acceptodds
Under review as a conference paper at ICLR 2027

SceShortcut: Efficient Scene-Aware Motion Synthesis via Shortcut Modeling

Abstract

Previous works on scene-aware human motion synthesis have mainly focused on motion-scene compatibility, with less attention to runtime efficiency. In this work, we present SceShortcut, a shortcut-based framework that accelerates scene-aware motion synthesis through one-step or few-step sampling while producing plausible and scene-compatible motion. To preserve fine-grained scene adaptation under a limited sampling budget, we introduce a global-to-local architecture with two stages trained jointly end-to-end. Specifically, at each sampling step, conditioned on the text instruction and global geometric-semantic scene features, the global stage produces intermediate motion features and predicts a root transport velocity. This velocity yields a coarse clean root motion estimate for querying fine-grained local scene geometry from a dense height field. The local stage incorporates text conditioning and the queried geometry into these intermediate features to predict transport velocities for the remaining motion components and a residual correction to the global-stage root transport velocity. To improve coarse root motion estimates and thereby enhance generation quality, we augment shortcut training with additional root-transport-velocity supervision and root-residual regularization. On TRUMANS and HUMANISE, one-step sampling with SceShortcut achieves approximately speedups over recent advanced open-source baselines while producing plausible motion with comparable motion-scene alignment.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Scene-aware human motion generation, diffusion acceleration, controllable generation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_7W6oOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the computational efficiency challenge in scene-aware human motion synthesis by proposing SceShortcut, a framework based on shortcut modeling that enables one-step or few-step sampling. The key innovation is a global-to-local two-stage architecture: the global stage predicts coarse root transport velocity to guide spatial movement, while the local stage performs geometry-informed queries to generate full-body motion sequences conditioned on local scene geometry. Experiments on TRUMANS and HUMANISE datasets demonstrate 50-100× speedup compared to diffusion-based methods while maintaining comparable motion-scene alignment quality. The paper includes comprehensive ablation studies validating each architectural component and demonstrates the framework's extensibility to multi-step sampling for quality-efficiency trade-offs.

Critical strengths

1.The 50-100× speedup over diffusion baselines enables one-step or few-step scene-aware motion synthesis, transforming the task from computationally prohibitive to real-time viable for VR and robotics applications.

2.The global-to-local design effectively decouples coarse spatial navigation (root transport velocity) from fine-grained motion generation, reducing computational dimensionality while preserving geometric guidance—ablation studies convincingly validate each component's contribution.

3.Evaluation on two distinct datasets (TRUMANS, HUMANISE) demonstrates generalization across motion tasks, with appropriate motion quality and scene alignment metrics, complemented by thorough comparisons to relevant baselines.

4.The method supports flexible sampling steps from 1 to 4 and can be extended to downstream control tasks such as motion inbetweening and endpoint control.

Critical weaknesses

1.A well-known trade-off in shortcut and consistency-based generative models is the potential sacrifice of motion diversity for inference speed, as they compress complex multi-step trajectories into direct mappings. While the paper highlights a 50–100× speedup, it lacks a comprehensive quantitative evaluation of motion diversity (e.g., diversity scores or variance across different stochastic realizations for the same text-scene pair). It remains unclear whether 1-step shortcut generation produces over-deterministic or averaged motions compared with multi-step diffusion baselines.

2.The proposed global-to-local architecture relies on predicting an estimated end-root state R^1G\hat{R}_1^G within a single step to query local scene geometry. However, real motion trajectories evolve dynamically over both space and time. Using fixed spatial conditions derived from R^1G\hat{R}_1^G to guide the full motion sequence introduces spatio-temporal asymmetry. The paper lacks in-depth analysis of whether this static geometry conditioning induces subtle foot-sliding or trajectory drift, especially for motions with long spatial transitions.

3.While the limitations of extreme 1-step generation (e.g., jitter, penetration) are acknowledged, the paper lacks a systematic quantitative breakdown (such as failure frequencies or error distributions across the test set).

Questions for authors

1.The global-to-local architecture queries local geometry based on a single-step estimated end-root state R^1G\hat{R}_1^G. How does this static spatial conditioning handle long-range or complex spatial transitions? Specifically, did the authors observe or measure any trajectory drift or foot-sliding in intermediate frames before reaching R^1G\hat{R}_1^G?

2.Since shortcut and consistency models compress multi-step trajectories into direct 1-step mappings, how does the model's generative diversity compare against multi-step diffusion baselines? Could the authors provide diversity scores or variance metrics across different stochastic realizations under the same text-scene condition?

3.While Appendix E acknowledges minor artifacts (e.g., jitter and penetration) under extreme 1-step generation, could the authors provide a systematic quantitative breakdown over the full test set? What are the exact collision/penetration rates or error distributions when reducing sampling steps from 4-step down to extreme 1-step inference?

