Cosmos-HumanEval-v1: AI Video Generation Evaluation Prompts

HUE-PaiBench v1.2 is an open-source subset of NVIDIA's Cosmos Human-Eval question bank for evaluating generated videos. It consists of 197 prompts split into two tasks: 100 for image-to-video and 97 for text-to-video generation. The dataset was published by NVIDIA on Hugging Face.

Use Cases

• Benchmarking text-to-video AI models based on the 97 text-to-video prompts.
• Evaluating image-to-video AI models based on the 100 image-to-video prompts.
• Comparing model performance on the standardized PaiBench prompt set mentioned in the description.

Strengths

• Contains 197 total prompts for video generation evaluation.
• Specifically split into 100 image-to-video and 97 text-to-video prompts.
• Published by NVIDIA, a major AI research institution.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.

Provenance

Source

NVIDIA

Collection Method

Subset of a proprietary question bank curated for AI video evaluation.

Freshness

Last updated 2026-05-31 18:58:33; freshness should be verified.