Researchers at Seoul National University presented Rest2Art — a method that recovers the geometry of movable parts and joint parameters of everyday objects from a single photograph of the object in a closed state. The work was accepted at ECCV 2026, and reconstruction quality is comparable to methods that observe real object motion.

What happened

Daeun Lee and colleagues from Seoul National University published Rest2Art — a 3D reconstruction method for articulated objects (cabinets, doors, laptops) from a single static image in a resting state. The architecture consists of three sequential stages: a vision-language model together with a segmentation model refines the object's part hierarchy, a video diffusion model synthesizes hypotheses of possible motions, and joint parameters are resolved through optimization of geometric consistency with the resulting mesh. The research was published on arXiv (2607.27749) on July 30, 2026, and accepted at the ECCV 2026 conference. The project page with interactive results is live, the dataset is available on HuggingFace, and the code is hosted on GitHub.

Context

Reconstruction of articulated objects is a fundamental task in computer vision and robotics, necessary for creating interactive 3D assets. Before Rest2Art, solving this task required observing the object in multiple states: motion capture, photo series, or video recordings of opening/closing. Without such observation, the task was considered ill-posed — from a single static image, it is impossible to unambiguously recover how the object's parts move. Rest2Art solves this problem in a fundamentally new way: a video diffusion model, trained on massive video corpora, is used not for image generation, but as a source of physically plausible articulation hypotheses, which are then filtered by geometric optimization.

Why this matters for industry

Rest2Art eliminates the most expensive stage in the pipeline for creating simulation assets for robotics — motion capture of articulated objects. Instead of a multi-step process with multiple cameras or manual 3D modeling, a single photograph is sufficient. This turns the Real-to-Sim-to-Real pipeline (real object, reconstruction, simulation, manipulation policy learning, execution on a real robot) from a research demonstration into a scalable process. The combination of VLM, video diffusion, and geometric optimization sets a new direction: using generative models not for content creation, but as a source of structural hypotheses that are refined by optimization. Quality on par with methods that observe real motion means that for many robotics tasks, Rest2Art can completely replace motion capture. With successful adaptation, the method could become a standard stage in digital twin creation pipelines and commercial simulation platforms such as Isaac and MuJoCo.

Why this matters for users

The project demonstrates a practical pattern: from a single photograph, an interactive 3D asset is created with fully working joint parameters. The dataset is available on HuggingFace for study and experiments. The project page contains interactive results and demonstration videos, allowing assessment of reconstruction quality on various objects. For researchers and engineers, this means that the VLM plus video diffusion plus geometric optimization methodology can be applied to related tasks right now by studying the paper and dataset.

What is still unknown / limitations

The GitHub repository at the time of analysis contains only a README — without training and inference scripts, model weights, and dataset processing pipeline, making reproduction of results difficult. The architecture combines VLM, segmentation, and a video diffusion model, which implies high computational consumption — no data on inference latency or GPU requirements has been published. Quality on complex objects with non-standard joint mechanisms was not evaluated. The research has just been published, and integration into production is not possible.

Sources

Author

Look at AI, editorial team