Comment by coderenegade
Comment by coderenegade 2 days ago
I don't see why this is an issue? The prompts imply obvious and well-known characters, and don't make it clear that they want an original answer. Most humans would probably give you similar answers if you didn't add an additional qualifier like "not Indiana Jones". The only difference is that a human can't exactly reproduce the likeness of a famous character without significant time and effort.
The real issue here is that there's a whole host of implied context in human languages. On the one hand, we expect the machine to not spit out copyrighted or trademarked material, but on the other hand, there's a whole lot of cultural context and implied context that gets baked into these things during training.
I think the point is that for a lot of them there are endless possible alternatives to the character design, but it still generates one with the exact same design. Why can't, for example, the image of Tomb Raider have a different colored tank top? Why is she wearing a tank top and not a shirt? Why does she have to have a gun? Why is she a busty, attractive brunette? These are all things that could be different but the dominance of Lara Croft's image and strong association with the words "tomb raider" in popular culture clearly influences the model's output.