acceptodds
Withdrawn from ICLR 2027

ArtiReal: Upgrading Your Coarse Articulated Models into High-Fidelity Articulated Assets

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

High-quality articulated 3D assets are increasingly valuable for robotic interaction, physics-based simulation and digital-twin modeling. Recent articulated-object generators can produce structured models with executable articulation, but still face limitations in fine-grained geometry, realistic appearance, and overall visual coherence. Meanwhile, recent static 3D generation models can synthesize highly realistic assets, yet lack explicit motion properties and have not been fully exploited by existing articulated-object generation pipelines. We introduce ArtiReal, a generator-agnostic agentic framework that bridges this gap by upgrading visually coarse articulated assets into high-fidelity ones while preserving their original structure and articulation. To achieve this goal, we first develop an articulation-aware visual evidence planning module that actively selects informative observation configurations and constructs coherent and photorealistic object- and part-level references. We then introduce a structure-preserving adaptive 3D realization module that adaptively composes geometry and appearance operations to produce high-fidelity parts while preserving articulation constraints. Moreover, an action-conditioned provenance critic monitors the entire process, traces failures to their upstream causes, and enables targeted rollback and repair. We evaluate ArtiReal on articulated assets produced by multiple existing generators, with comprehensive comparisons covering visual realism, structural and articulation consistency, efficiency, and systematic ablations. The results demonstrate that ArtiReal consistently improves the visual quality of diverse articulated assets while maintaining their structural and functional validity. Please refer to the appendix and our anonymous supplementary page for more details and demonstrations.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Articulated 3D Assets, High-Fidelity 3D Generation, Agentic Framework

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

This paper introduces ArtiReal, an agentic framework that transforms structurally valid yet visually coarse articulated 3D assets into high-fidelity counterparts while preserving their original part structure, joint relationships, and kinematic attributes. ArtiReal bridges the gap between articulated asset generators, which produce valid articulations but often lack visual fidelity, and static 3D generators, which achieve realistic appearance but do not support articulated motion. It integrates three core modules into a closed-loop refinement process. The articulation-aware visual evidence planning module actively selects viewpoints and articulation states and synthesizes globally coherent photorealistic references at both the object and part levels. The structure-preserving adaptive 3D realization module adaptively selects and composes operations from a 3D generation toolbox, including voxel-constrained refinement, normal-guided refinement, and mesh simplification. Finally, the action-conditioned provenance critic evaluates intermediate results against action-specific objectives, traces defects to their upstream causes, and enables targeted rollback and correction.

Critical strengths

None

Critical weaknesses
  1. The work presents limited technical novelty, with its contribution residing primarily at the application level. Although the authors list “applications to robotics” as a keyword, robotic simulation typically demands accurate physical parameters—mass, inertia, friction, and contact properties—rather than visual fidelity and articulation alone. Recent methods such as PhysX-3D, PhysX-Anything, PhysX-Omni, PhysX-CoT, and PhysForge address this gap by generating simulation-ready assets rather than merely articulated ones. In contrast, the proposed method assumes a predefined articulation structure and confines its refinement to geometry and appearance. Consequently, its claimed value for robotics applications remains insufficiently substantiated.

  2. To demonstrate its application to robotics, the paper should compare against the more relevant sim-ready methods (PhysX-3D, PhysX-Anything, PhysX-Omni, PhysX-CoT, and PhysForge) mentioned above and validate its usefulness by training policies in simulation with sim-ready assets. Only such experiments can substantiate the benefits for robotics, but the paper does not conduct them.

Questions for authors

In my view, this paper primarily addresses an engineering problem, i.e., how to obtain higher-quality 3D assets through an agentic pipeline without any training. Its technical novelty is limited. If the contribution is to be judged purely as an engineering contribution rather than for its novelty, the authors need to justify whether the token cost incurred by this refinement pipeline is worthwhile. Given that all currently available trained 3D generative models are open-source and free to use, what advantages do the refined assets actually offer? Are they worth the tokens consumed?

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

The paper addresses a clear problem: existing articulated asset generators often produce valid articulation but limited visual quality, whereas static 3D asset generators produce realistic appearance without explicitly preserving articulation validity. The proposed method serves as a post-processing framework that refines the geometry and appearance of coarse articulated assets. The author compares the articulation asset postprocessing with SAM3D, Blender, and Part-I2-3D to evaluate the effectiveness of the proposed method.

Critical strengths

The paper presents a promising post-processing framework for adaptively refining the geometry and textures of articulated assets. It is designed to work with assets produced by different generation approaches, including feed-forward models and template-based methods. The quantitative and qualitative evaluations demonstrate improvements in visual quality while largely preserving the original structure and articulation.

Critical weaknesses

The video demo lacks comparisons with the baselines, making it difficult to assess their relative articulation quality and visual appearance. For the computer example, demonstrating button-press interactions would be helpful. All demonstrations are rendered from a single viewpoint, making it difficult to inspect articulated motion from novel views. In addition, it would be better to show both untextured meshes and textured results to clarify whether the improvements come from geometric refinement, texture refinement, or both.

