PhysAg3D, a Method for Creating Movable 3D Assets from a Single Image, Published on arXiv

What happened

The paper "PhysAg3D," posted to arXiv and dated October 29, 2026, proposes a method that creates, from a single image, a 3D asset that looks realistic, stays faithful to the original object, and moves in a physically plausible way. The authors are eight people, including Bi'an Du, Fanpeng Meng and Jiemin Fang.

The starting point is the problem that, even when a shape is detailed, shape alone does not determine how an object moves or deforms. The method therefore combines a 3D generative model with an agent that performs reasoning and tool use.

In the pipeline, a vision-language agent estimates semantic parts, possible motions and physical priors from the image. In parallel, a 3D generative model creates a textured mesh of the whole object from the same image. The project page is listed as supra-3d.github.io/PhysAg3D.

How it works

The core is a "hierarchical interaction representation." It ties revisable physical priors to the motion mechanisms and to the generated mesh surfaces those mechanisms control, and uses this to organize the construction and refinement of motion.

Guided by this representation and by observations of the mesh, the agent uses geometry tools and physics tools. These are used to build parts, joints, deformable regions and bindings to the surface. The paper says it preserves the useful parts of the generated geometry and materials.

After construction, the agent looks at rendered results of the motion and runtime measurements, and makes targeted corrections within a fixed budget (such as a number of iterations). The paper says it supports both rigid-body joint motion and soft deformation.

Why it matters

In game and XR production, it could reduce the effort of modelers setting up movable objects such as doors and drawers one by one. However, practical accuracy and how it would fit into workflows cannot be confirmed from the source material.

Robot simulation needs articulated and deformable objects to manipulate. If such assets can be made from photos, it points toward helping prepare them, but the paper's abstract does not explicitly mention robotics use.

Caveats and open questions

• The source material is only the paper's abstract, so quantitative results, comparison baselines, processing time, and the types of 3D generative model and vision-language model used cannot be confirmed.
• The availability of code and data cannot be confirmed from the source material beyond the project page listing.
• The evaluation is reported by the authors. The abstract says the method "showed reasonable results on real photos and an external benchmark," but third-party verification has not been confirmed.
• This may be a pre-peer-review arXiv paper; the peer-review status is not known from the source material.

Glossary

Agent (agentic)

A system in which a language model assesses the situation and carries out work while calling tools.

Vision-language model

An AI model that can understand and reason over both images and text.

Joint

A connection in a 3D model that defines how parts move relative to each other, like a door hinge.

Surface binding

A setting that ties the mesh surface to the parts or mechanisms that control its motion.

Sources

Primary sources (publisher, paper, repository)

• PhysAg3D: Bridging 3D Generative and Agentic Models for High-Fidelity and Physically-Interactive AssetsarXiv（arxiv.org）・en

The X post that surfaced this story

• View @CSVisionPapers's post on Xx.com

This article was generated in Japanese by AI from the sources listed (primarily primary sources) and published automatically, without prior human review, after automated checks such as source matching (how we use AI). Figures and specifications are as stated by the publishers and have not been independently verified by us. Please consult the original sources. To report an error, use our contact form; we will review it promptly and correct it. This page is an AI translation of the Japanese original. If anything differs, the Japanese version prevails. Japanese original