Training data at scale unavoidably taints models with vast amounts of references to the same widespread ideas that appear repeatedly in said data, so because the model has "seen" probably millions of photos of Indiana Jones, if you ask for an image of an archeologist who wears a hat and uses a whip, it's weighted averages are going to lead it to create something extremely similar to Indiana Jones because it has seen Indiana Jones so much. Disintegrating IP into trillions of pieces and then responding to an instruction to create it with something so close to the IP as to barely be distinguishable is still infringement.
The flip-side to that is the truly "original" images where no overt references are present all look kinda similar. If you run vague enough prompts to get something new that won't land you in hot water, you end up with a sort of stock-photo adjacent looking image where the lighting doesn't make sense and is completely unmotivated, the framing is strange, and everything has this over-smoothed, over-tuned "magazine copy editor doesn't understand the concept of restraint" look.
The flip-side to that is the truly "original" images where no overt references are present all look kinda similar. If you run vague enough prompts to get something new that won't land you in hot water, you end up with a sort of stock-photo adjacent looking image where the lighting doesn't make sense and is completely unmotivated, the framing is strange, and everything has this over-smoothed, over-tuned "magazine copy editor doesn't understand the concept of restraint" look.