24 comments
AuthorsOpenReview ↗

Response to Reviewer 7W6o

We sincerely thank the reviewer for recognizing the significance of efficient scene-aware motion synthesis, the effectiveness of our global-to-local architecture, and the substantial inference speedups achieved by our method. We also appreciate the constructive feedback regarding motion diversity under different sampling budgets and stochastic realizations, potential trajectory drift and foot sliding associated with within-step geometry conditioning during long-range motion generation, and motion artifact frequencies and error distributions under different sampling budgets. To address these concerns, we provide additional experimental analyses and clarifications below.

AuthorsOpenReview ↗

Q1. Evaluation of Motion Diversity

Comparison of motion diversity. We agree that maintaining motion diversity is important for one-step and few-step generation. Our original submission already reports motion diversity on TRUMANS and HUMANISE in Tables 1 and 2, using the Diversity (Div.) and Average Pairwise Distance (APD) metrics, respectively. These metrics are computed using the official evaluation code provided by SceMoS and Afford@2, respectively. For clarity, we summarize these previously reported results in Table 1 of this rebuttal, covering our method under different sampling budgets and existing baselines.

Table 1: Comparison of motion diversity across different sampling budgets and baseline methods on TRUMANS and HUMANISE. With the exception of the GT diversity score on TRUMANS, all results are summarized from Tables 1 and 2 of our original submission. Div. and APD are the respective diversity metrics for the two datasets and are not directly comparable.

DatasetMethodDiv.APD
TRUMANSGT2.81/
TRUMANS2.71/
TeSMo2.02/
SceMoS (Officially reported)2.67/
SceMoS (Our evaluation)2.59/
SceShortcut-1step (Ours)2.69/
SceShortcut-2step (Ours)2.72/
SceShortcut-4step (Ours)2.70/
HUMANISEGT//
HUMANISE/6.271
Afford@2/2.625
MoST/2.108
SceShortcut-1step (Ours)/2.356
SceShortcut-2step (Ours)/2.275
SceShortcut-4step (Ours)/2.301

As shown in Table 1, on TRUMANS, our one-step model achieves a diversity score of 2.69, only slightly lower than that of the GT (2.81), and comparable to our 2-step (2.72) and 4-step (2.70) variants, as well as existing baselines such as TRUMANS (2.71) and SceMoS (2.67). On HUMANISE, our one-step model achieves an APD of 2.356, comparable to our multi-step variants and other baselines such as Afford@2 (2.625) and MoST (2.108). These results suggest that reducing the sampling budget from four steps to one does not substantially compromise the dataset-level motion diversity of our method. Moreover, apart from the HUMANISE baseline, which exhibits a substantially higher APD, our one-step model achieves broadly comparable dataset-level motion diversity to the other evaluated baselines, including diffusion-based and autoregressive methods.

Diversity under identical text-scene conditions. Following the reviewer's suggestion, we further evaluate motion diversity across different stochastic realizations under identical text-scene conditions on the TRUMANS dataset. Specifically, for each fixed text-scene condition, we generate 10 motion sequences using stochastic sampling while keeping all other conditioning inputs unchanged. We then measure their diversity using mean pairwise distance and motion variance, as reported in Table 2.

Table 2 Evaluation of conditional motion diversity under identical text-scene conditions on TRUMANS. For each condition, 10 motion sequences are generated using independent stochastic samples. We report the mean pairwise distance and motion variance across the generated sequences.

MethodPairwise DistanceMotion Variance
TRUMANS13.450.0041
SceMoS10.010.0031
SceShortcut-1step (Ours)15.040.0050
SceShortcut-2step (Ours)15.490.0052
SceShortcut-4step (Ours)15.870.0054

As shown in Table 2, increasing the number of sampling steps leads to a modest improvement in conditional motion diversity. Specifically, compared with one-step generation, four-step generation increases the mean pairwise distance by approximately 5.5% and motion variance by 8.0%. These results indicate that reducing the number of generation steps from four to one preserves most of the conditional motion diversity, with only a limited reduction in stochastic variation.

Furthermore, our one-step model achieves higher pairwise distance and motion variance than both TRUMANS and SceMoS under identical text-scene conditions. The relatively low conditional diversity of SceMoS may be related to its autoregressive token prediction process, which could produce more concentrated motion samples under fixed conditions. Meanwhile, TRUMANS requires additional conditioning on motion history, target pelvis positions, and action labels, which may further constrain the variability of generated motions.

AuthorsOpenReview ↗

Q2. Spatio-Temporal Geometry Conditioning, Trajectory Drift, and Foot Sliding

As described in Section 3.3 and Eq. (5), our global stage predicts a coarse clean root trajectory covering the entire motion sequence, rather than a single terminal root position. Based on this trajectory, we query local heightmap patches at different spatial locations along the motion sequence. Therefore, although the coarse trajectory is estimated within a single sampling step, the queried local geometry varies spatially along the predicted motion trajectory rather than remaining fixed at a single location.

