acceptodds
Withdrawn from ICLR 2027

GLoRe: Detail-Preserving Text-Driven Two-Person Motion In-Betweening via Global Guidance and Local Refinement

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Motion in-betweening is fundamental to controllable character animation for virtual humans, games, and embodied interaction. While recent two-person motion in-betweening approaches have achieved good performance for generating coordinated interaction between two persons, they often result in degraded individual motion details. We introduce GLoRe, a novel three-stage detail-preserving framework containing an individual motion initialization stage that leverages a singleperson motion in-betweening model to generate detailed individual motions, a global guidance stage that provides global context for motion refinement, and a controlled local refinement stage that adaptively applies residual corrections. Specifically, in the global guidance stage, we propose a new global interaction condition that models sequence-level interaction demands from the initialized motions and input constraints; in the controlled local refinement stage, we introduce a new local residual gate that regulates the person- and frame-level corrections based on cross-person evidence. Extensive experiments on InterHuman and Inter-X demonstrate that our method achieves superior performance on text-driven twoperson motion in-betweening, effectively balancing individual motion quality and interaction coordination.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Human Motion Generation, Two-Person Motion In-Betweening, Human–Human Interaction, Controllable Generation, Global-to-Local Refinement

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_ZtS8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a three-stage framework for text-driven two-person motion in-betweening. First, a retrained CondMDI model is employed to generate initial motions for each of the two individuals. Next, a Global Interaction Condition integrates the initial inter-person motion relationships, textual descriptions, and keyframe constraints to provide sequence-level interaction guidance. Finally, a Local Residual Gate is used to modulate the intensity of residual corrections based on cross-person evidence, evaluated on a per-person and per-frame basis. The input keyframes are precisely recovered in the final output.

Critical strengths

The proposed GIC provides sequence-level target information, while the LRG regulates the magnitude of adjustments at both the character and frame levels. Cross-character evidence is used solely for estimating the gating mechanism, rather than directly overwriting current motion features, serving as an intuitive strategy for preserving fine-grained details. The paper evaluates the model across two datasets based on motion error, temporal dynamics, distribution quality, text alignment, generation diversity, and interaction coordination. Furthermore, it presents comprehensive analyses, including component and feature ablation studies, an examination of the gating scope, experiments with varying numbers of keyframes, and a user study.

Critical weaknesses
  1. There is a lack of direct evaluation regarding the core claim of detail preservation. Current metrics, L2P, L2Q, NPSS, and FID, fail to explicitly measure whether fine-grained, single-person motion details are retained. The paper should report metrics such as joint-wise displacement relative to the initial motion, local pose error, or kinematic derivatives (velocity, acceleration, or jerk). Otherwise, the core contribution rests primarily on indirect metrics and a limited set of visualizations.

  2. The comparison with the most relevant prior work on two-person motion in-betweening is insufficient. Cross-Space In-Betweening was excluded from quantitative comparisons due to issues with its code and evaluation protocol, while other key baselines were adapted by the authors from text-to-motion generation methods. Given that the quality of these adaptations and the selection of hyperparameters can significantly influence results, the current experiments do not adequately demonstrate an advantage over methods specifically designed for two-person in-betweening.

  3. The methodological innovation is largely incremental. Techniques such as single-person diffusion-based initialization, explicit pairwise features, cross-attention, residual refinement, and sigmoid gating are already established in the field. The paper’s primary contribution appears to be an effective combination of these existing techniques tailored for a specific task.

  4. The expressive capacity of the LRG module is limited; each refinement block applies the same global vector to all characters and frames, with its magnitude adjusted only by a scalar gate. Consequently, different frames are essentially restricted to scaling the same refinement direction, and cross-character information influences only the intensity, not the content, of the refinement. It is recommended to compare this approach against alternatives such as frame-wise vector gating, vector-valued gating, or schemes where local residuals are predicted directly from cross-character information.

