acceptodds
Withdrawn from ICLR 2027

PRISM: Articulated Object Modeling with Precise Structural Control

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Articulated objects are fundamental to robotics, simulation, and interactive virtual environments, where generated assets must satisfy not only visual realism but also structural and functional validity. Although recent generation methods have made substantial progress through 2D diffusion priors, sparse-view reconstruction, and native 3D generative models, their semantic or appearance-level conditions leave part topology and articulation underdetermined. Consequently, visually plausible assets can exhibit incorrect connectivity, joint placement, or functional organization. We present Prism, an Analytic-Concept-guided framework for articulated object modeling with strict structural control, together with an automated Analytic Concept construction module that instantiates the executable concepts used for conditioning. Prism uses Analytic Concept as a procedural control interface that specifies part geometry, topological relations, articulation constraints, and affordance functions in an executable form. Given a target brief, procedural templates, and a small set of valid instances, the construction module infers parameter dependencies, proposes an instance in dependency order, and repairs only the parameters implicated by failed execution checks. A validated concept and an appearance reference image are then converted into a multi-modal prior representation from concept-rendered images, concept-derived language descriptions, and image-based appearance features. Adaptive modulation and cross-attention integrate these complementary modalities before a conditional diffusion transformer maps the resulting representation into a latent space. The predicted latent is then decoded into a mesh that preserves the specified articulated organization while maintaining realistic geometric details. We evaluate Prism on PartNet-Mobility and downstream robotic manipulation tasks. Under their respective input conditions, comparisons with general-purpose and articulated-generation baselines show improved geometric quality and structural fidelity. Controlled condition ablations further show that gains depend on how Prism integrates the explicit concept with appearance evidence, rather than on structural specification alone. Evaluating automatically constructed concepts across unseen categories remains a separate question.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Controllable 3D generation, Structural Concept, Articulated Objects.

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_u6ECOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PRISM, a framework for articulated object generation that combines executable Analytic Concepts with multimodal conditioning and a diffusion-based 3D generator. The method integrates concept-rendered images, concept-derived textual descriptions, and appearance references to improve geometric quality and structural fidelity. The paper also introduces an automated concept construction module based on parameter dependencies, execution checks, and scoped repair. Experiments on PartNet-Mobility demonstrate improvements over several general-purpose and articulated-generation baselines, along with downstream robotic manipulation evaluations.

Critical strengths
  1. The paper addresses an important problem in articulated object generation. Unlike conventional 3D generation methods that primarily focus on geometric appearance, PRISM explicitly incorporates structural and articulation information. This is particularly valuable for robotics applications, where functional consistency is essential.

  2. The generation results demonstrate substantial improvements over existing baselines. PRISM achieves better FID, Chamfer Distance, and concept similarity scores than the evaluated general-purpose generators. The conditioning ablations and robotic manipulation experiments further provide useful evidence that combining structural and appearance information can improve the generated assets.

Critical weaknesses
  1. The automated concept construction module lacks quantitative experimental validation. Although dependency induction, execution auditing, and scoped repair are presented as key methodological contributions, Section 4.2 and Appendix A.2 provide only an evaluation protocol without numerical results. Consequently, the effectiveness, efficiency, and generalization of this module remain unverified. This is a significant limitation because automated concept construction is explicitly identified as one of the paper's contributions.

  2. The baseline comparisons do not sufficiently isolate the contribution of the proposed generation architecture. PRISM receives an explicit Analytic Concept containing part organization and articulation information, whereas the evaluated baselines primarily rely on image or semantic conditions. Although the paper acknowledges these different input interfaces, the resulting performance improvements cannot be attributed solely to the proposed multimodal fusion architecture. The existing condition ablations are useful, but they do not include an appearance-only configuration with the same PRISM backbone, making it difficult to determine how much of the improvement comes from the additional structural information versus the proposed conditioning mechanism.

Questions for authors
  1. The automated concept construction module is presented as a major contribution, but I could not find any quantitative results for it. How does it compare with a simpler LLM-based parameter generation approach?