Furthermore, within the same sampling step, our local stage incorporates the queried geometry to predict full-body motion and a residual correction to the coarse root trajectory. The additional root supervision and root residual regularization are designed to improve coarse root estimation and limit discrepancies between the queried geometry and the final generated trajectory. For multi-step inference, the coarse root trajectory and corresponding local geometry conditions are recomputed at every sampling step.

To investigate potential trajectory drift and foot sliding associated with within-step geometry conditioning during long-range motion generation, we evaluate the final root-trajectory accuracy, coarse-to-final trajectory consistency, and foot sliding across different spatial transition lengths on TRUMANS. Specifically, we rank all test motions according to the cumulative horizontal travel distance of their ground-truth root trajectories and divide them into three approximately equal-sized groups: Short (bottom 33.3%), Medium (33.3%–66.7%), and Long (top 33.3%).

We evaluate four root-trajectory metrics:

  • FTE (Final Root Translation Error) and FRE (Final Root Rotation Error): Measure the translation and rotation errors between the final generated root trajectory and the ground-truth root trajectory, respectively.

  • CF-TE (Coarse-to-Final Root Translation Error) and CF-RE (Coarse-to-Final Root Rotation Error): Measure the translation and rotation discrepancies between the coarse root-trajectory estimate and the final generated root trajectory, respectively.

Translation errors are measured in meters, while rotation errors are measured in degrees. We additionally evaluate the foot-sliding rate, defined as the percentage of foot-contact frames exhibiting horizontal foot displacement.

To analyze potential trajectory drift over time, we divide each motion sequence into two equal temporal halves and compute these metrics separately for the first and second halves. In particular, changes in FTE and FRE between the two halves allow us to assess whether trajectory deviations relative to ground truth become more pronounced in later frames, providing an indication of potential drift-related error growth. Meanwhile, CF-TE and CF-RE characterize the consistency between the queried coarse root trajectory and the final generated trajectory. We also examine whether trajectory drift-related errors and foot sliding become more pronounced as spatial transition length increases. The corresponding results are reported in Table 3.

Table 3 Analysis of potential trajectory drift, coarse-to-final root-trajectory consistency, and foot sliding across different spatial transition lengths on TRUMANS. All metrics are evaluated separately for the first and second halves of each motion sequence.

Spatial Transition LengthTemporal HalfFTE ↓FRE ↓CF-TE ↓CF-RE ↓Foot-Sliding Rate ↓
GT/////0.14
ShortFirst half0.1613.40.021.840.15
Second half0.2014.60.031.930.16
MediumFirst half0.2718.50.031.920.16
Second half0.3321.90.041.870.17
LongFirst half0.3527.40.031.860.16
Second half0.4334.30.031.920.16

As shown in Table 3, FTE and FRE increase with spatial transition length and are consistently higher in the second half of each motion sequence. In particular, for long-range motions, FTE increases from 0.35 to 0.43 m and FRE from 27.4° to 34.3°, indicating larger trajectory deviations in later frames and suggesting potential drift-related error growth. However, the coarse-to-final trajectory discrepancies remain small and relatively stable across different spatial transition lengths and temporal halves, with CF-TE ranging from 0.02 to 0.04 m and CF-RE from 1.84° to 1.93°. Furthermore, the foot-sliding rates remain close to the ground-truth reference (0.14), ranging from 0.15 to 0.17, without a clear increasing trend for longer spatial transitions. These results suggest that, although longer motions exhibit greater trajectory drift-related errors, our method maintains reasonable coarse-to-final trajectory consistency and foot stability, demonstrating a degree of robustness during long-range motion generation.

Reviewer_vDfqOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SceShortcut, an efficient framework for scene-aware human motion synthesis using shortcut modeling. The method introduces a within-step global-to-local architecture that first predicts coarse root motion to query local scene geometry and subsequently refines the root motion while generating full-body motion. Additional root-transport supervision and residual regularization are introduced to improve generation quality. Experiments on TRUMANS and HUMANISE demonstrate substantial inference speedups while maintaining competitive motion quality and scene compatibility. Overall, the paper addresses a practical and meaningful problem, and the reported efficiency improvements are promising. However, several aspects of the evaluation protocol and methodological validation could be strengthened.

Critical strengths

Meaningful problem and significant efficiency improvements. The paper addresses the computational limitations of existing scene-aware motion generation methods. The reported one-step inference achieves substantial speedups while maintaining competitive motion quality and scene consistency.

Reasonable and effective architecture. The within-step global-to-local design provides an intuitive solution to incorporating fine-grained local scene geometry into one-step motion generation. The coarse root prediction and subsequent local refinement are well motivated.

