EditReward-Data: 200K Human Preference Pairs for Instruction-Guided Image Editing

EditReward-Data is a large-scale human preference dataset for instruction-guided image editing, introduced by TIGER-Lab. It comprises over 200,000 manually annotated preference pairs curated by trained experts following a standardized protocol. The dataset was last updated on October 12, 2025.

Use Cases

• Training reward models for instruction-guided image editing based on human preference annotations.
• Benchmarking the alignment of image editing models with human judgment based on preference pairs.
• Fine-tuning generative image models to follow complex editing instructions more accurately.
• Studying human preferences and quality assessment criteria for AI-generated image edits.

Strengths

• Contains over 200,000 manually annotated preference pairs, indicating substantial scale.
• Annotations were curated by trained experts following a rigorous and standardized protocol, suggesting high quality.
• The dataset is associated with a published research paper, providing academic context and validation.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown beyond the 'over 200K' figure, which may limit suitability assessment.
• Freshness should be verified as the last update timestamp is from 2025.

Provenance

Source

TIGER-Lab

Collection Method

Manually annotated preference pairs curated by trained experts following a standardized protocol.

Freshness

Last updated 2025-10-12 05:48:42