acceptodds
Withdrawn from ICLR 2027

Int3Rcept: Physics-Driven Structural Redesign via 3D Generative Intermediate Representations

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Modern 3D generative models synthesize visually compelling assets, but physics-driven redesign must support substantial structural changes without sacrificing visual quality. Existing approaches struggle to achieve both: learned-feature optimization preserves generative priors but provides only indirect control over structural changes, while explicit grid-based optimization offers direct spatial control but operates outside the native synthesis space of pretrained generators. We introduce Int3Rcept, a physics-driven 3D redesign framework that intercepts and redesigns a pretrained generator's intermediate spatial representation under physical objectives. Our key idea is to use a spatially explicit representation as a direct design space while remaining compatible with downstream synthesis. Int3Rcept expands the intermediate structure into a volumetric design domain, optimizes a continuous density field under volume constraints, and converts it back through connectivity-aware discretization, followed by downstream synthesis. Int3Rcept achieves stronger physical improvement than learned-feature optimization and higher visual quality than explicit grid-based optimization. It also shows strong controllability across target material volumes, enabling support pruning, reinforcement, and opening creation. We further demonstrate the interface across multiple physical objectives and pretrained 3D generators. Together, these results establish intermediate spatial representations as a practical bridge between physical optimization and 3D generative priors.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D generative model, physics-aware optimization

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

The paper introduces Int3Rcept, a framework for physics-driven 3D redesign through spatial intermediate representations of pretrained 3D generators. It combines density-based optimization, connectivity-aware discretization, and frozen downstream synthesis to improve physical performance while retaining visual quality. Experiments cover compliance, equilibrium draft, and thermal resistance, with additional evaluations across generators.

Critical strengths
  1. The method has practical value in improving the physical performance of generated 3D assets while preserving their visual quality. Its support for different physical objectives and material budgets makes it potentially useful for physics-aware 3D design.
Critical weaknesses
  1. The pipeline requires manual inspection and refinement of semantic regions. The amount of intervention is not clearly quantified, making it difficult to assess the effort required to apply the method to new objects.
Questions for authors

Ref to Weakness

Reviewer_zKE4OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper offers a training-free method to incorporate physics-driven redesign into a 3D generative model pipeline. The topology optimization is enforced on the intermediate sparse structure of a pretrained two-stage generator (TRELLIS), whereas other stages of the pipeline are preserved. The proposed method addresses three different objectives: compliance minimization, equilibrium-draft minimization and thermal-resistance maximization. It then compares to the baseline PhysiOpt and explicit SIMP, and demonstrates the gain in performance.

Critical strengths
  • Their proposed method effectively enforces physically plausible structure of the designs, while preserving the ability of 3D generative prior.
  • It is practically efficient with end to end runtime around 85 s, much lower than the explicit grid SIMP-64.
  • It enables topology changes that cannot be achieved by latent method such as PhysiOpt. Because the explicit spatial domain is exposed, the method can add or remove parts.
Critical weaknesses
  • The redesigned structure will likely deviate dramatically from the original shape. I don't see any constraints in formulation that can counteract this. There is also no analysis on how close the output shape is to the original shape.
  • I am concerned that adding interior voxels to the intermediate surface mesh produced by TRELLIS would make the structure out-of-domain for the stage 2.
Questions for authors
  • How did the authors decide on which voxels are editable through SIMP?
  • It is not clear how high-density neighbor is added to the intermediate object surface.
  • The external load enforced is not shown in the visualizations (Fig. 3 (a)-1, (a)-2), so the optimization constraints are unknown here.
  • In connectivity discretization, how is KK determined?
  • For stage 2 synthesis stage, is the generative model conditioned on the original image or re-rendered image from the optimized structure?
Reviewer_ng6gOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

In this paper, the authors propose a framework for 3D geometry redesign which uses the intermediate spatial representation of 3D generator and performs physics-driven redesign. The authors show multiple use cases where the physics-driven redesign is very useful and the results also demonstrate the same.

Critical strengths
  • The idea of using an interpretable, spatially controllable intermediate spatial representation is very nice.
  • The authors show its use in multiple applications such as structural topology optimization, thermal redesign etc. all of which shows its ubiquitous applicability.
  • comparisons with learned features and explicit-grid alternatives show that INT3RCEPT performs much better.
Critical weaknesses
  • The key weakness I see is that not much discussion is there on the size of the intermediate spatial representation... if it is too coarse, the optimization will be equally difficult.
  • Ablation studies on the intermediate spatial representation is missing
  • The major weightage of the paper is on the physics-driven redesign, but I dont understand the need for the G_pre and G_post. G_post still makes sense based on the ablation studies, but why do we need G_pre is not clear to me. May be I didnt read this properly, but I feel that all you need is an explicit representation to do the redesign which you can get even without G_pre.
  • Comparisons with other NN based topology optimization approaches (may not be exactly possible, but some similar baselines would make sense). It is not clear to me as to what component of the entire pipeline is really benefitting the most for this approach of redesign.
Questions for authors

See above points in the weaknesses. A thorough analysis on the intermediate spatial representation and its effect on the redesign would be nice to look at and ablation studies as well. Without that, it seems like the authors tried out a bunch of stuff and something worked and they dont even know why it worked. Seeing a better presentation of results in this aspect would help strengthen the paper.

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

Related papers

Discussion (0)

Sign in to comment.