PhysGen: Physically Grounded 3D Shape Generation for Industrial Design

Yingxuan You, Chen Zhao, Hantao Zhang, Mingda Xu, Pascal Fua · 2025

Read PhysGen: Physically Grounded 3D Shape Generation for Industrial Design by Yingxuan You, Chen Zhao, Hantao Zhang, Mingda Xu, Pascal Fua (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2512-00422.txt.

Introduction

Generative models can design three-dimensional shapes that look plausible yet fail as real products. Yingxuan You and co-authors present PhysGen in 2025, a framework for physically grounded shape generation aimed at industrial design. For engineered shapes, realism is tied to physical properties like aerodynamic efficiency.

A car body that ignores airflow may render beautifully and still be a bad design. PhysGen couples shape generation to the underlying physics so outputs respect engineering constraints. Designers could explore more concepts automatically while keeping each candidate grounded in how the object must actually perform. Industrial design is the test case, where a pretty shape that fails physics is worthless. Automobiles are the running example, since drag decides their efficiency.

Catalog record

CREATORYingxuan You, Chen Zhao, Hantao Zhang, Mingda Xu, Pascal Fua
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE51 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2512-00422

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

PhysGen: Physically Grounded 3D Shape Generation for Industrial Design Yingxuan You, Chen Zhao, Hantao Zhang, Mingda Xu, Pascal Fua Abstract Existing generative models for 3D shapes can synthesize high-fidelity and visually plausible shapes. For certain classes of shapes that have undergone an engineering design process, the realism of the shape is tightly coupled with the underlying physical properties, e.g., aerodynamic efficiency for automobiles. Since existing methods lack knowledge of such physics, they are unable to use this knowledge to enhance the realism of shape generation. Motivated by this, we propose a unified physics-based 3D shape generation pipeline, with a focus on industrial design applications. Specifically, we introduce a new flow matching model with explicit physical guidance, consisting of an alternating update process. We iteratively perform a velocity-based update and a physics-based refinement, progressively adjusting the latent code to align with the desired 3D shapes and physical properties. We further strengthen physical validity by incorporating a physics-aware regularization term into the velocity-based update step. To support such physics-guided updates, we build a shape-and-physics variational autoencoder (SP-VAE) that jointly encodes shape and physics information into a unified latent space. The experiments on three benchmarks show that this synergistic formulation improves shape realism beyond mere visual plausibility. Figure 1 : Physical knowledge is crucial for realistic and functionally efficient 3D shape generation. Without physics, (a) generated shapes may appear visually plausible yet violate basic physical feasibility, such as car wheels intersecting the body or chairs with broken or unstable legs, and (b) aerodynamic shapes produce wide turbulent wakes, indicating low aerodynamic efficiency. (c) In contrast, physics-guided generation produces shapes with coherent flow and reduced drag, achieving both aesthetic appeal and physical efficiency. 1 Introduction Generative AI has achieved remarkable success in text Singh et al. 2025 and image Yazdani et al. 2025 processing. Meanwhile, 3D content creation has gained increasing attention with applications across a wide range of fields, including virtual reality Ma et al. 2025 , gaming Chen et al. 2025d , retail Chen et al. 2025c , and engineering design Vatani et al. 2025 . The existing 3D generative models Zhang et al. 2023 ; Chen et al. 2025b ; Chan et al. 2022 ; Li et al. 2025a excel in creating apparently realistic objects. However, such realism is only skin-deep. Close examination often reveals inconsistencies or implausibilities, which might be acceptable in some entertainment applications where the objective is merely to produce visually appealing results, but become detrimental when the end goal is engineering-oriented. For example, as illustrated in Fig. PhysGen: Physically Grounded 3D Shape Generation for Industrial Design , the wheels of generated cars Li et al. 2025a touch the body of the car, which makes no sense when designing functional cars. Similarly, the chairs Lai et al. 2025 generated from images contain feet with the wrong topology, making them unsuitable to