> again trained (without permission) on copyrighted work
So far there is no solid proof of that. They didn't disclose the sources or the methodology. Except for 'trained' and 'copyrighted' the rest is questionable. Otherwise they would be already paying royalties.
They could have used the output from prev version 2 with prompts generated by GPT, and then corrected by humans based on the produced image. Also they could use CV to analyze new/old images. I.e. if there is a new feature in the image add it to prompt and train again.
So far there is no solid proof of that. They didn't disclose the sources or the methodology. Except for 'trained' and 'copyrighted' the rest is questionable. Otherwise they would be already paying royalties.
They could have used the output from prev version 2 with prompts generated by GPT, and then corrected by humans based on the produced image. Also they could use CV to analyze new/old images. I.e. if there is a new feature in the image add it to prompt and train again.