Rest2Art Recovers Object Kinematics from a Single Snapshot

The Rest2Art method from Seoul National University, accepted to ECCV 2026, recovers the part geometry and joint parameters of articulated objects from a single static snapshot without motion examples.

Compiled by Sergey KostenchukPublished 2026-08-13Updated 2026-09-18

2026-08-13 Research HuggingFace

🤖 Rest2Art Recovers Object Kinematics from a Single Snapshot

The method from Seoul National University (ECCV 2026) recovers part geometry and joint parameters — doors, cabinets, laptops — from a single static photograph in a closed state. A VLM refines the part hierarchy, a video diffusion model synthesizes motion hypotheses, and joint parameters are resolved through geometric consistency.

🌍 Previously, reconstruction of articulated objects required capture in multiple states. Rest2Art removes this requirement — the result is comparable to methods that observe motion. It accelerates the Real-to-Sim-to-Real pipeline for robotics.

👤 The combination of VLMs, segmentation, and video diffusion models solves a problem that seemed ill-posed. Dataset on HuggingFace, code on GitHub.

Source 1: https://da-eun07.github.io/rest2art/ Source 2: https://arxiv.org/abs/2607.27749

Sources