acceptodds
Withdrawn from ICLR 2027

Closely Interacting 3D Clothed Human Generation with a Tightness Signal

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

We introduce interacting clothed human generation, generating a pair of 3D clothed humans in close contact such as hugging, holding hands, or dancing, from text and a target interaction pose. Existing work either generates interacting bodies without clothing or reconstructs clothed people in contact from a real image, while single clothed-human generators produce an isolated person with limited text control and fidelity. No prior method generates novel interacting clothed humans. We propose a training-free pipeline that generates each clothed person in 2D, so a text prompt controls diverse subjects and outfits at high fidelity. It then reaches the interaction pose in 3D for exact contact, and resolves the garment interpenetration. The tightness signal drives both the reposing and the penetration solution through one confidence-weighted variational solver. First, tightness-aware skinning uses the signal's direction to recover cloth-to-body correspondences and its magnitude as a per-vertex confidence in the variational solver for the skinning weights, enabling template-free reposing without the loose-garment artifacts of naive weight transfer. Second, penetration-aware combination solves for a smooth per-vertex displacement field that pushes interpenetrating garments apart without distorting them, with the tightness signal making the looser garment yield more. In the experiments on diverse datasets, our reposing lowers cloth distortion on every measure, reducing out-of-plane bending by 5.6× over naive skinning, and our combination removes 94% of the interpenetration volume without distorting the garments. As a complete pipeline, it generates diverse, text-driven pairs of interacting clothed humans whose individual people surpass state-of-the-art single-person generators in appearance and geometry.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D human generation, clothed human, human-human interaction, reposing, penetration

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

This paper proposes a training-free pipeline for generating two interacting 3D clothed humans from text prompts and a target pose. The method combines existing 2D generation and 3D reconstruction models with tightness-aware skinning and penetration-aware optimization to reduce clothing distortion and interpenetration. Experimental results show that the proposed method can generate diverse, text-driven pairs of interacting clothed humans.

Critical strengths

The paper addresses an interesting and relatively underexplored problem of generating 2 interacting 3D clothed humans.

The proposed tightness-aware skinning and penetration optimization are technically reasonable, with quantitative improvements over existing alternatives.

Critical weaknesses

Limited technical novelty. The proposed framework largely integrates existing models for image generation, 3D reconstruction, and tightness estimation. The main technical contributions are two geometry processing steps based on a confidence-weighted variational solver, whose smoothness formulation is largely inherited from prior work. While the integration is useful, the technical novelty appears relatively limited.

Insufficient evaluation of consistency. The paper mainly presents static results without video demonstrations. It is difficult to assess whether the generated clothed humans maintain consistent geometry and appearance across different poses and viewpoints.

Limited 3D generation quality. Several generated examples (eg. Fig.4) exhibit noisy surfaces and visible geometric artifacts. The overall 3D quality appears less convincing than suggested by the reported metrics.

Questions for authors

How are the target interaction poses obtained in practice? Can the framework be combined with generated interaction poses rather than relying on poses sampled from an existing dataset?

How sensitive is the proposed method to the quality of the initial 3D reconstruction? Do reconstruction artifacts affect the reposing and penetration correction?

Reviewer_44NBOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

Task Description

  • This paper addresses interacting clothed human generation: generating a pair of 3D clothed humans in close contact from two text prompts and a target interaction pose.

Research Motivation

  • Existing interaction generators produce only unclothed bodies, clothed multi-person reconstruction cannot synthesize novel subjects, and single clothed-human generators produce only an isolated person with limited text control and fidelity.

Main Contributions

  • A training-free pipeline that generates each person in 2D, lifts it to 3D, and assembles the pair via tightness-aware skinning for reposing and penetration-aware combination for resolving garment interpenetration, both driven by a per-vertex tightness signal through a shared confidence-weighted solver.
Critical strengths
  • The paper is well written and easy to follow. The overall pipeline and each component are clearly illustrated, which makes the method easy to understand.
  • The paper addresses a practically relevant problem. Populating 3D representation is relevant to downstream applications such as games and film production.
  • The proposed approach is conceptually simple. The pipeline is training-free and composed of off-the-shelf models.
Critical weaknesses

The scope of the proposed task is ill-defined, and its significance is not convincingly motivated. The technical contributions also do not seem to address the core challenges of the newly introduced task. Instead, they read as components introduced to handle auxiliary problems that arise from the chosen pipeline design.