bear the weight of a person. These failings can be traced to the fact that these methods are trained solely on static datasets Deitke et al. 2023 ; Chang et al. 2015 of 3D shapes, without regard to the engineering design process that was used to create them. For instance, automobile and airplane designs Regenwetter et al. 2022 ; Song et al. 2024 are usually optimized for aerodynamic efficiency, which conditions their shape beyond just aesthetics. Such physical awareness is absent from the current 3D shape generation pipelines Chan et al. 2022 ; Li et al. 2025a ; Lai et al. 2025 . To bridge this gap, we demonstrate that physical knowledge can be used to define regularization constraints, which are then used to improve the quality and realism of 3D shape generation. To this end, we present PhysGen , an approach to enforcing physical realism on generated 3D shapes, by using information around the physics of the shape. Specifically, we propose a physics-guided flow matching model, built upon an alternating update algorithm that switches between a velocity-based update and a physics-based refinement. Given the desired physical properties, we further enhance the physical guidance by introducing a physics-aware regularization in the velocity-based update. We update the latent code leveraging gradients derived from the discrepancy between the predicted and target physical values. Notably, 3D generative models typically rely on VAE-learned latent spaces, but existing shape VAEs Zhang et al. 2023 ; Chen et al. 2025b encode no physics, making physical properties unrecoverable from their latent codes. To enable effective physics-based regularization, we propose a shape-and-physics variational autoencoder (SP-VAE) that embeds both 3D shape and physics information into a unified latent space. In this paper, we focus on the application of 3D shape generation in automobile design, where physical performance is pivotal to ensuring that generated shapes satisfy real-world engineering requirements. We also showcase additional applications, such as airplane design, which demonstrates the generalization ability of our approach. We evaluate our PhysGen across three benchmarks Elrefaie et al. 2024a ; Chang et al. 2015 ; Sung et al. 2025 , including unconditional 3D generation, 3D generation conditioned on a sketch, and 3D generation conditioned on a real single-view image. The results demonstrate that by utilizing physics-based regularization, PhysGen produces shapes that are more geometrically plausible and physically efficient than previous approaches. We also conduct comprehensive ablation studies on the key components in our pipeline. In summary, our contributions are as follows: • We investigate the physical realism in 3D shape generation models and propose a novel physics-guided flow matching model that generates physically efficient and aesthetically pleasing 3D shapes. • We achieve the physical guidance by alternating between a velocity-based update with physics-aware regularization and a physics-based refinement. • We build a shape-and-physics variational autoencoder that encodes 3D shape and physics in a unified latent space, enabling the physical guidance in the presented flow matching model. 2 Related Work 2.1 3D Shape Generation Inspired by advances in image and video generation, generative models for 3D shape generation Zhang et al. 2023 ; Chen et al. 2025b ; Li et al. 2025a ; Wu et al. 2024b ; Zhang et al. 2024 ; Zhao et al. 2023 have recently made rapid progress. One line of work distills knowledge from powerful 2D diffusion models into 3D domains Chen et al. 2023 ; Li et al. 2023 ; Lin et al. 2023 ; Liu et al. 2024 ; Poole et al. ; Qiu et al. 2024 ; Tang et al. , alleviating the scarcity of high-quality 3D data. While effective at producing visually appealing shapes, these methods often exhibit slow convergence and unstable optimization. Another line of work builds native 3D generative models Zhang et al. 2023 ; Chen et al. 2025b ; Li et al. 2025a ; Zhang et al. 2024 . Methods such as 3DShape2VecSet Zhang et al. 2023 and Dora Chen et al. 2025b commonly adopt a two-stage paradigm, where a 3D VAE first encodes geometry into a latent space, followed by a diffusion model for 3D shape generation. While achieving visually plausible shapes, neither the VAE nor the diffusion models encodes physics information, which limits their usefulness in physics-sensitive domains. To address this, we propose a joint shape-physics VAE that embeds geometry and physics information into a unified latent space, capturing their intrinsic correlations and enabling more physically

Read the complete text