Synthetic Dataset DALL·E 3 + Captions
This dataset brings together over a million images generated by AI, mainly with DALL·E 3, accompanied by rich captions created via CogVLM and Dolphin 2.6.
Description
Synthetic Dataset DALL·E 3 + Captions is a massive corpus of images generated by artificial intelligence, mostly from DALL·E 3, Midjourney v5/v6 and Stable Diffusion. Each image is accompanied by one or more captions generated automatically using advanced models (CogVLM, Llama 3, Dolphin Mistral), with error detection and continuous improvement.
What is this dataset for?
- Training captioning and vision-language models
- Pre-training of generative models (diffusion, GaNS) using high quality data
- Studies on the human perception of AI-generated images
Can it be enriched or improved?
Yes, possible enhancements include: reclustering by theme or visual style, improving captions with human feedback, or finer filtering according to creative or aesthetic axes. This dataset can also be cross-referenced with human judgments of quality or preference.
🔎 In summary
🧠 Recommended for
- AI vision engineers
- Multimodality researchers
- Captioning assessment projects
🔧 Compatible tools
- Hugging Face Datasets
- OpenCV
- BLIP
- CogVLM
- Lava
💡 Tip
To create a thematic sub-corpus (e.g. landscapes, objects, urban scenes), use captions as an automated semantic sorting criterion.
Frequently Asked Questions
Can this dataset be used to train an image generation model?
Indirectly yes, it can be used as a basis for inspiration or quality score, but it does not include the original prompts needed for training.
Can the images in this dataset be used commercially?
Yes, the MIT license allows free use, including in commercial contexts, with attribution if desired.
Are the captions generated reliable for NLP use?
Yes, they are produced using specialized models with error detection, but may require validation for critical cases.
