metadata
license: openrail
task_categories:
- image-classification
tags:
- art
size_categories:
- 100K<n<1M
Used to detect severely stylistically corrupted anime images generated by Stable Diffusion.
There are two classes: corrupted and normal, comprising 61,100 and 63,004 images, respectively.
For the corrupted type, it includes several cases such as:
- Entirely black images
- Mosaics or stylistic anomalies due to low sampling steps
- Stylistic anomalies due to excessively weighted tags
- Stylistic anomalies due to embedding with excessive weights
It's important to note that the following content is not included:
- Distortions in facial and hand details caused by resolution issues
- Distortions in human and object structures
For the normal type, it contains regular AI-generated images as well as approximately 15,000 images hand-drawn by humans.