Relatively comprehensive experimental validation. The paper evaluates the method on two datasets and provides quantitative comparisons, qualitative results, and several ablation studies. The experiments generally support the effectiveness of the proposed architecture and additional training objectives.

Critical weaknesses

Evaluation fairness and inference-time reporting need further clarification. The compared methods have different conditioning requirements and inference procedures. In particular, Appendix B reports substantial discrepancies between the offline and online inference modes of SceMoS, making the interpretation of the baseline results challenging. Although the authors discuss these discrepancies, the exact reproduction configurations and their implications for fair comparison should be clarified. Furthermore, Appendix C indicates that the TRUMANS runtime measurement covers fewer autoregressive windows than those used for the full motion evaluation. A standardized end-to-end latency comparison, explicitly accounting for all preprocessing, scene encoding, and generation operations, would strengthen the efficiency claims.

The key mechanisms could be more directly validated. While the ablation studies demonstrate the benefits of the overall architecture and additional losses, the reliability of coarse root extrapolation under different transport times and step sizes remains insufficiently analyzed. Since local geometry is queried using the predicted coarse root trajectory, errors in this estimate could affect subsequent motion generation. More controlled comparisons using alternative local geometry query strategies or direct supervision of the global-stage root prediction would help clarify the specific advantages of the proposed design.

Limited evaluation of robustness and physical plausibility under challenging interactions. Although the qualitative results are generally plausible, the paper acknowledges occasional penetration and jitter during one-step generation. The evaluation could be strengthened with additional analysis of challenging scene configurations and the relationship between coarse trajectory errors and local interaction failures. Furthermore, the supplementary video mainly presents motion visualizations without directly demonstrating the claimed inference speedups. A side-by-side generation-time comparison would make the central efficiency contribution more transparent.

Questions for authors

Can the authors clarify the exact inference configurations used for the SceMoS comparisons and provide a standardized end-to-end latency comparison across methods? How much of the reported runtime includes scene encoding, geometry queries, and other preprocessing operations?

How reliable is the coarse root estimation across different transport times and step sizes? Can the authors compare the proposed root-guided geometry queries against alternative strategies, such as queries based on noisy root positions or explicitly supervised coarse root predictions?

How does the method perform in scenes with complex vertical geometry or challenging human-object contacts? Can the authors provide additional robustness evaluations and a supplementary demonstration that directly compares actual generation latency with existing approaches?

7 comments
AuthorsOpenReview ↗

Response to Reviewer

We sincerely thank the reviewer for recognizing the significance of our research problem, the substantial improvements in scene-conditioned motion generation efficiency, the effectiveness and reasonable design of our architecture, and the relatively comprehensive experimental evaluation. We also appreciate the constructive feedback regarding the fairness of evaluation and inference-time reporting, the need for more direct validation of our key mechanisms, particularly root-guided local geometry queries and additional root supervision, and further robustness analysis under different scene configurations. We address these concerns with additional experiments and clarifications below.

AuthorsOpenReview ↗

Q1. SceMoS Reproduction Configurations and Inference Latency Comparison

SceMoS reproduction configurations. As clarified in the final paragraph of Appendix B, we adopt the offline inference mode of the official SceMoS implementation for the comparisons reported in Table 1 and Figure 3 of our paper. This mode uses precomputed geometry conditions derived from ground-truth motion, as discussed in Appendix B. We explicitly distinguish this setting from the online inference mode, which queries local geometry based on previously generated motion. The offline setting is adopted to reproduce the performance reported in the original SceMoS paper, although its use of ground-truth-derived conditions means that the conditioning requirements differ from those described in the original SceMoS paper.

Inference latency comparison. Different scene-conditioned motion generation methods employ different native motion sequence lengths and inference procedures. Changing these configurations may introduce inconsistencies between training and inference, potentially degrading generation quality. Therefore, we preserve the official inference configurations of the baselines to maintain their established generation behavior.

For all compared methods, we measure online inference latency starting from preprocessed scene inputs. The reported times include all model-side computations, covering text encoding, global scene-feature encoding, and the local geometry queries and encoding performed at every sampling step. We exclude time-consuming scene preprocessing, such as scene point-cloud cropping in Afford@2 and BEV-map rendering in SceMoS. These operations can typically be completed during scene construction, with their outputs reused across subsequent motion-generation requests in the same scene. All measurements use batch size 1, with no text embeddings, global scene features, or local geometry features cached across generation runs.

To account for differences in native output sequence lengths, Table 1 additionally reports frame throughput in frames per second (FPS), defined as the number of valid motion frames divided by the corresponding inference time. This provides a sequence-length-aware complement to latency under each method’s native evaluation setting.

Table 1 Inference latency and frame throughput comparisons on TRUMANS and HUMANISE. FPS is defined as the number of valid motion frames divided by the corresponding inference time measured under each method's native inference configuration. Higher FPS indicates greater generation throughput.

