Bilingual Keypoint Evaluation Benchmark for Text-to-Image Models

T2I-Keypoints-Eval is a bilingual evaluation dataset for assessing text-to-image models. It was created by researchers from Tencent Hunyuan and released by PromptEnhancer in September 2025. The dataset includes Chinese and English text prompts paired with keypoint-annotated images for model assessment.

Use Cases

• Benchmark model performance on generating images with correct keypoint placements from bilingual text prompts.
• Evaluate spatial reasoning capabilities of text-to-image models by analyzing generated images against ground-truth keypoint annotations.
• Compare model robustness across Chinese and English language prompts using the provided parallel text data.
• Assess the accuracy of human or object pose generation in synthetic images based on textual descriptions of keypoints.

Strengths

• Bilingual dataset supporting evaluation in both Chinese and English languages.
• Specifically designed for a focused evaluation task on keypoint generation.

Limitations

• Specific row count, column details, and sample data are unavailable for assessing size and granularity.
• The dataset's scope is limited to the specific task of keypoint evaluation, not general text-to-image generation.

Provenance

Source

Tencent Hunyuan research team.

Collection Method

Constructed as an evaluation benchmark, likely involving curated text prompts and corresponding keypoint-annotated images.

Freshness

Last updated on September 16, 2025.