CoW-Bench: 1K-10K Prompts for Generative World Model Consistency

OpenRaiser released CoW-Bench in March 2026 as a benchmark containing between 1,000 and 10,000 records for evaluating generative world models. It measures model performance across spatial relationships, object interactions, and temporal dynamics in text-to-image and video generation tasks.

Use Cases

• Benchmarking text-to-video models on their ability to maintain temporal dynamics using CoW-Bench prompts
• Evaluating spatial reasoning in image generation models by measuring object placement accuracy
• Testing object interaction consistency in video-to-video or image-to-video generation workflows

Strengths

• Apache 2.0 licensed for open research and commercial use
• Targets three specific consistency dimensions: spatial, interaction, and temporal
• Includes support for both image and video generation modalities

Limitations

• Specific column names and schema details are not provided in the metadata
• Small sample size (under 10,000 records) compared to general-purpose generative datasets

Provenance

Source

OpenRaiser, associated with the paper 'The Trinity of Consistency as a Defining Principle for General World Models'

Freshness

Last updated March 12, 2026.