deepghs/ai_image_corrupted

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.