Any movement, rebuilt in 3D

Film someone walking, slipping or falling. Claude rebuilds the scene, poses the figure from what it sees, and plays it beside your clip.

On hold: development resumes with the next major version of Claude. This release keeps working.

Gait Bench 2.0.1 · by Nulytica · free

Three steps. No suit, no studio.

Film

Any phone, portrait or landscape. Feet in frame matters most.
Ask Claude

Attach the video and say "recreate this in 3D". Claude reads the lens, the place and the person, poses the body at the moments that matter, compares its render with your clip, and says why.
Compare

Your clip and the rebuild play side by side, in sync. Overlay them, orbit the scene, or pop it out to a second screen.

Built by reasoning about your footage, not by measuring it. Broadly the same place, person and movement.

The sceneGround, sky, sun, trees and buildings on the horizon, and the camera's own movement: turns, dollies and hand-held shake.
The movementA figure posed by Claude from the frames: steps, slips, hands on the ground, falls. Planted feet stay put; sliding feet are marked.
Terrain and real shoesBanks, kerbs and walls on a half-metre grid. Shoes traced from real photos, never generated.
A page to share with yourselfOne HTML file with both videos, overlay, orbit and pop-out, plus Claude's reasons for every choice.

Free. Runs on your own Claude plan.

Most people

Desktop or web. Needs a paid Claude plan.

Developers

Paste this into your terminal:
claude plugin marketplace add erdoar/gait-bench-plugin && claude plugin install gait-bench@gait-bench
Updates later: /plugin marketplace update.

Needs Node.js 18+ and ffmpeg.

Agents

Give this to Claude Code or any AI agent:
Install the Gait Bench plugin by following https://gait.nulytica.com/llms.txt
It reads the agent instructions and does the rest.

A reconstruction, not a measurement: speeds, distances and heights are Claude's best judgement from looking, and it tells you which guesses mattered.
Best clips: one person walking or standing, feet visible, up to about 15 seconds. Falls, running and crowds aren't shown yet.
Small and offline: plain JavaScript and ffmpeg on your own computer. No GPU, no AI image or 3D generators, nothing uploaded.
Your footage stays yours: the result page carries your clip so it can play beside the rebuild, so keep it as private as the video.

The deal: AMD is buying Fei-Fei Li's World Labs for about $8.2 billion in stock (announced 28 Sep 2026).
The people: Li becomes AMD's chief scientist, reporting to Lisa Su; closing is expected by the end of 2026.
The models: Atlas, a 3D world model; Marble, for building 3D scenes; Spark 2.0, for rendering them.
Lisa Su: “…requires a deep understanding of how models are evolving.”
For Gait Bench: it tried World Labs for scenery in 0.x; 1.x uses no outside models. Source: CNBC