DatasetMethodTime (s) ↓FPS ↑
TRUMANSTRUMANS6.04811.90
TeSMo31.062.58
SceMoS-( Our evaluation)2.96926.95
SceShortcut-1step (Ours)0.0581379.31
HUMANISEHUMANISE0.158390.44
Afford@23.34718.42
MoST3.20219.19
SceShortcut-1step (Ours)0.0591043.94

Runtime breakdown of our method. We further report the average inference latency breakdown of SceShortcut-1step on the TRUMANS test set in Table 2. The total inference time is 0.058s, as reported in Table 1 of our paper. This includes text encoding, global scene condition encoding (including global geometric feature encoding, DINOv3 semantic feature extraction, and feature fusion), local geometry queries and encoding, and the remaining motion generation operations. These measurements clarify the computational costs of the individual components and the overall inference efficiency of our method. As shown in Table 2, local geometry queries and encoding account for only 0.003s (approximately 5.2%) of the total 0.058s inference latency, demonstrating that fine-grained local geometry conditioning introduces only limited computational overhead.

Table 2 Breakdown of the average inference latency of SceShortcut-1step on TRUMANS.

ComponentTime (s)
Text encoding0.017
Global scene condition encoding (geometry encoding, DINOv3, and feature fusion)0.013
Local geometry queries and encoding0.003
Motion generation network and remaining operations0.025
Total inference latency0.058
AuthorsOpenReview ↗

Q2. Reliability of Coarse Root Estimation and Root-Guided Geometry Queries

Reliability of coarse root estimation. Following the reviewer's suggestion, we evaluate the reliability of coarse root extrapolation across different transport times tt and step sizes dd. Specifically, we report the coarse root-trajectory translation error (CTE) and rotation error (CRE) relative to the ground-truth root trajectory in Table 3.

Table 3 Reliability of coarse root estimation under different transport times tt and step sizes dd. CTE and CRE measure translation and rotation errors relative to the ground-truth root trajectory, in meters and degrees, respectively.

Transport Time ttStep Size ddCTE ↓CRE ↓
00.250.3424.5
00.50.3223.8
010.3123.4
0.250.250.3223.5
0.250.50.3123.1
0.50.250.2922.6
0.50.50.2822.3
0.750.250.2721.9

As shown in Table 3, coarse root estimation generally becomes more accurate as the transport time tt approaches 1 for a fixed step size dd. Similarly, for a fixed tt, the estimation accuracy improves as t+dt+d approaches 1. Notably, even under the challenging one-step generation setting (t=0,d=1)(t=0, d=1), the CTE and CRE are only approximately 15% and 7% higher, respectively, than those under the setting with the lowest errors (t=0.75,d=0.25)(t=0.75, d=0.25). These results support the reliability of our coarse root extrapolation across different transport times and step sizes.

Furthermore, as described in Section 3.3, we incorporate the embeddings of tt and dd into the encoded local geometry features, allowing the model to explicitly account for different transport times and step sizes during local geometry conditioning.

Further comparisons of root-guided geometry queries. We additionally compare our proposed coarse-root-guided geometry queries with alternative query strategies on the TRUMANS dataset, including querying local geometry using noisy root positions and ground-truth root trajectories. We also evaluate a variant that applies additional root supervision directly to the global-stage coarse root prediction, rather than to the combined root prediction. All variants are independently trained using the same shortcut training protocol, except for the explicitly modified geometry query strategies or root supervision targets. The results are reported in Table 4.

Table 4 Comparisons of different local geometry query strategies and root supervision configurations on the TRUMANS dataset.

ConfigurationFID ↓PFC ↓Cont. ↑P_mean ↓P_max ↓FTE ↓FRE ↓CTE ↓CRE ↓
Noisy root-guided geometry queries0.390.640.961.9312.030.3323.80.3625.9
GT root-guided geometry queries (Oracle)0.290.610.991.7411.390.3021.80.3223.4
Additional supervision on coarse root prediction0.340.610.991.7811.530.2921.50.3223.3
Coarse root-guided geometry queries (Ours)0.320.610.981.7711.480.2921.70.3123.4

As shown in Table 4, our proposed coarse-root-guided geometry queries achieve better motion quality and scene compatibility than queries based on noisy root positions, while approaching the performance obtained using ground-truth root trajectories. Furthermore, applying the additional root supervision directly to the coarse root prediction does not yield obvious improvements in motion quality or scene compatibility. These results support the effectiveness of our proposed root-guided local geometry query strategy and the current design of our additional root supervision.

Reviewer_U7Q9OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SceShortcut combines shortcut modeling with a within-step global-to-local architecture for scene-conditioned motion generation. A coarse root estimate guides local geometry queries, followed by body prediction and root refinement. It reports useful one-step quality and latency on TRUMANS and HUMANISE.

