy	$\Delta_\tau z(i)=z(i+\tau)-z(i)$ 1	(Wu et al., 14 May 2026)
Latent conditioning without explicit delta update	Motion latent via AdaLN	(Li et al., 16 Oct 2025)

This heterogeneity matters. X-NeMo explicitly presents its fused motion latent as a delta-like control signal, but not as an explicit difference to a source latent (Zhao et al., 30 Jul 2025). RoboGhost, by contrast, uses motion latent guidance through conditioning of a diffusion policy and does not define an explicit latent-delta update in motion space (Li et al., 16 Oct 2025). Latent-Delta Motion Guidance is therefore best understood as an umbrella label for latent residualization of motion, not as a universally adopted term of art.

2. Formalizations of latent motion change

The clearest formulation is the explicit temporal-difference view. AnimateAnything defines motion strength in latent space by averaging frame-to-frame latent differences,

$\Delta_\tau z(i)=z(i+\tau)-z(i)$ 2

and aligns the predicted clean latent with the target strength through

$\Delta_\tau z(i)=z(i+\tau)-z(i)$ 3

Its targeted motion-area mechanism is not an explicit masked delta penalty; instead, supervision is constructed by freezing the non-movable region,

$\Delta_\tau z(i)=z(i+\tau)-z(i)$ 4

which enforces negligible latent changes outside the mask through the target itself (Dai et al., 2023).

LaMo generalizes the same idea from a scalar motion-strength proxy to a conditional latent motion prior. It defines the lagged latent change

$\Delta_\tau z(i)=z(i+\tau)-z(i)$ 5

with default $\Delta_\tau z(i)=z(i+\tau)-z(i)$ 6, and models a prior $\Delta_\tau z(i)=z(i+\tau)-z(i)$ 7. The prior is exposed through two readouts: a macro drift obtained by spatial averaging and trained with a Motion Drift Loss, and a micro motion field used during sampling as Motion Prior Guidance (Jiang et al., 22 May 2026). PhaseLock uses an even more direct operator,

$\Delta_\tau z(i)=z(i+\tau)-z(i)$ 8

to extract a few-step motion prior from an early denoising trajectory, then computes a residual

$\Delta_\tau z(i)=z(i+\tau)-z(i)$ 9

and injects it back into frames $\Delta$ 0 during the first half of denoising (Han et al., 4 Jun 2026).

A second formalization uses predicted latent trajectories rather than clean latent pairs. FlowMotion defines frame-wise latent deltas on predicted clean latents and aligns both the predicted latent itself and its temporal differences:

$\Delta$ 1

The difference-alignment term is the explicit latent-delta component, while a velocity regularizer constrains the update direction to remain close to the accumulated flow (Wang et al., 6 Mar 2026). FlowMo instead derives an appearance-debiased temporal representation by taking the distance between consecutive frame-wise velocity predictions, computes patch-wise temporal variance, and performs a latent refinement step that reduces the most incoherent patch (Shaulov et al., 1 Jun 2025).

A third formalization dispenses with framewise latent differences and treats motion control as structured coefficient editing. SLD learns orthogonal semantic directions $\Delta$ 2 and expresses a future-motion latent as $\Delta$ 3. Controllability is obtained by adding a coefficient delta $\Delta$ 4,

$\Delta$ 5

so the “delta” is a semantic latent offset rather than a temporal difference (Xu et al., 2024). PMGF uses the same residual logic for athlete-specific motion. Interpolation constructs

$\Delta$ 6

while local optimization applies a bounded latent shift $\Delta$ 7 on a hypersphere; both are explicit latent deltas decoded into guidance trajectories (Takamidoa et al., 12 Oct 2025).

3. Injection mechanisms inside generative and control backbones

Latent-delta formulations differ as much in their control pathway as in their mathematics. AnimateAnything applies all guidance in latent space but does not introduce gradient-based energy guidance at sampling time. Instead, a 3D U-Net is conditioned on a binary motion mask concatenated as an extra latent channel and on a motion-strength embedding concatenated with the timestep embedding; the mask channel is zero-initialized following ControlNet-style practice, and shared-noise initialization is used to reduce train-test mismatch (Dai et al., 2023). This is a conditioning-only realization of latent-delta supervision.

X-NeMo pushes that separation further. Motion is distilled into a 1D identity-agnostic latent descriptor $\Delta$ 8, augmented by a relative translation/scale triplet

$\Delta$ 9

fused into a global motion vector, and injected purely through cross-attention after each spatial transformer block. In a faithful formulation consistent with the design,

$\Delta$ 0

The paper’s design rationale is that motion must arrive without any 2D spatial control maps, because aligned spatial guidance leaks driver identity (Zhao et al., 30 Jul 2025).