  • Overstated task novelty. The paper is framed as introducing a new task, "interacting clothed human generation". In practice, the proposed system chains three existing capabilities into a single pipeline: (1) text-to-image generation, (2) animatable 3D human reconstruction from a single image, and (3) localization of the reconstructed 3D humans according to a given target pose. I therefore find the claim of a new task, and of "the first pipeline" for it, to be overstated.

  • The target scope of the proposed method is ambiguous and not well-aligned to the task. As I understand it, the task assumes a target pose is given. The goal is then to produce textured avatars whose identities match the reference images generated from the text prompts, and to place them according to that pose. However, Fig. 8 suggests that the main effect of the third contribution is to correct an implausible target pose. In the "before" column, the given target pose already exhibits interpenetration between the two people, independent of the clothing. The "ours" column then shows the solver pushing the penetrating parts out. In effect, this component acts as a collision detector and solver that repairs an implausible input condition into a plausible one. If the target pose were provided as a valid, penetration-free configuration, would this problem arise at all? The method thus seems to address an issue outside the scope of the task as the paper defines it, where the target pose is assumed to be given correctly.

  • The core of the claimed task is under-addressed. If the paper aims to propose a method tailored to the task it introduces, I would expect the technical focus to be on avatar texture synthesis. However, the paper provides almost no technical description of this component, which is delegated entirely to off-the-shelf models (FLUX, SyncHuman), and presents no technical novelty there.

The experimental evaluation does not establish reliable benchmark/comparison, so it is difficult to well support the author's arguments.

  • Unclear baseline selection and fairness of comparison. The authors state that, since no baseline exists for this task, they compare against methods designed for other tasks. However, the paper does not explain the criteria used to select these baselines, nor what was done to ensure a fair comparison across methods that differ in input conditions, training data, and output representations. While App. D.1 describes the scoring protocol, it does not justify why these baselines were selected or address the training-data mismatch.

  • Inconsistent advantages from the choice of reference dataset. Single-human generation is evaluated on SHHQ. GSM was trained on SHHQ and should therefore be favored by FID/KID computed against this dataset. Yet the proposed method substantially outperforms GSM (FID 38.0 vs. 53.2). It is unclear how this is possible, and the paper offers no analysis of it. Conversely, the remaining baselines (GETAvatar, E3Gen, LayerAvatar) were not trained on SHHQ and are disadvantaged on SHHQ-based metrics by the domain gap. As a result, the reference dataset benefits some baselines and penalizes others, so the reported numbers do not constitute a fair comparison.

  • Reposing is evaluated without ground truth. The reposing evaluation (Sec. 4.3, Tables 3–4) relies only on intrinsic distortion metrics, without comparing against ground-truth geometry. 4D-DRESS provides real posed scans of the same outfits. Using target poses from the same sequence would therefore have allowed a direct comparison with the ground-truth scans (e.g., Chamfer or point-to-surface distance). If a ground-truth comparison was not intended, the dataset choice for evaluation is not well justified.

Questions for authors

In addition to the concerns raised in the Weaknesses, I would like the authors to address the following questions:

  • Why is the proposed task needed as a standalone task? Please clarify why "interacting clothed human generation" should be treated as an independent task, rather than as a combination of existing tasks such as text-to-3D human generation and the reposing and placement of the generated humans.

  • Why must the two proposed methods be evaluated and validated within the newly introduced task? What specifically prevents them from being handled within existing tasks? In my view, tightness-aware skinning naturally falls within the scope of animatable 3D clothed avatar modeling, since it concerns human–cloth interaction. Penetration-aware combination likewise falls within the scope of differentiable cloth simulation or physics-aware clothed avatar modeling, since it concerns cloth–cloth interaction. Please explain what makes it necessary to validate and evaluate these two components under the proposed task, not existing tasks and its conventional benchmarks.