  2. I wonder how much of the improvement comes from the proposed architecture rather than the additional structural information. What would the results look like if PRISM used only the appearance image with the same backbone?

  3. Since Uni3D-C mainly measures feature similarity, could the authors also report direct structural metrics, such as part-count accuracy and part-connectivity accuracy, to better evaluate whether the generated objects preserve the intended topology?

  4. I noticed that PRISM reports different FID and CD values in Tables 1 and 3. Are these results evaluated on different subsets, or is there another reason for the difference?

  5. In Table 4, combining concept images with text actually performs worse than using concept images alone. Do the authors have an explanation for why adding the text condition hurts performance in this setting?

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

PRISM introduces an Analytic-Concept-guided framework for structurally controllable 3D generation. Compared with the previous version, it adds automated concept construction, articulated-generation benchmarks, and downstream manipulation evaluations. However, several fundamental concerns remain unresolved.

Critical strengths
  1. Executable structural concepts provide explicit control over object geometry, topology, and affordances.
  2. Additional articulated-generation baselines, joint-level metrics, and manipulation experiments improve the evaluation.
Critical weaknesses
  1. The automated concept construction module lacks quantitative evaluation, ablations, and comparisons with simpler alternatives.
  2. Reliance on manually designed ConceptFactory templates limits generalization to unseen structures and novel categories.
  3. Joint parameters and trajectories are provided by Analytic Concepts, leaving the generated meshes' connectivity, articulation, and collision-free motion insufficiently validated.
  4. The dataset size and diversity seem limited, with no evaluation on complex multi-joint objects or unseen structural configurations.
Questions for authors
  1. Quantitative evaluation of the automated concept construction module.
  2. Evaluate the generalization of the framework.
Reviewer_bmCmOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents PRISM, an articulated object generation framework that uses executable Analytic Concepts for precise structural control. The method constructs Analytic Concepts through parameter dependency analysis, constraint checking, and scoped repair. It then fuses concept-rendered images, concept-derived text, and appearance reference images to condition a DiT, which generates 3D latents decoded into meshes by a pretrained Dora-VAE. Experiments on PartNet-Mobility demonstrate improvements in geometric quality and structural consistency over the reported baselines, along with promising results on several downstream robotic manipulation tasks.

Critical strengths
  1. The paper addresses the important problem of maintaining geometric and structural consistency in articulated object generation. Using executable Analytic Concepts to specify part topology, articulation constraints, and affordance functions provides an intuitive approach to fine-grained structural control.
  2. The method effectively combines pretrained 3D generative models with explicit structural priors, achieving competitive results in geometric quality and structural similarity. The paper also includes condition ablations and downstream robotic manipulation experiments, providing supporting evidence for the effectiveness of the proposed generation framework.
Critical weaknesses
  1. Analytic Concept builds on ConceptFactory, while 3D generation relies on Dora-VAE and DiT, with established AdaLN and cross-attention mechanisms. The generative framework primarily integrates existing components, without comparisons to simpler structure-conditioned generation approaches.
  2. Although Analytic Concept specifies structural and articulation constraints, the paper does not clearly explain how these constraints are enforced on the final mesh. Figure 3 includes a Consistency Loss, but its formulation is not provided in Section 3.5. Direct visual comparisons of joint axes associated with generated assets and ground truth are also missing.
  3. PRISM receives explicit part topology, joint, and affordance information, whereas the baselines mainly rely on image or semantic conditions. Although the authors acknowledge this difference, the results in Tables 1 and 3 do not clearly disentangle the benefits of additional structural priors from those of the proposed generation architecture.
  4. The paper states that PRISM's joint errors measure fidelity to the specified articulation. How are the joint parameters used in this evaluation obtained and associated with the generated meshes? Are they inherited from the input Analytic Concept or evaluated from the generated assets? This distinction is important for interpreting the reported joint-axis errors as evidence of kinematic consistency.
Questions for authors

Just see weeknesses

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

Related papers

Discussion (0)

Sign in to comment.