A different branch uses explicit gradients during sampling. “Motion Guidance” for image editing computes a guidance loss through RAFT and backpropagates it through the decoder to the diffusion latent, yielding a direct latent update

$\Delta$ 1

or, equivalently, a guided noise prediction in the DDIM update. This is the most literal instance of latent-delta motion guidance as a gradient-induced delta in latent space (Geng et al., 2024). LaMo also guides at sampling, but in noise space rather than by editing $\Delta$ 2 directly. After classifier-free guidance mixing, it computes a motion-consistency loss on the current data-side projection and updates the predicted noise by subtracting the gradient of that loss; the paper explicitly argues that direct latent edits desynchronize the trajectory from the scheduler and empirically underperform (Jiang et al., 22 May 2026).

PhaseLock occupies an intermediate position. It keeps the sampler unchanged, but after each denoising step applies an additive correction to latent frames $\Delta$ 3 that pulls the current inter-frame latent difference toward the two-step prior. Wan-Move performs an even simpler intervention: it edits the model’s own latent condition once, before sampling, by projecting user trajectories to the latent grid and scattering first-frame latent features along those trajectories so that the condition becomes motion-aware. The edited condition can be written as $\Delta$ 4, with $\Delta$ 5 nonzero only at future latent positions specified by trajectories (Han et al., 4 Jun 2026, Chu et al., 9 Dec 2025).

Not all latent motion guidance is inference-time steering. Delta Forcing compares teacher and student semantic feature deltas,

$\Delta$ 6

measures their discrepancy $\Delta$ 7, and uses a sigmoid trust weight $\Delta$ 8 to blend a DMD loss with a continuity loss during streaming long tuning. Here the delta acts as a reliability estimator for teacher supervision rather than as a direct motion command (Wu et al., 14 May 2026). RoboGhost should be distinguished from all of these: its language-grounded motion latent conditions a diffusion student policy via a trainable latent encoder and AdaLN, but the paper explicitly states that it does not introduce classifier guidance, classifier-free guidance in motion latent space, or any explicit $\Delta$ 9 steering rule (Li et al., 16 Oct 2025).

4. Application regimes

Open-domain image animation provides one of the most direct uses of latent temporal differences. AnimateAnything targets static-image animation in diverse real environments, and its two controls—targeted motion area guidance and motion strength guidance—allow users to specify both where motion may occur and how strong that motion should be. The method is explicitly open-domain, text-conditioned, and interactive, while remaining within a latent video-diffusion framework (Dai et al., 2023).

Portrait reenactment uses latent deltas for disentanglement rather than for generic dynamics. X-NeMo frames the central problem as motion transfer under identity preservation. Its solution is to make the motion path structure-agnostic by replacing spatially aligned guidance with a compact motion latent and a relative translation/scale triplet, while appearance remains in a separate reference branch. The explicit target is to mitigate identity leakage while preserving subtle and extreme expressions (Zhao et al., 30 Jul 2025).

Video generation and motion transfer have produced the broadest set of latent-delta mechanisms. LaMo learns a self-supervised latent motion prior for physical realism in diffusion video generators. PhaseLock preserves an early motion prior because phase coherence erodes during long denoising trajectories. FlowMo and FlowMotion are training-free guidance methods that extract temporal structure from model predictions themselves: FlowMo from predicted velocity fields at each step, and FlowMotion from early latent predictions in flow-based text-to-video backbones. Wan-Move turns user-specified point trajectories into a sparse motion-aware latent condition without any motion encoder or architecture change (Jiang et al., 22 May 2026, Han et al., 4 Jun 2026, Shaulov et al., 1 Jun 2025, Wang et al., 6 Mar 2026, Chu et al., 9 Dec 2025).

Human motion modeling uses structured latent deltas rather than visual latent transport. SLD constrains stochastic human motion prediction to a span of orthogonal semantic latent directions, then controls outcomes by editing direction coefficients. PMGF learns athlete-specific latents with a Transformer-VAE and generates guidance motions either by interpolating toward a target athlete or by applying a locally optimized latent shift tied to biomechanical objectives (Xu et al., 2024, Takamidoa et al., 12 Oct 2025). These methods do not require diffusion and show that latent-delta guidance is not specific to image or video synthesis.

Humanoid control occupies a related but distinct category. RoboGhost eliminates explicit motion decoding and retargeting by conditioning a diffusion policy directly on language-grounded motion latents. The latent is the semantic anchor for denoising executable actions, but the paper states that the mechanism is conditioning rather than latent-delta steering in the strict sense (Li et al., 16 Oct 2025).