Critical strengths
  1. Efficient scene-conditioned motion generation is a meaningful problem, and the reported quality/latency operating point is useful.
  2. Table 3 provides relevant same-shortcut, comparable-parameter architecture and root-supervision controls.
Critical weaknesses
  1. The scene-specific contribution needs clearer attribution beyond existing efficient generation. Shortcut learning is established by Frans et al.. The remaining contribution is the particular integration of within-step root-guided queries, feature sharing, root correction, and supervision. Table 3 is positive evidence: the full model improves FID from 0.46 to 0.32 over the single-stage model at similar latency. However, the exact inputs, applicable root losses, and budgets retained by each variant are insufficiently explicit. The large speedup against iterative generators does not independently isolate this architectural contribution. I am open to raising my rating if the rebuttal establishes that contribution through matched controls
Questions for authors
  1. Table 3 uses the same shortcut protocol and comparable parameter counts. Which comparison most directly isolates within-step coupling? Please clarify whether conditioning inputs, applicable root losses, and training budgets are matched apart from the components explicitly ablated.
  2. Does comparing the separate root/body variant with the no-residual variant isolate intermediate feature passing, with other settings matched? Beyond the reported CTE/CRE against ground truth, what evidence supports better alignment between the queried local geometry and the final generated root trajectory?
  3. Does the reported latency include text/scene encoding and per-step local geometry queries and encoding? Please specify the inference batch size, cached features, and sequence lengths used for timing, and explain how the baseline speedups account for the output-length differences disclosed in Appendix C.
14 comments
AuthorsOpenReview ↗

Response to Reviewer U7Q9

We sincerely thank the reviewer for recognizing the importance of efficient scene-conditioned motion generation and the favorable trade-off between generation quality and inference latency achieved by our method. We also appreciate the constructive feedback regarding the attribution of our scene-specific contributions beyond established shortcut modeling. We agree that clarifying the conditioning inputs, applicable root-related losses, and training budgets of individual variants is essential for evaluating our architectural contribution. To address these concerns, we provide detailed experimental configurations and additional controlled ablations to better isolate the contributions of our proposed components.

AuthorsOpenReview ↗

Q1. Architectural Contributions and Matched Ablation Controls

We agree that shortcut modeling is an established generative framework. Our primary contribution beyond shortcut modeling is a scene-specific architecture that enables fine-grained scene adaptation within a limited sampling budget. Specifically, our within-step global-to-local architecture integrates coarse root estimation, root-guided local geometry queries, intermediate feature sharing, and root refinement into a single transport update. Accordingly, the speedups over iterative generators primarily demonstrate the efficiency of shortcut-based sampling, while our architectural ablations aim to isolate the benefits of the proposed scene-specific design. We first reorganize the original Table 3 into Table 1 and provide detailed configurations of all ablation variants in Table 2, including the enabled architectural components, applicable additional root-related losses, trainable parameter counts, and training epochs.

Among the original architectural ablations, the comparison between Separate root and body models and w/o local root residual correction most directly isolates the effect of intermediate feature sharing. Both variants employ local geometry queries and additional root supervision on the coarse root prediction, without root residual correction or residual regularization. Their primary architectural difference is whether intermediate motion features are passed from the global stage to the local stage. As shown in Table 1, enabling intermediate feature sharing improves FID, Cont., P_mean, and P_max. These improvements provide initial evidence for the effectiveness of within-step intermediate feature sharing.

However, our full architecture additionally incorporates root residual correction and root residual regularization. To further isolate the architectural contributions under the complete root-refinement setting, we introduce two additional ablations in group (a++) of Tables 1 and 2. Specifically, the w/o feature sharing variant retains local geometry queries, root residual correction, additional root supervision, and root residual regularization, while removing intermediate feature sharing between the global and local stages. Compared with our full model, this variant therefore provides a more direct evaluation of intermediate feature sharing when the remaining architectural components and training objectives are retained. As shown in Table 1, our method with within-step intermediate feature sharing achieves better FID, PFC, P_mean, and P_max.

Additionally, we introduce the w/o local geometry query variant, which removes local geometry conditioning while retaining intermediate feature sharing, root residual correction, and the corresponding root-related training objectives. This comparison allows us to isolate the contribution of local geometric information to scene-aware motion synthesis. Both the newly introduced and original ablation variants follow the same base shortcut training protocol and are trained for 300 epochs, with comparable numbers of trainable parameters. They employ the same text and global scene conditions, as well as the same optional initial-state conditioning. The results in Table 1 demonstrate the effectiveness of our proposed scene-specific architectural components and additional root-related training objectives, with our full model achieving better overall motion quality and scene compatibility while maintaining comparable inference efficiency relative to the corresponding ablation variants with individual components or training objectives removed.

AuthorsOpenReview ↗

Q1. Architectural Contributions and Matched Ablation Controls

Table 1 Ablation studies of (a) architecture design, (a++) additional architecture ablations, and (b) training loss for our method. The best and second-best results are bold and italics, respectively.

