False positives
#6
by
Dolfik
- opened
Hello. Thanks for this great model. I use this model to validate images that users generate, and I found that the model is prone to false positives when there are red tones in the colors in the image.
Here is some examples:
Looks like the training is causing skin-tone (color) to be oversensitive to false positives.
Yes flesh tones trigger false positives for obvious reasons. Lol
It’s obvious that this data and model are very brittle. As soon as it looks like a woman it’s unsafe, but men are barely considered unsafe