Text-to-Image DPO Human Preferences: 416,360 Pairwise Judgments

DatapointAI created a dataset of 416,360 pairwise human judgments comparing AI-generated images. The data was collected from approximately 20,000 annotators, focusing on prompt alignment and overall preference. The full, unfiltered version was last updated on March 30, 2026.

Use Cases

• Training reward models for text-to-image generation based on human preference data.
• Fine-tuning image generation models via Direct Preference Optimization using pairwise comparisons.
• Analyzing human preferences for AI-generated art across different evaluation dimensions.
• Benchmarking the alignment of image generation models with textual prompts.

Strengths

• Contains 416,360 pairwise human judgments, providing a large-scale preference signal.
• Sourced from approximately 20,000 annotators, suggesting a diverse set of human evaluators.
• Explicitly evaluates images on two dimensions: prompt alignment and overall preference.

Limitations

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

Provenance

Source

datapointai

Collection Method

Human annotators provided pairwise judgments comparing AI-generated images.

Freshness

Last updated 2026-03-30 06:41:17; freshness should be verified.