acceptodds
Under review as a conference paper at ICLR 2027

AeroGen: Parametric 3D Aircraft Generation from Structured Design Specifications

Abstract

Recent advances in generative models have created new opportunities for design automation in graphics, manufacturing, robotics, and games. However, conceptual designs of aircraft require rigorous physical plausibility, functional constraints, and controllable edibility. Such precision remains a bottleneck for common text- or image-conditioned generative models. Motivated by a standard aircraft design workflow, we therefore adopt structured parameters as a complete and interpretable conditioning interface to describe component hierarchies, topological structures, and continuous geometric parameters. We propose a parameter-conditioned aircraft generation framework, where the Parameter Extractor encodes variable-length hierarchical design parameters into conditioning tokens for Diffusion Transformer, while an attribute predictor improves the model's sensitivity to aircraft-level parameters by recovering them from generated geometry. Experiments on our built benchmark with 40,348 samples across 71 classes show that our method outperforms image-conditioned baselines in geometric fidelity and shows promising schema-consistent and parameter-sensitive editing.

Primary area:
generative models
Keywords:
Aircraft Generation, Structured Design

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

AeroGen conditions a pretrained 3D diffusion backbone on structured aircraft specifications using type-specific parameter tokens and structural/spatial embeddings. An auxiliary attribute predictor encourages parameter-sensitive generation. The paper also introduces AeroJSON-40K, built from OpenVSP parameter–mesh pairs, and demonstrates reconstruction and single-parameter editing.

Critical strengths
  • Explicit component parameters offer an interpretable control interface for aircraft geometry, with a sensible encoding of heterogeneous fields and structural relationships.
  • The paired dataset and auxiliary attribute supervision provide useful resources for studying parameter-conditioned shape generation and controlled edits.
Critical weaknesses
  • The learned generator's added value over direct construction is unclear. Appendix A creates the target OBJ from the same JSON through OpenVSP. A learned surrogate could offer speed, flexibility, or completion of unspecified details, but the paper does not establish such a benefit against that route.
  • The comparisons only partly isolate the contribution. AeroGen receives explicit component geometry and aircraft-domain training, while Appendix C evaluates image baselines with default pipelines. The gains therefore combine richer inputs, domain adaptation, and the proposed encoder. The raw-parameter ablation supports the added features, but does not resolve those broader alternatives.
  • Precise editing is less established than parameter sensitivity. Figure 4 provides useful qualitative edits, but whole-shape metrics and a latent predictor's recovery on one example do not quantify requested-versus-realized changes on decoded meshes or preservation of unedited attributes. This limits the strength of the editing claim.
  • Engineering framing exceeds demonstrated geometry control. Figure 1 claims physical plausibility and functional geometry, whereas the method supervises shape and selected attributes. Appendix D explicitly acknowledges that aerodynamics, strength, manufacturability, and stability are untested.
Questions for authors
  • What capability or efficiency benefit does learned generation provide over the OpenVSP construction used to create targets, and what geometric details are unspecified by the JSON?
  • What evidence distinguishes the encoder's benefit from aircraft-domain adaptation, and how accurately do decoded meshes follow edits while preserving unedited attributes?
  • Is the DiT frozen as stated in Sec. 4.3, or jointly optimized as stated in Appendix C?
Reviewer_9s5SOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

AeroGen proposes a framework for controllable 3D aircraft generation given structured design parameters. It introduces a Parameter Extractor to condition a diffusion transformer and an auxiliary Attribute Predictor to improve parameter consistency. Experiments on a dataset of 40,348 aircraft across 74 classes show improved geometric fidelity over image-conditioned baselines and promising single-parameter editing capabilities.

Critical strengths
  1. The paper addresses a problem with clear practical value, and its framework is well suited to aircraft design, where users need to control specific geometric parameters.
Critical weaknesses
  1. The motivation for using diffusion is not fully clear. Appendix A shows that the input parameters can already be used to construct the target geometry with OpenVSP. What benefit does the proposed model offer over this direct pipeline?
  2. Although the paper mainly focuses on aircraft, additional experiments on other engineering assets with similar parametric structures would help assess the method’s effectiveness and applicability beyond aircraft.
Questions for authors
  1. Refer to Weakness.
  2. How are the generated latents used by the Attribute Predictor obtained during training? The predicted value is velocity in Flow Matching, so how this predicted value is converted so as to input to the Attribute Predictor.
Reviewer_TTS9OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes AeroGen, a method that conditions TRELLIS.2 on structured aircraft parameters to control the positions, dimensions, and shapes of the fuselage and wings. A Parameter Extractor converts these specifications into conditioning tokens using type-specific projections and structural and spatial embeddings. An auxiliary Attribute Predictor recovers aircraft parameters from generated 3D latents, providing additional supervision to improve parameter-level controllability. To train and evaluate the method, the authors introduce AeroJSON-40K, a dataset containing 40,348 paired aircraft specifications and 3D models. Quantitative and qualitative comparisons with image-conditioned baselines show improved agreement between generated and reference geometry, measured by Chamfer Distance, Normal Consistency, and F1 score.

Critical strengths
  • Interesting research direction. As methods for generating 3D models from visual and textual inputs improve, the need for new ways to control their outputs grows. Although this method demonstrates control specifically for aircraft, I believe a similar approach could be applied to other types of parametric models and inspire further research. The release of AeroJSON-40K could also provide a valuable resource for the community.
  • Convincing results. The quantitative and qualitative evaluations against image-conditioned baselines show the benefit of the proposed conditioning. Specifically, the method substantially improves the agreement between the generated and reference geometry, as measured by Chamfer Distance.
Critical weaknesses

Writing and clarity. The writing and clarity of the paper could generally be improved. In the following lines I include some examples. The main paper does not explain how conditioning is injected into TRELLIS.2, which is an important detail. The training setup is inconsistent: Sec. 4.3 states that the DiT is frozen, whereas Appendix C describes joint optimization with the Parameter Extractor. The last two rows of Tables 2 and 3 report different versions of the proposed method alongside the baselines. For clarity, it would be better to move these comparisons to Section 5.2, which covers ablation studies.

Questions for authors
  • Could you provide more information in the main paper about how conditioning is injected into TRELLIS.2? This is an important detail that is currently explained only briefly in Appendix B of the supplementary material.
  • Could you clarify whether the TRELLIS.2 DiT is frozen or fine-tuned? Lines 361–362 state that it is frozen, whereas line 654 suggests that it is jointly optimized with the Parameter Extractor.
  • Minor details. I found the following typos in the manuscript:
    • L42: edibility → editability
    • L470 and Fig.1 (top, green box): Paramter → Parameter
    • L197: sturctured → structured

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.