  5. There is potential ambiguity regarding character semantics. Since two independent CondMDI initializers receive the same description of a two-person interaction, which often involves asymmetric roles. The paper does not sufficiently explain how these roles are assigned to the respective characters. This is particularly prone to causing role confusion when keyframes are highly sparse.

Questions for authors

See weakness.

Reviewer_yU9MOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies motion in-betweening jointly constrained by text descriptions and sparse two-person keyframes, aiming to improve interaction coordination while preserving individual motion details. GLoRe first initializes each person's motion separately using a retrained single-person CondMDI model. It then constructs global interaction conditioning from the current two-person motions, text, and keyframes, and uses cross-person information to estimate gating strengths for each person at each frame, modulating global guidance and predicting motion residuals. After the residuals are added to the initial motions, the input keyframes are preserved exactly. The paper evaluates in-betweening quality, interaction coordination, and text alignment on InterHuman and a filtered version of Inter-X, and provides module ablations, experiments with different numbers of keyframes, and a user preference study.

Critical strengths
  1. In the five-keyframe setting, it outperforms the listed baselines in FID and NPSS on both datasets and achieves the best mean value on the relative-orientation coordination metric.
  2. The method combines the current two-person motions with text and keyframe requirements to construct global guidance that depends on the initialization state, then modulates its influence through gating for each person at each frame.
  3. The paper examines different information sources in the global conditioning and compares fixed, sequence-level, and framewise gating, as well as variants with and without cross-person information.
Critical weaknesses
  1. Performing individual in-betweening before incorporating interactions, and preserving motion quality through refinement, are utilized in previous works.
  2. The supplementary materials provide some source code but do not include generated animations that can be viewed directly.
  3. Although the manuscript describes the results as competitive, the opening of the results analysis still makes a broad claim about advantages in text-conditioned generation quality. I recommend explaining these values in relation to the suitability of the pretrained evaluator for Inter-X, the retrieval candidate set, and the specific results, and distinguishing more clearly between advantages in in-betweening quality and trade-offs in text alignment.
Questions for authors
  1. Compared with existing approaches that combine individual in-betweening with interaction modulation, such as Cross-Space In-Betweening, what are GLoRe's key differences in information flow and refinement mechanisms? Which experiments most directly demonstrate the effects of these differences?
  2. How do the authors explain the low text retrieval accuracy for both ground-truth and generated motions on Inter-X? What is the size of the candidate set, and how was the evaluator adapted, if at all? How should these results be used to assess improvements in text alignment?
  3. I recommend centering the contributions on the design of global interaction conditioning and local gating, and providing a clear comparison with the closest prior work.
Reviewer_978oOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses text-guided two-person motion in-betweening, where improving interpersonal coordination can disrupt individual motion details, by independently initializing each person’s motion and then refining the interaction through global guidance and local gating.

Critical strengths
  1. The approach reuses established single-person motion priors and treats interaction coordination as a subsequent correction task, allowing the refinement stage to start from individually plausible motions.
  2. The method preserves prescribed keyframes while completing the interaction, making it useful for animation editing that requires control over specific poses.
Critical weaknesses
  1. The paper attributes the limitations of existing two-person methods to lost individual details, but its comparisons do not clearly distinguish detail loss from interaction errors or general reconstruction errors, leaving the central motivation insufficiently demonstrated. (Sections 1 and 4.2.)

  2. The local gate modulates an internal guidance signal rather than the final motion correction, so its design and average quality gains do not establish that reliable regions are preserved while poorly coordinated regions are selectively corrected. (Section 3.5.)

  3. The improvement over a plain refiner establishes the effectiveness of the full configuration rather than a distinct detail-preservation mechanism, particularly because the relationship between that baseline and the variant with local gating disabled is unclear. (Section 4.3.)

Questions for authors

No additional questions.

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

Related papers

Discussion (0)

Sign in to comment.