← News

Agents & MCPSources: arXiv · 16 Nov 2026

TelePhysica turns a single photo into a verifiable, physics-based 3D world

The arXiv paper describes a closed-loop system that reconstructs an executable physical scene from one uncalibrated photograph, verifies it for gravitational stability, and uses the simulation to steer video generation.

In plain words

The authors propose a system that takes one ordinary photo and builds a simulated 3D version of the scene with objects, joints, and cloth that obey physics. A checking loop rejects unstable reconstructions before use, and the running simulation then guides a video-generation model. In their tests this made the generated motion more accurate than direct image-to-video methods.

What was announced

The paper argues that existing image-to-video diffusion models mix pixel synthesis with implicit motion, which prevents persistent 3D mechanical equilibrium, mid-rollout intervention, and counterfactual force control, while single-pass 3D reconstruction from an uncalibrated photo leaves depth and support clearance underconstrained so scenes interpenetrate or collapse under gravity. TelePhysica addresses this with a vision-language coding agent that produces separable visual meshes and declarative physical models inside a two-stage verification loop combining pixel-aligned silhouette critique with a zero-force gravitational equilibrium preflight. The verified world then compiles into a unified solver covering rigid bodies, articulated chains, cloth shells, and tetrahedral elastodynamics, whose rollouts condition a pretrained video diffusion model through pixel-aligned structural streams.

On the authors' 100-scene single-image benchmark, closed-loop execution raised equilibrium-verified world recovery from 0% with single-pass synthesis to 81%. Their ablations also report an order-of-magnitude larger localized corner displacement and out-of-plane pleating in finite-element cloth than in matched rigid-plate proxies, and conditioning video diffusion on simulated trajectories increased dynamic motion-mask overlap by 2x-5x while preserving motion direction under counterfactual force reversals where direct image-to-video models failed. These figures come from the paper's own evaluation and have not been independently confirmed.

What it does not show

  • Evidence is limited to the paper's abstract; the full methodology, experimental setup, and peer-review status are not available in the supplied passages.
  • The benchmark results and ablation figures are the authors' own reported numbers; no independent replication or confirmation is described.
  • The abstract does not state whether code, models, or a working demo are publicly available.

For security and IT teams

Scope
The supplied material covers only the paper's abstract, so claims about system behavior rest on the authors' description rather than the full article.

Sources

Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 17 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.