Text 2 Video Human Preferences: 29,283 Motion Quality Labels

29,283 pairwise human preference labels comparing human motion quality across four frontier video generation models, released by Datapoint AI in February 2026. The dataset captures evaluations from 4,349 unique annotators focusing on three specific quality dimensions of AI-generated video.

Use Cases

• Training reward models for RLHF using the 29,283 pairwise preference labels
• Benchmarking text-to-video models against three specific human motion quality dimensions
• Analyzing annotator agreement across 4,349 unique contributors to improve video evaluation metrics

Strengths

• 29,283 pairwise human preference labels
• Evaluations from 4,349 unique annotators
• Covers 4 frontier video generation models

Limitations

• Narrow focus on human motion quality rather than general video fidelity
• Lack of explicit column documentation in the provided metadata

Provenance

Source

Datapoint AI

Collection Method

Human annotation via the Datapoint AI platform

Freshness

Last updated February 2026.

Geography

United States