5. Empirical characteristics

Across domains, the strongest empirical results typically occur when the delta signal is used to constrain motion while leaving appearance synthesis or later refinement relatively unconstrained.

System	Task	Representative result
AnimateAnything	Open-domain image animation	Zero-shot MSR-VTT FVD 443
X-NeMo	Cross-identity portrait reenactment	ID-SIM 0.787; AED/APD 0.039/3.42; EMO-SIM 0.65
SLD	Stochastic human motion prediction	HumanEva-I ADE/FDE 0.193/0.209; Human3.6M ADE/FDE 0.348/0.436
LaMo	Physics-aware video generation	CogVideoX-2B VideoPhy 60.5/25.6 $\delta z^i$ 0 67.2/31.4 SA/PC
PhaseLock	Physics-consistent image-to-video	Average Physics-IQ gain +6.2; $\delta z^i$ 1 time; $\delta z^i$ 2 memory
FlowMotion	Training-free video motion transfer	Text 0.347; Motion 0.850; Temporal 0.986; 19.3 GB
FlowMo	Training-free motion coherence	Final Score +6.20% on Wan2.1; +5.26% on CogVideoX
Wan-Move	Motion-controllable I2V	MoveBench single-object FID 12.2; FVD 83.5; EPE 2.6

AnimateAnything’s ablations show how explicit latent-delta supervision sharpens controllability. Motion Mask Precision rises from $\delta z^i$ 3 with no control to $\delta z^i$ 4 with mask guidance and to $\delta z^i$ 5 when mask guidance is combined with the frozen-target strategy. Motion Strength Error drops from $\delta z^i$ 6 with no control to $\delta z^i$ 7 with motion-strength embedding and to $\delta z^i$ 8 when the motion-strength loss is added (Dai et al., 2023). This isolates two distinct effects: spatial localization benefits from freezing non-movable regions in the target latent, while temporal magnitude benefits from an explicit latent-difference loss.

X-NeMo’s key empirical claim is not merely higher expressiveness but improved identity-motion disentanglement. Replacing motion cross-attention with a spatial control map lowers identity similarity from $\delta z^i$ 9 to $\Delta_\tau z(i)$ 0 while offering no advantage in motion accuracy, and removing augmentations also hurts identity similarity and emotion transfer (Zhao et al., 30 Jul 2025). The data support the specific claim that structure-agnostic latent attention mitigates identity leakage.

For physical realism, LaMo reports consistent gains across scales. On VideoPhy, CogVideoX-2B improves from $\Delta_\tau z(i)$ 1 SA / $\Delta_\tau z(i)$ 2 PC to $\Delta_\tau z(i)$ 3 / $\Delta_\tau z(i)$ 4, and CogVideoX-5B from $\Delta_\tau z(i)$ 5 / $\Delta_\tau z(i)$ 6 to $\Delta_\tau z(i)$ 7 / $\Delta_\tau z(i)$ 8. On VBench, Total Score rises from $\Delta_\tau z(i)$ 9 to $\mathcal{T}(\mathbf z)$ 0, with a large gain in Spatial Relationship from $\mathcal{T}(\mathbf z)$ 1 to $\mathcal{T}(\mathbf z)$ 2 (Jiang et al., 22 May 2026). PhaseLock reports an average Physics-IQ gain of $\mathcal{T}(\mathbf z)$ 3 points across CogVideoX-5B, LTX-Video, Wan 2.1, and Wan 2.1 distill, while remaining close to baseline efficiency (Han et al., 4 Jun 2026). Both results support a common pattern: latent-delta constraints are most effective when they stabilize motion structure without replacing the backbone’s appearance model.

Training-free methods show a similar separation. FlowMotion achieves the highest reported motion fidelity and temporal consistency among compared methods while using much less memory than methods based on intermediate features or inversion. FlowMo improves motion smoothness and overall quality on Wan2.1 and CogVideoX, but with runtime increased by approximately $\mathcal{T}(\mathbf z)$ 4 because it adds backpropagation at selected denoising steps (Wang et al., 6 Mar 2026, Shaulov et al., 1 Jun 2025). Wan-Move demonstrates that sparse residual condition editing can scale efficiently: on MoveBench single-object control it reports FID $\mathcal{T}(\mathbf z)$ 5, FVD $\mathcal{T}(\mathbf z)$ 6, PSNR $\mathcal{T}(\mathbf z)$ 7, SSIM $\mathcal{T}(\mathbf z)$ 8, and EPE $\mathcal{T}(\mathbf z)$ 9, and human 2AFC results indicate competitiveness with Kling 1.5 Pro’s commercial Motion Brush (Chu et al., 9 Dec 2025).

