ImageRewardDB: 137k Text-to-Image Human Preference Comparisons

ImageRewardDB is a text-to-image human preference dataset containing 137,000 expert comparison pairs. It was created by zai-org and uploaded to Hugging Face on June 21, 2023. The dataset is built from text prompts and corresponding model outputs sourced from DiffusionDB.

Use Cases

• Training reward models for text-to-image generation based on expert preference comparisons.
• Benchmarking the alignment of generative model outputs with human aesthetic judgment.
• Analyzing the relationship between text prompts and perceived image quality based on the comparison data.

Strengths

• Contains 137,000 expert comparison pairs, providing a substantial sample for analysis.
• Designed with a specific pipeline, including criteria for quantitative assessment and annotator training.
• Focuses on text-to-image human preference, a key metric for generative model evaluation.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is known, but the specific data structure and file formats are unknown.
• Last updated 2023-06-21 06:36:29; freshness should be verified.

Provenance

Source

zai-org via Hugging Face.

Collection Method

Constructed from text prompts and model outputs from DiffusionDB, using a designed pipeline with expert comparisons.

Time Range

null

Freshness

Last updated 2023-06-21 06:36:29.

Geography

null