Text 2 Image Rich Human Feedback
Dataset of detailed human evaluations of images generated by AI, with identification of errors in the correspondence between text and image, collected via Rapidata.
Description
The dataset Text 2 Image Rich Human Feedback contains annotations from more than 150,000 human reviewers who judged the quality, consistency, and fidelity of images generated from text prompts. Each annotation identifies the words misrepresented in the image or the defects observed, thus allowing a detailed analysis of the performance of the generation models.
What is this dataset for?
- Improve and refine text-image generation models through detailed human feedback
- Analyze common representation errors in generated images
- Designing qualitative assessment systems for the multimodal generation
Can it be enriched or improved?
This dataset can be extended by adding new human annotations according to other quality or aesthetic criteria, or by cross-referencing with automatic annotations. Adaptation to different styles or image areas is also possible.
🔎 In summary
🧠 Recommended for
- Researchers in image generation
- Text-to-image template developers
- AI assessment teams
🔧 Compatible tools
- Hugging Face Datasets
- Pandas
- Apache Parquet
- Rapidata API
💡 Tip
Use the streaming function to quickly manipulate this large dataset without a complete download.
Frequently Asked Questions
How were human annotations collected?
Through Rapidata API, over 150,000 reviewers annotated images generated by indicating errors in the text-image correspondence.
Is this dataset suitable for improving text-to-image models like Stable Diffusion?
Yes, it provides accurate feedback that is essential to refine the quality of generations.
What is the format of the data and how can it be handled effectively?
Data in Parquet format (17.9 GB); streaming via Hugging Face Datasets makes it easy to handle without a complete download.
