|
|
--- |
|
|
license: apache-2.0 |
|
|
--- |
|
|
## Distorted-5K Dataset |
|
|
### Overview |
|
|
|
|
|
The Distorted-5K dataset comprises 5,000 synthetic images generated by text-to-image (T2I) models, as introduced in the paper "Evaluating and Predicting Distorted Human Body Parts for Generated Images". |
|
|
|
|
|
Each entry includes: |
|
|
* **{i}.png**: A high-resolution synthetic human-related image that may exhibit common distortions. |
|
|
* **{i}_label.json**: Detailed annotations labeling distorted regions, including type, position, and multi-annotator consensus. |
|
|
|
|
|
### Potential Use Cases |
|
|
* Benchmarking T2I models. |
|
|
|
|
|
* Training/fine-tuning detectors to identify human distortions in AI-generated images. |
|
|
|
|
|
* Fine-tuning T2I models to generate distortion-free human images. |
|
|
|
