Gokayfem releases open Video 2 Physics workflow to correct AI video physics

• Video 2 Physics rebuilds faulty AI-video motion as solver-driven Blender scenes, according to gokayfem's demo.
• The workflow is open source: the repository announcement provides code for measurement, simulation and independent checks.
• A companion garden demo reconstructs a Gaussian splat using procedural objects, Blender nodes and captured textures, as gokayfem demonstrated.

In the five-case example pack, a pendulum's swing grows 62% without anything driving it, and a vortex street runs about 3.5 times too fast. The independent verifier is the best part: deliberately wrong inputs must fail its checks.

Gokayfem says Claude Opus 5.5 measured the clips, solved computational fluid dynamics and rod mechanics, checked results with an independent finite-element solver, and rendered corrections into the original scenes.

Five worked cases ship with the workflow:

• Brass pendulum: compound-pendulum dynamics with damping at the pivot.
• Spring pendulum: a five-degree-of-freedom model that includes a rocking weight.
• Torsion disc: twisting motion combined with sway and tilt.
• Buckling ruler: elastic bending under an off-center tip load and the ruler's own weight.
• Vortex shedding: two-dimensional fluid flow at Reynolds number 150, with dye tracking.

The agent follows six requirements:

1. Measure before modeling. Account for camera movement and refraction through water or glass; track motion with stated uncertainties.
2. Declare the model. Record each parameter's provenance and identify the physically coherent opening segment used to anchor the reconstruction.
3. Benchmark and refine. Check solvers against analytical or published results, then run three resolutions while tracking energy and mass budgets.
4. Verify independently. Recompute key quantities with a second implementation that shares no code with the first.
5. Audit the Blender bake. Drive the scene from solver files and compare its evaluated state against those files at frames and subframes.
6. Deliver review evidence. Produce same-time comparisons without retiming, a full-resolution render and a report, followed by human approval.

The vortex case reports drag agreement within 0.3% between implementations in the published results.

The README explicitly disclaims real-world accuracy for all five examples. They start from generated footage and demonstrate agreement between the rendered scene and a declared, benchmarked model.

Video 2 Physics is MIT-licensed and distributed as a coding-agent skill. Its installation instructions cover Claude Code, Codex, Cursor, Copilot and Gemini CLI.

• Skill installation: npx skills add gokayfem/video-to-physics.
• Base runtime: Python 3.10+, NumPy, OpenCV, Pillow, FFmpeg and Blender 5.2+.
• Solvers: installed per case, with examples including FEniCSx for finite elements and Basilisk for fluid dynamics.
• Optional clip generation: requires a fal API key in FAL_KEY.

Gokayfem also gave Claude Opus 5.5 a Gaussian splat of a garden and asked it to rebuild the scene in Blender.

Asked about manual intervention, Gokayfem described the run as fully automatic. The published toolkit requires project-specific source assets, camera registration and configuration; it bundles no automatic semantic segmentation.

In the released garden demonstration, captured appearance is baked onto the table and paving, while other reconstructed parts retain node-based construction. The source scan and fitted hybrid assets are excluded from the repository.

The appearance-mapping documentation adds an editing constraint: mapped surfaces require rebaking after geometry or UV changes. Neither supported appearance mode recovers clean physical albedo.