VisionFoundry-10K

# VisionFoundry-10K  **VisionFoundry-10K** is a synthetic visual question answering (VQA) dataset with 10,000 image-question-answer triples spanning 10 vision-centric tasks. The data is produced by the VisionFoundry pipeline: an LLM generates task-aware questions, answers, and detailed text-to-image prompts; a text-to-image model synthesizes images; and a strong multimodal verifier filters samples for alignment. [](https://zlab-princeton.github.io/VisionFoundry) [](https://arxiv.org/abs/2604.09531) [](https://github.com/zlab-princeton/VisionFoundry) ## VisionFoundry: Teaching VLMs Visual Perception with Synthetic Images ## Authors [Guanyu Zhou](https://the-martyr.github.io/)¹, [Yida Yin](https://davidyyd.github.io/)¹, [Wenhao Chai](https://wenhaochai.com/)¹, [Shengbang Tong](https://tsb0601.github.io/)², [Xingyu Fu](https://zeyofu.github.io/)¹, [Zhuang Liu](https://liuzhuang13.github.io/)¹ ¹Princeton University, ²New York University ## Tasks (10) - Color_and_Appearance - Depth_Order - Orientation_and_Direction - Positional_and_Relational_Context - Real_World_Spatial_Understanding - Relative_Distance - Spatial_Relationship - State_and_Condition - Structural_and_Physical_Characteristics - Viewpoint_and_Perspective ## Files - `annotations.json` - `images.parquet` - `restore_images_from_parquet.py` ## Schema Each record in `annotations.json` contains: - `messages`: a 2-turn chat with user question and assistant answer - `images`: a relative image path like `Spatial_Relationship/q00796.png` - `qid`: sequential integer ID starting at 1 `images.parquet` contains: - `path` (string): relative path matching `images[0]` - `image_bytes` (binary): raw PNG bytes ## Examples  ## Restore Images To rebuild the original PNG folder structure from `images.parquet`: ```bat python restore_images_from_parquet.py ``` This will recreate the 10 task folders and write each PNG to its relative path. ## Citation ```bibtex @article{zhou2026visionfoundry, title={VisionFoundry: Teaching VLMs Visual Perception with Synthetic Images}, author={Zhou, Guanyu and Yin, Yida and Chai, Wenhao and Tong, Shengbang and Fu, Xingyu and Liu, Zhuang}, journal={arXiv preprint arXiv:2604.09531}, year={2026} } ``` > **⚠️ Disclaimer**: This dataset is **fully AI-generated with no human intervention**. This means there may be cases of image-text mismatch. While a VLM verifier filters out many misaligned samples, some may still pass through.