ConfigurationFID ↓PFC ↓Cont. ↑P_mean ↓P_max ↓FTE ↓FRE ↓CTE ↓CRE ↓Time (s) ↓
(a) Architecture Design
Single-stage shortcut model0.460.670.952.0112.720.3224.6//0.055
Separate root and body models0.380.610.971.8511.790.3020.8//0.061
w/o local root residual correction0.350.620.991.8311.510.3123.2//0.058
(a++) Additional Architecture Ablations
w/o feature sharing0.340.620.981.8011.630.3022.10.3223.70.059
w/o local geometry query0.410.640.961.9312.170.3123.10.3424.90.056
(b) Training Loss
w/o additional root supervision0.560.710.942.1713.350.3728.30.4030.70.059
w/o root residual regularization0.360.620.961.8711.740.3122.70.3526.30.058
Full model (Ours)0.320.610.981.7711.480.2921.70.3123.40.058

Table 2 Design configurations of the ablation variants in Table 1. We report the enabled architectural components, applicable additional root-related losses, numbers of trainable parameters, and training epochs.

ConfigurationLocal Geometry QueryFeature SharingRoot Residual CorrectionAdditional Root SupervisionRoot Residual RegularizationParamTraining Epoch
(a) Architecture Design
Single-stage shortcut model✗✗✗✓✗60.14M300
Separate root and body models✓✗✗✓✗66.93M300
w/o local root residual correction✓✓✗✓✗63.77M300
(a++) Additional Architecture Ablations
w/o feature sharing✓✗✓✓✓66.94M300
w/o local geometry query✗✓✓✓✓60.43M300
(b) Training Loss
w/o additional root supervision✓✓✓✗✓63.78M300
w/o root residual regularization✓✓✓✓✗63.78M300
Full model (Ours)✓✓✓✓✓63.78M300

Comments on the paper

3 comments
AuthorsOpenReview ↗

Summary of Rebuttal and Updated Manuscript — Looking Forward to the Reviewers' Feedback

Summary of Rebuttal and Updated Manuscript — Looking Forward to Your Feedback

We sincerely thank all reviewers and Area Chairs for their careful evaluations and constructive feedback on our work. We have provided detailed responses to the concerns raised by the three reviewers, conducted additional experiments, and incorporated the corresponding clarifications and results into our revised manuscript, which has now been uploaded.

To facilitate further discussion, we summarize the major revisions and additional evidence below:

  1. Architectural contributions and controlled ablations. We clarify the contribution of our within-step global-to-local architecture beyond established shortcut modeling. We provide detailed configurations of all ablation variants, including conditioning components, root-related losses, parameter counts, and training budgets. Additional controlled ablations further isolate the contributions of intermediate feature sharing, local geometry queries, and root-related supervision.

  2. Inference efficiency and evaluation fairness. We clarify the reproduction configurations of the baselines, particularly the inference settings of SceMoS. We provide detailed inference-time measurement protocols, runtime breakdowns, and frame-throughput comparisons (FPS), explicitly accounting for differences in native sequence lengths and inference procedures.

  3. Validation of key mechanisms. We analyze the reliability of coarse root estimation across different transport times and step sizes, introduce coarse-to-final trajectory and geometry consistency metrics, and evaluate alternative geometry query strategies and root supervision targets.

  4. Motion diversity and robustness. We provide additional evaluations of conditional motion diversity under identical text-scene conditions, robustness across scenes with varying vertical complexity, and the relationship between coarse root estimation errors and generation quality. We also analyze potential trajectory drift and foot sliding over different spatial transition lengths, together with scene penetration and motion jitter distributions under different sampling budgets.

  5. Supplementary generation-time demonstration. We have additionally provided a side-by-side demonstration to make the inference-time differences between our method and the baselines more intuitive.

These revisions are reflected in the updated main paper and appendices, providing more comprehensive evidence for our architectural design, evaluation protocol, and generation quality.

We would greatly appreciate it if the reviewers could revisit our detailed responses and the updated manuscript. We hope that the additional experiments and clarifications have adequately addressed your main concerns. If any questions or concerns remain, we would be grateful for your feedback and would be happy to provide further clarification. We also hope that these additional results and revisions will be taken into consideration when reassessing our submission.

Thank you again for your time and consideration. We look forward to your feedback and continued discussion.

AuthorsOpenReview ↗

Second Updated Manuscript — Additional Comparisons and Looking Forward to Your Feedback

Dear Reviewers and Area Chair,

We sincerely thank all reviewers and the Area Chair for their careful evaluations and constructive feedback on our work. We have provided detailed responses to the concerns raised by all three reviewers, conducted extensive additional experiments, and incorporated the corresponding clarifications and results into our revised manuscript. We have now uploaded the second updated version of our manuscript and supplementary material.

