A new report warns that the proliferation of child sexual abuse images on the internet could become much worse if something is not done to put controls on artificial intelligence tools that generate deepfake photos.
Mixing with a basic prompt isn’t where the AI’s strengths lie, though. Selecting different parts of the image and giving each part a different description is the best way to get AI art out of the models that you can distribute yourself.
This AI has been fed so much porn that it can combine concepts like “big breasts” as easily as “a tree”. This leads to some hilarious mistakes (like adding multiple rows of breasts when it tries to combine “woman” and “big breasts”) but with basic masking and inpainting you can correct all that.
I’m all for banning AI models trained on child abuse imagery¹, but the assertion that AI generated child porn must be based on a model trained on existing illegal material is simply not true.
¹ unlike many AI enthusiasts, I consider AI models to contain significant parts of the training set, which is why I think models trained on illegal material shouldn’t be redistributed and generated AI artwork should be considered derivative works of the many images that are part of the original dataset. Same goes for ChatGPT.
Mixing with a basic prompt isn’t where the AI’s strengths lie, though. Selecting different parts of the image and giving each part a different description is the best way to get AI art out of the models that you can distribute yourself.
This AI has been fed so much porn that it can combine concepts like “big breasts” as easily as “a tree”. This leads to some hilarious mistakes (like adding multiple rows of breasts when it tries to combine “woman” and “big breasts”) but with basic masking and inpainting you can correct all that.
I’m all for banning AI models trained on child abuse imagery¹, but the assertion that AI generated child porn must be based on a model trained on existing illegal material is simply not true.
¹ unlike many AI enthusiasts, I consider AI models to contain significant parts of the training set, which is why I think models trained on illegal material shouldn’t be redistributed and generated AI artwork should be considered derivative works of the many images that are part of the original dataset. Same goes for ChatGPT.