Human Preference Labels for Video Generation Motion Quality

115,732 pairwise human preference labels compare outputs from 4 frontier video generation models. Annotators from Datapoint AI evaluated videos across 3 quality dimensions, using 417 unique prompts and 11 motion categories. The dataset was last updated in March 2026.

Use Cases

• Training reward models for RLHF using pairwise preference labels on motion quality dimensions.
• Benchmarking video generation models by comparing their performance on 11 distinct motion categories.
• Analyzing human preferences across 3 specific quality dimensions for AI-generated video.
• Investigating prompt design by studying preferences across 417 unique text prompts.

Strengths

• 115,732 pairwise preference labels provide a substantial basis for model training.
• Covers 4 different frontier video generation models for comparative analysis.
• Expanded scope includes 417 unique prompts and 11 motion categories.

Limitations

• Specific column names, data structure, and sample size per model are unknown.
• Potential for annotator bias or inconsistency inherent in human preference data.
• Geographic and demographic coverage of the annotators is unspecified.

Provenance

Source

datapointai via Hugging Face.

Collection Method

Pairwise human preference labels collected from real annotators via Datapoint AI.

Freshness

Last updated March 2026.

Geography

Region tag suggests 'us', but specific coverage is unknown.