BiComp: Large-Scale Text-to-Image Preference Dataset with Region-Level Annotations

BiComp is a large-scale, high-quality text-to-image preference dataset containing 57,474 original and 94,502 edited images. The dataset is annotated with region-level information and filtered through a VQA-based quality control step. It was introduced by anzeameol in a 2026 paper on compositional text-to-image generation.

Use Cases

• Training text-to-image reward models based on human preferences across six compositional dimensions.
• Benchmarking image generation models on compositional attributes like color, shape, and texture.
• Developing region-aware image editing tools based on the provided region-level annotations.
• Studying numeracy and spatial relationships in multimodal AI systems.

Strengths

• Large scale with 151,976 total images (57,474 original + 94,502 edited).
• High-quality annotations include region-level information for detailed analysis.
• Incorporates a VQA-based quality control step for data filtering.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Last updated 2026-06-03 08:57:11; freshness should be verified.

Provenance

Source

huggingface

Collection Method

Introduced in the paper 'Compositional Text-to-Image Generation Via Region-aware Bimodal Direct Preference...'; specific collection method not detailed.

Freshness

Last updated 2026-06-03 08:57:11.