Compared with our first manuscript update, this second revision primarily addresses the additional concerns raised by Reviewer 7W6o during the recent discussion, particularly regarding (1) comparisons with more recent scene-aware motion generation methods and (2) the evaluation of our scene-specific contributions against alternative one-step architectural designs.

1. Additional Comparisons with Recent Scene-Aware Motion Generation Methods.

To address Reviewer 7W6o's concerns, we have provided detailed clarifications and additional experimental evidence. We systematically documented our investigation of 11 recent scene-aware motion generation methods published in 2025–2026, including the availability of their official implementations, training and preprocessing code, and pretrained inference checkpoints. We found that publicly available pretrained checkpoints and sufficiently complete reproduction pipelines were generally unavailable, making strictly controlled comparisons difficult. We also examined the baseline selection practices of two recent studies, DeSeG (2026) and MDNet (ECCV 2026), which similarly rely primarily on representative methods published in 2022–2024.

Despite these reproducibility challenges, we made every effort to provide additional comparisons. Specifically, we included exploratory quantitative comparisons with HSI-GPT2 (CVPR 2026), IntentMotion (AAAI 2026), MDNet (ECCV 2026), and DeSeG (arXiv 2026) using their paper-reported results. We additionally fine-tuned LINGO (SIGGRAPH Asia 2024) on TRUMANS and evaluated it under our established evaluation protocol. Furthermore, we conducted architecture-based inference-time measurements for DIP (AAAI 2026), SSOMotion (AAAI 2026), and IntentMotion (AAAI 2026) using their publicly available network implementations.

We explicitly disclose the limitations of these supplementary comparisons, including differences in conditioning inputs, training/test splits, evaluation protocols, and the use of randomly initialized network parameters for architecture-based timing. These additional results and methodological clarifications have been incorporated into the updated supplementary material.

2. Clarification of Our Contributions Beyond Shortcut Learning.

Regarding the concern that our method was not evaluated against alternative one-step designs, we further clarified that our original Table 3 already includes several alternative one-step architectural designs under the same shortcut learning framework, including a single-stage shortcut model, separate root and body models, and variants without local root residual correction. Moreover, in our rebuttal to Reviewer U7Q9, we provided additional matched architectural controls, including variants without intermediate feature sharing or local geometry queries, together with detailed training configurations and geometry-alignment analyses. These variants follow the same shortcut training protocol, with comparable parameter counts and matched training budgets. We respectfully believe that these controlled experiments provide direct evidence for the effectiveness of our proposed scene-specific within-step global-to-local architecture, beyond the established shortcut learning framework.

3. Looking Forward to Your Feedback.

We would greatly appreciate it if all reviewers could revisit our detailed responses and the second updated manuscript, including the newly added experimental results and supplementary analyses. With less than 21 hours remaining in the discussion period, we sincerely hope to receive your feedback while there is still an opportunity for further clarification. In particular, we would be especially grateful to Reviewers U7Q9 and vDfq, from whom we have not yet received follow-up feedback on our rebuttal, if you could kindly review our responses and share your assessments.

If any concerns remain unresolved or any additional clarification is needed, we would greatly appreciate it if you could let us know before the discussion period concludes, so that we can make every effort to address them. We sincerely thank all reviewers and the Area Chair again for your valuable time, constructive feedback, and continued consideration of our work. We hope that our original experimental evidence, extensive rebuttal, and two rounds of manuscript revisions will be fully considered in your final assessments.

AuthorsOpenReview ↗

Third Updated Manuscript — Additional Comparison with Consistency-Model Distillation

Dear Reviewers and Area Chair,

We sincerely thank all reviewers and the Area Chair for your continued engagement and constructive feedback throughout the rebuttal process. We have now uploaded the third updated version of our manuscript and supplementary material. Compared with the second revision, the main addition is a new experimental comparison with an independently constructed baseline based on consistency-model (CM) distillation, an alternative approach for accelerating diffusion-based generation. Specifically, we applied CM distillation to the pretrained TRUMANS model and evaluated the distilled model under 1-step, 2-step, and 4-step inference, comparing it with SceShortcut under the corresponding sampling budgets.

Our results show that, under one-step inference, SceShortcut achieves approximately 40% faster inference (0.058s vs. 0.081s), while also demonstrating better motion quality (FID: 0.32 vs. 0.48) and scene compatibility. Similar advantages are observed under two-step and four-step inference. Importantly, this new comparison goes beyond our previous architectural ablations within the shortcut learning framework by introducing a baseline based on a different generative acceleration paradigm.

We sincerely hope that this additional evidence will further strengthen the experimental validation of our method and address the remaining concerns regarding alternative efficient generation approaches. We would be very grateful if all reviewers could consider this latest update, together with our previous responses and additional experiments, when determining your final assessments.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.