AIML-TUDA/t2i-diversity-captions

📄 Synthetic Captioned LAION Subset

🧾 Dataset Summary

This dataset consists of over 39 Million synthetic image captions generated for 1 Million curated images from LAION Aesthetics. Images have an aesthetics score >6, at minimum resolution of 512p, and have been screened for NSFW, CSAM and watermarks. We also removed exact duplicates.

📦 Data Structure

• strID - Unique string identifier for the sample
• intID - Unique integer identifier for the sample
• imageURL - Original LAION image URL
• captionX - Up to 10 synthetic captions (caption0, ..., caption9). Please note that some captions might be None

As of June 2025, approximately 74% of the image URLs were still accessible.

🧪 Subsets / Configurations

Name	Model	Caption Prompt	Target Caption Length	Note
llava_1.6_7B_base_30	LLaVA 1.6 7B	base	30
llava_1.6_7B_base_60	LLaVA 1.6 7B	base	60
llava_1.6_34B_base_30	LLaVA 1.6 34B	base	30
llava_1.6_34B_base_60	LLaVA 1.6 34B	base	60
internvl_2_llama3_76B_base_10	InternVL-2 LLaMA3 76B	base	10
internvl_2_llama3_76B_base_20	InternVL-2 LLaMA3 76B	base	20
internvl_2_llama3_76B_base_30	InternVL-2 LLaMA3 76B	base	30
internvl_2_llama3_76B_base_40	InternVL-2 LLaMA3 76B	base	40
internvl_2_llama3_76B_base_50	InternVL-2 LLaMA3 76B	base	50
internvl_2_llama3_76B_base_60	InternVL-2 LLaMA3 76B	base	60	caption0 is sampled at default temperature. caption1 - caption9 are sampled at t=1.5

Base prompting refers to the VLM captioning prompt from App. A.1.

Additionally we provide a internvl_2_llama3_76B_persona_all with persona based prompts of different lengths. Here column names describe the persona used in caption prompting (App. E) and the respective target length. Please note that the majority of values will be None, but each image has at least 15 different captions.

📚 Intended Use

Designed for use in training and evaluating text-to-image generation models. Ideal for research on the impact of caption length, model choice, and prompt structure.

⚠️ Caveats

Image availability is subject to web persistence; only 74% of links were functional as of June 2025. Captions are synthetic and model-dependent, potentially introducing model-specific biases or hallucinations.

📤 Citation

If you use this dataset, please cite:

@article{brack2025howtotrain,
  title={How to Train your Text-to-Image Model: Evaluating Design Choices for Synthetic Training Captions},
  author={Manuel Brack and Sudeep Katakol and Felix Friedrich and Patrick Schramowski and Hareesh Ravi and Kristian Kersting and Ajinkya Kale},
  journal={arXiv preprint arXiv:2506.16679},
  year={2025}
}