- Rating
- 3
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper presents surface-sync, a method for generating human–object interaction motions from text and object geometry. The main idea is to explicitly model where and how the human and object surfaces make contact. Built on a dual-branch diffusion model, it jointly predicts surface-marker motions, contact activation, and contact locations on the object. It adds objectives for surface alignment, contact coherence, and force-closure-inspired stability, and other constraints. The authors evaluate the method on InterAct and test cross-dataset transfer. The reported results show improvements over the compared baselines in motion quality, text alignment, and several contact-related metrics. Ablation studies and qualitative examples further support the contribution of the main components.
Critical strengths
-
The model predicts both whether contact occurs and where it occurs on the object surface, together with human and object motion. This gives the geometric and stability objectives a shared representation to work with. I see this as the paper’s main strength: it directly addresses interaction quality, which plausible-looking motion trajectories alone do not ensure.
-
On InterAct, the method improves FID, text alignment, contact distance, and realism over the compared baselines. Cross-dataset results also show consistent gains over HOI-Diff and InterDiff, while the ablations support the usefulness of the representation, physical objectives, and body coordination. Together, these results make the proposed design more convincing than qualitative examples alone would.
Critical weaknesses
-
The evidence is limited to relatively simple interaction settings. The formulation focuses on one person interacting with one object. It is unclear how it would handle multiple objects or people, where several human–object and object–object contacts must be coordinated. The video examples also appear to use short, single-action instructions, leaving performance on longer, multi-step instructions untested. These are common scenarios, and the paper does not explain or demonstrate how its contact representation and coordination objectives would extend to them. This limits how broadly I can interpret the claimed generality of the framework.
-
The visual results still show noticeable contact errors and low-quality motion. In the supplementary videos, I observed hand–object misalignment, sliding, and contact at incorrect locations. These issues directly concern the paper’s central claim of improving surface contact and interaction stability. Although the quantitative results are encouraging, the remaining artifacts make it unclear how reliably the proposed methods can be used in real applications.
Questions for authors
- Which parts of the current formulation would carry over to scenes with multiple people and objects, and which would require substantial changes? In particular, how would it represent and coordinate simultaneous human–object and object–object contacts? This matters to my assessment of its practical applicability beyond the single-person, single-object setting.
- In my own work, I have found that a VLM-guided refinement loop. Rendering a 3D scene in Blender, asking a model such as Qwen (or even Astra) to identify semantic and physical inconsistencies, and iteratively adjusting the motion can substantially improve physical plausibility. The results are even better than yours shown in the supplementary video. What advantages does your specialized formulation offer over such an approach, for example in contact accuracy, temporal consistency, computational cost, or reliability? A controlled comparison or concrete evidence would help justify the added complexity of the method.
- Can the current model follow a multi-step instruction, such as “pick up the object, turn around while holding it, and place it on a nearby surface,” starting from a specified initial human–object configuration? Please clarify whether initial-state conditioning is supported and, if so, provide a representative result showing contact transitions and continuity between actions. This would help establish whether the method extends beyond the short, single-action examples shown.