  • Do the authors provide video results of interacting clothed humans? The paper states that the generated pairs "can be animated and rendered into video" (Sec. 1), yet all results appear to be static. It is unclear what practical value static interacting pairs have without temporal consistency. Could the authors provide video results over interaction motion sequences (e.g., from Hi4D) and discuss the temporal stability of the reposing and the interpenetration solve across frames?

Reviewer_4awSOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents a training-free framework for generating closely interacting 3D clothed humans from text descriptions and target interaction poses. The proposed pipeline first generates individual clothed humans using pretrained 2D generation and 3D reconstruction models, and then introduces two geometry-processing components: tightness-aware skinning for reposing clothed humans and penetration-aware combination for resolving interpenetration between interacting garments. Both components leverage a cloth-to-body tightness signal. Experiments demonstrate improvements in reposing distortion, interpenetration reduction, and the visual quality of generated human pairs.

Critical strengths

1. A meaningful and relatively underexplored task. The paper addresses the generation of closely interacting 3D clothed humans from text descriptions and target interaction poses. While prior works have studied human interaction generation and clothed human reconstruction separately, jointly generating novel clothed subjects with controllable appearances, interaction poses, and garment-level contact remains relatively underexplored. This task formulation is meaningful and has practical relevance to 3D character creation and animation.

2. A practical framework with well-supported improvements in cloth reposing. The proposed framework integrates existing generative models with tightness-aware geometric processing without requiring additional training or garment templates. In particular, leveraging cloth-to-body correspondences for skinning weight assignment is well motivated, and the improvements over existing skinning approaches are consistently demonstrated across multiple distortion metrics and a large set of reposing experiments. The penetration-resolution module also substantially reduces garment intersections, providing a practical approach to assembling independently generated clothed humans.

Critical weaknesses

1. Insufficient evidence for the benefit of tightness-weighted penetration resolution.

The proposed penetration-resolution method uses cloth-to-body tightness to determine how much each garment should yield during contact. However, garment looseness does not necessarily reflect its physical compliance, and weighted deformation allocation has also been explored in prior garment untangling methods (e.g., Implicit Untangling, ACM TOG 2019). More importantly, the ablation in Appendix C.2 shows that a symmetric interface achieves comparable, and even slightly lower, residual penetration than the proposed tightness-weighted interface. Although the weighting makes looser garments move more, this does not establish that the resulting deformation is more plausible or better preserves garment fit. Therefore, the specific advantage of the proposed tightness-based weighting remains insufficiently demonstrated.

2. Insufficient quantitative evaluation of interaction and contact fidelity.

The paper aims to generate closely interacting clothed humans with accurate target poses and contact relationships. However, the evaluation primarily measures appearance, cloth distortion, and interpenetration volume, none of which directly establishes whether the intended contacts are preserved. Moreover, the body-alignment step in Appendix C.1 allows corrections to the target SMPL poses and relative placements. Although these adjustments are regularized, their actual impact on pose accuracy and contact preservation is not quantified. Since penetration-free configurations may still contain unrealistic contact gaps or displaced contact locations, additional evaluation of contact fidelity and pose preservation is important to substantiate the paper's central claims.

Questions for authors
  1. Benefit of tightness weighting. Appendix C.2 shows that the symmetric interface achieves comparable, even slightly lower, residual penetration than the proposed tightness-weighted interface. Beyond making looser garments move more, could the authors provide quantitative evidence that tightness weighting better preserves garment fit or produces more plausible contact geometry?

  2. Contact fidelity. The current evaluation focuses primarily on interpenetration reduction, which does not necessarily guarantee correct contact. Could the authors quantitatively evaluate whether the intended contact relationships are preserved after penetration resolution, for example using contact-region distances or contact preservation rates?

  3. Target pose preservation. Appendix C.1 allows adjustments to the SMPL poses and relative placements to resolve body penetration. How large are these corrections in practice? Could the authors report pose or joint-position errors before and after body alignment to demonstrate that the target interaction poses remain sufficiently accurate?

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

Related papers

Discussion (0)

Sign in to comment.