Regarding the method, most adaptive decisions appear to involve selecting operations from the 3D Generation Toolbox. However, the paper does not specify an explicit scoring criterion for determining the refinement order. An explicit decision tree or policy defining the conditions for selecting each operation would improve clarity and reproducibility. A controlled comparison with fixed refinement strategies would further help validate the effectiveness of adaptive tool selection if we have.

Moreover, distinguishing geometric discrepancies from texture differences also appears challenging and it is unclear that how the proposed method handles this. For example, a dark line in a reference image could represent either a physical groove requiring geometric refinement or a printed pattern requiring only texture refinement. How does the agent reliably distinguish these cases and select the appropriate operation? The current ablations primarily evaluate observation planning, the global appearance reference, and the critic, but do not directly validate adaptive tool selection. It would be better to include ablations and comparisons in : a). comparisons with fixed tools and fixed operation sequences under the same computational budget, and b). evaluation using annotated geometric, texture, and mixed defects to assess whether the agent selects the appropriate refinement operations.

Questions for authors

Regarding the method, most adaptive decisions appear to involve selecting operations from the 3D Generation Toolbox. However, the paper does not specify an explicit scoring criterion for determining the refinement order. Does the method follow a fixed sequence of stages, with adaptation limited to tool selection and rollback?

In addition, given Astra’s release, could the authors include a baseline that directly uses Astra to refine individual part meshes or have some additional discussion?

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

The paper presents an agentic pipeline for improving the geometry and appearance of existing articulated 3D assets while retaining their original kinematic structure. An articulation-aware visual evidence planner selects camera viewpoints and articulation states that expose relevant parts, then generates a global appearance reference and corresponding part-level references.

Guided by these references, a structure-preserving adaptive 3D realization module selects and combines operations according to each part’s geometric, visual, and articulation requirements. These operations include part regeneration, local surface refinement, texture generation, and mesh simplification. Throughout the pipeline, an action-conditioned provenance critic evaluates intermediate outputs. When defects are detected, it identifies their likely source and guides the agent to roll back to the appropriate stage.

Evaluation on 400 assets from four articulated-object generators demonstrates improved visual quality over competing refinement pipelines, with generally comparable structural and articulation consistency.

Critical strengths
  1. Refining existing articulated assets through coordinated image and 3D generation is a interesting and timely direction. Planning observations over both viewpoint and articulation state is well motivated, since important surfaces may remain hidden in the initial configuration.

  2. The visual improvements are supported by comparisons across four source generators, with ArtiReal achieving the highest human realism scores for each source. The ablations further support global appearance guidance and iterative correction. Removing this component substantially reduces the visual fidelity and realism of the generated assets.

Critical weaknesses
  1. Several central agent decisions are described at a high level without sufficient implementation detail. Although Appendix A explains individual toolbox operations, the policy governing their selection and correction remains unclear. In particular, the paper does not sufficiently specify how articulation sensitivity is assessed, how failures are attributed to earlier stages, or how retrospective adjudication resolves conflicting assessments.

  2. The evaluation does not sufficiently isolate the benefits of adaptive tool selection and provenance-aware correction. Table 3 lacks a comparison against a fixed policy using the same toolbox. Although the full critic improves the reported visual scores over local correction, the FID improvement from 72.109 to 71.388 and realism improvement from 4.37 to 4.45 are relatively small.

Questions for authors
  1. How are the part visibility, object completeness, and joint observability criteria in Section 3.2 computed? What thresholds determine whether an observation is accepted or additional viewpoints and articulation states are required?

  2. In lines 214–215, what does it mean for a part’s appearance to be “not adequately revealed” when its geometry is visible? Please clarify how the planner identifies this condition and distinguishes it from geometric occlusion.

  3. How are part geometry and kinematic relationships represented and supplied to the agent? What determines whether a part is articulation-sensitive and how much geometric change is acceptable? Please clarify whether tool selection follows explicit rules or is inferred through prompting.

  4. What information from the generation history is available to the critic, and how does it distinguish an incorrect visual reference from an unsuitable reconstruction operation? How does retrospective adjudication determine which assessment to revise? An example execution trace showing rejection, attribution, and rollback would clarify this process.

  5. Could the authors provide precise definitions and evaluation protocols for Motion-CD, AID, and AOR, including normalization and articulation-state sampling? Are geometric compatibility and overlap checked throughout the motion range?

  6. Why was the evaluation set selected from four representative generators rather than PartNet-Mobility or ArtiCraft-10K? Please clarify how the 100 assets per generator were selected, including category coverage and any filtering criteria.

  7. In Figure 3, apart from the second row, the results appear geometrically similar to the input, with differences mainly in texture and materials. Could the authors provide examples that more clearly demonstrate the geometric detail introduced by ArtiReal relative to both the input and the baselines?

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

Related papers

Discussion (0)

Sign in to comment.