HPDv3: 1.08M Text-Image Pairs with Human Preference Annotations

Mizzen AI, CUHK MMLab, and academic partners released the Human Preference Dataset v3 (HPDv3) in August 2025. It comprises 1.08 million text-image pairs and 1.17 million annotated pairwise comparisons for modeling human preferences. The dataset is associated with the ICCV 2025 paper 'HPSv3: Towards Wide-Spectrum Human Preference Score'.

Use Cases

• Training reward models for text-to-image generation based on pairwise human preference annotations.
• Benchmarking the alignment of AI-generated images with human preferences based on the annotated comparisons.
• Developing and evaluating wide-spectrum human preference scoring models based on the dataset's scale and annotation structure.

Strengths

• Contains 1.08 million text-image pairs, providing a substantial base for training.
• Includes 1.17 million annotated pairwise comparisons for direct preference modeling.
• Associated with peer-reviewed research (ICCV 2025) and developed by multiple academic and industry labs.

Limitations

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

Provenance

Source

MizzenAI, CUHK MMLab, King's College London, Shanghai Jiaotong University, Shanghai AI Laboratory, CPII InnoHK.

Collection Method

Likely collected and annotated for research on human preference scoring.

Freshness

Last updated 2025-08-26 08:37:34.