World Labs' Atlas Tops Camera Control Rankings 127 Points Ahead of Google

World Labs’ Atlas has taken first place on Rapidata’s camera movement leaderboard, which tests whether video and world models follow instructions such as pan left, tilt down, and roll clockwise. In the published v1 results, Atlas scored 1469 Elo and won 76.7% of its evaluated pairwise matches, based on majority votes from human raters.

Atlas finished 127 Elo points ahead of Google’s Gemini Omni Flash 1.2, which scored 1342. Accurate camera control matters for tools that depend on repeatable paths through a scene, including virtual production, game development, simulation, robotics, and 3D reconstruction.

Atlas opens a 127-point lead

The leaderboard covers 19 video generators and world models. Selected placements show how Atlas compares with systems from Google, Alibaba Cloud, NVIDIA, Kuaishou, and xAI.

Google’s Veo 3.1 ranked 14th, while NVIDIA’s Cosmos Predict 2.5 ranked 18th. The placements of Veo and Kling show that visual polish is an unreliable proxy for camera-instruction accuracy.

Human votes turn into Elo

Rapidata evaluates models through pairwise comparisons judged by people. Raters compare two outputs generated from the same request and select the one that follows the requested camera movement more accurately. Each match receives at least five responses, with the majority determining the winner.

Pairwise results are fitted with a Bradley-Terry model, which estimates each system’s relative strength from its wins and losses. Rapidata then rescales those estimates to an Elo-style range similar to the one used by Chatbot Arena. The scores compare models within this leaderboard; they are not absolute measures that transfer to other benchmarks.

The fixed test set contains 600 prompts built from 30 camera-movement instructions across 22 spaces. Instructions use scene-independent language, reducing the influence of content-specific cues and making results easier to reproduce.

Quality controls include gold-standard screening questions for annotators. The v1 results draw on 513,079 votes across 102,595 pairwise matches, and Rapidata publishes the judgments on Hugging Face.

Compound paths expose weak control

Camera movement often fails in ways that polished output can conceal. A prompt requesting forward motion through a library may yield a sideways drift, a zoom, or animated objects viewed from a stationary camera. The benchmark focuses its judgments on whether the requested camera action occurred.

Its movement vocabulary contains 14 primitives: move forward, backward, up, down, left, or right; tilt up or down; turn left or right; zoom in or out; and roll clockwise or counterclockwise. Prompts combine as many as three actions, requiring models to preserve scene geometry through a compound trajectory.

Where predictable motion pays off

World Labs describes Atlas as a generative world model designed to create and navigate spatial scenes. That design goal aligns closely with a benchmark centered on camera paths, although the leaderboard cannot isolate architecture from differences in training data, inference settings, and model scale.

Reliable execution of camera instructions supports several development workflows:

• Virtual production and previsualization tools that need repeatable shots
• Games and interactive environments with user-controlled cameras
• Robotics and simulation pipelines that generate geometrically consistent footage
• 3D reconstruction systems that depend on known camera trajectories

A narrow score with practical limits

The benchmark’s scope is camera-instruction following. Visual fidelity, motion realism, scene-content adherence, and physical plausibility require separate evaluation. Rapidata maintains dedicated physics and realism leaderboards, where the model order changes substantially.

Cost also sits outside the Elo ranking. The benchmark page includes a Pareto chart plotting score against price in U.S. dollars per generated video second. Models on the frontier have no competitor that is both cheaper and higher scoring, and several mid-ranked systems reach that frontier at lower prices.

Production evaluations should also cover latency, resolution, API quotas, licensing, output consistency, and performance on application-specific scenes. Atlas leads this camera-control test, while the best deployment choice depends on which of those constraints governs the product.