ma-xu-fine-t2i

# Fine-T2I: An Open, Large-Scale, and Diverse Dataset for High-Quality T2I Fine-Tuning [[arxiv]](https://arxiv.org/abs/2602.09439) by [Xu Ma](https://ma-xu.github.io/), [Yitian Zhang](https://bespontaneous.github.io/homepage/), [Qihua Dong](https://dddraxxx.github.io/), [Yun Fu](http://www1.ece.neu.edu/~yunfu/) **Northeastern Univeristy** # Please see our [[Dataset Explore]](https://huggingface.co/spaces/ma-xu/fine-t2i-explore) to view detailed samples (loading is slow, be patient). ---- ## 🆕 What's New - **[2026.02.20]**: **Fine-T2I** reaches the #1 spot among Hugging Face Datasets Trending list ⭐️⭐️⭐️ - **[2026.02.16]**: **Fine-T2I** tops the Hugging Face Datasets Trending list, reaching the #2 spot and #1 spot for image datasets! ⭐️⭐️⭐️ ---- Fine-T2I is a large-scale, high-quality, and fully open dataset designed to advance SOTA text-to-image (T2I) fine-tuning. Comprising **over 6 million text–image pairs (approximately 2 TB)**, Fine-T2I was constructed to bridge the performance gap between open community models and enterprise-grade models. The dataset distinguishes itself through a rigorous construction pipeline that combines high-fidelity synthetic data with professional real-world photography, ensuring exceptional visual quality and precise instruction adherence. <div align="center"> <img src="z_assets/teaser.jpg"> </div> ---- ## Key Features - **Massive Scale & Quality:** Contains ~6.15M synthetic samples generated by SOTA diffusion models (Z-Image, FLUX2) and ~168k curated real images from professional sources (Pexels, Pixabay, Unsplash). - **Diverse Distributions**: Synthetic samples span 10 task combinations, 32 prompt categories, and 11 distinct visual styles. - **Dual Prompt Annotations**: Each image is paired with two types of prompts (original and enhanced) to support diverse user behaviors (The prompt for generating synthetic image is saved in txt). - **High Resolution**: Fine-T2I supports randomized aspect ratios and high resolutions (mostly >1K). - **Rigorous Filtering**: The data underwent a strict cleaning pipeline involving prompt semantic deduplication, safety checks, aesthetic filter, a VLM-thinking-based "Strict Visual Quality Auditor" and other filtering steps, filtered out total >95% of candidates to ensure flawless text–image alignment and zero artifacts. <div align="center"> <img src="z_assets/example_one_sample.jpg"> **Example of one sample in our dataset (image file: top-left, json file: top-right, and txt file: bottom)** You can ignore the json file if just fine-tuning a T2I model without further advanced filtering or setting. </div> ---- ## Statistical Analysis | Folder | samples | Storage | |---------------------------------------------|-----------|---------| | synthetic_enhanced_prompt_random_resolution | 1,615,592 | 476G | | synthetic_enhanced_prompt_square_resolution | 1,538,253 | 517G | | synthetic_original_prompt_random_resolution | 1,686,498 | 479G | | synthetic_original_prompt_square_resolution | 1,305,350 | 436G | | curated | 168,424 | 259G | <div align="center"> <img src="z_assets/distribution.jpg"> </div> <div align="center"> <img src="z_assets/aesthetic.jpg"> </div> <div align="center"> <img src="z_assets/prompt_length.jpg"> </div> ---- ## Useage Fine-T2I is organized in the **WebDataset** format, each sample consists of three corresponding files: `jpg`, `txt`, and `json`. You can easily load the dataset in **streaming mode** (without downloading the entire 2TB) using the Hugging Face datasets library. ```python from datasets import load_dataset import json # we use synthetic_enhanced_prompt_random_resolution as an example url_pattern = "https://huggingface.co/datasets/ma-xu/fine-t2i/resolve/main/synthetic_enhanced_prompt_random_resolution/train-*.tar" dataset = load_dataset( "webdataset", data_files={"train": url_pattern}, split="train", streaming=True # please do streaming, or you will download the whole dataset ~2TB ) for i, sample in enumerate(dataset): if i >= 10: break print(f"\n\nImage info: {sample['jpg'].size}\n") print(f"Text: {sample['txt']}\n") print(f"Meta json data: {json.dumps(sample['json'], indent=4, ensure_ascii=False)}\n") print(f"=========="*5) ``` ---- # Cite If you find the data useful, please cite: ``` @misc{ma2026finet2i, title={Fine-T2I: An Open, Large-Scale, and Diverse Dataset for High-Quality T2I Fine-Tuning}, author={Xu Ma and Yitian Zhang and Qihua Dong and Yun Fu}, year={2026}, eprint={2602.09439}, archivePrefix={arXiv}, primaryClass={cs.CV}, url={https://arxiv.org/abs/2602.09439}, } ``` # Contact For any issues or questions, please feel free to contact Xu Ma: [e-mail] # Thanks We are very grateful to the other datasets（like LAION, BLIP3o, JourneyDB, etc） that laid the foundation for text-to-image generation, and we also appreciate the templates provided by BLIP3o. Thanks.