Structured latent-delta methods in human motion prediction and coaching also show strong controllability-accuracy trade-offs. SLD improves on prior ADE/FDE results while preserving diversity through learned motion queries (Xu et al., 2024). PMGF reports smooth transitions across all $\Delta(z_0(t))$ 0 pitcher pairs and significant increases in stride length and knee extension across all five runs of the local-optimization setting, with reconstruction RMSE of $\Delta(z_0(t))$ 1 cm (Takamidoa et al., 12 Oct 2025).

6. Misconceptions, limitations, and open directions

A common misconception is that latent-delta motion guidance is synonymous with optical flow guidance or trajectory conditioning. That is not the case. AnimateAnything does not require optical flow or explicit trajectories, instead using a mask channel and a scalar motion-strength embedding (Dai et al., 2023). X-NeMo avoids spatially aligned control maps entirely, specifically to reduce identity leakage (Zhao et al., 30 Jul 2025). Wan-Move does depend on dense point trajectories, but only after projecting them into latent space and turning them into a sparse residual condition (Chu et al., 9 Dec 2025).

A second misconception is that all such methods rely on sampling-time gradient guidance. AnimateAnything explicitly does not add gradient-based motion terms during sampling (Dai et al., 2023). LaMo uses noise-space guidance late in sampling, FlowMo performs latent refinement only at selected early–mid steps, Motion Guidance backpropagates a differentiable motion-estimation loss through the decoder, and PhaseLock applies a direct latent correction after each denoising step in the first half of the trajectory (Jiang et al., 22 May 2026, Shaulov et al., 1 Jun 2025, Geng et al., 2024, Han et al., 4 Jun 2026). The design space therefore includes conditioning-only, latent-edit, noise-space, and trust-region variants.

The term also should not be overextended. X-NeMo’s motion signal is delta-like rather than an explicit difference to the source latent. SLD and PMGF use latent deltas as semantic coefficient or interpolation shifts in a structured motion space. Delta Forcing measures a transition discrepancy in a semantic feature space to regulate teacher supervision. RoboGhost does not define a latent-delta motion mechanism at all, even though it is a latent-guided motion-control system (Zhao et al., 30 Jul 2025, Xu et al., 2024, Takamidoa et al., 12 Oct 2025, Wu et al., 14 May 2026, Li et al., 16 Oct 2025).

Limitations are correspondingly domain-specific. AnimateAnything is constrained by mask quality, prompt-image mismatch, and training at relatively low or medium resolution (Dai et al., 2023). X-NeMo is trained only on real human videos and struggles with out-of-domain styles such as 3D cartoons and with exaggerated expressions absent from the data (Zhao et al., 30 Jul 2025). SLD notes under-representation of rare or abrupt motions, difficulty with very long horizons, and lack of scene or interaction modeling (Xu et al., 2024). LaMo emphasizes that it is not a hard physics constraint or simulator, and that rare interactions and long-horizon conservation are not guaranteed (Jiang et al., 22 May 2026). PhaseLock can propagate an incorrect two-step motion prior if the early prior is already wrong (Han et al., 4 Jun 2026). FlowMotion reports failures on complex pose-dependent actions, many-object scenes, and large semantic gaps between source and target (Wang et al., 6 Mar 2026). FlowMo notes an oversmoothing risk if variance minimization is applied too aggressively (Shaulov et al., 1 Jun 2025). Wan-Move remains sensitive to occlusions, point visibility, crowded scenes, and tracking errors (Chu et al., 9 Dec 2025).

Several explicitly stated research directions converge on richer constraints rather than larger deltas. LaMo proposes 3D-aware priors, learned $\Delta(z_0(t))$ 2-adaptation, multi-lag ensembles, and explicit Newtonian constraints (Jiang et al., 22 May 2026). PMGF proposes a constrained general-PMGF with bodily, environmental, and task constraints, and a human-in-the-loop component (Takamidoa et al., 12 Oct 2025). Delta Forcing identifies alternative feature spaces and harder trust-region constraints as possible extensions (Wu et al., 14 May 2026). Wan-Move points to depth- and pose-derived trajectories for camera control and to optional regularizers in atypical domains (Chu et al., 9 Dec 2025). Taken together, these directions suggest that latent-delta motion guidance is evolving less toward a single canonical objective than toward a broader design principle: motion should be represented as a structured latent change that is easy to constrain, compose, and preserve without erasing the generative model’s native appearance capabilities.