It’s a bit different for MidjourneyV6, previous AI models would create their own original images based on patterns learned from the data. MidjourneyV6 on the other side reproduces the original images to such a degree where they look identical to the originals for the average observer, you have to see them side by side to even spot the differences at all. DALLE3 has that problem as well, but to a much lesser degree.

That means there is something going wrong in the training, e.g. some images end up being duplicated so often in the training data that the AI remembers them completely. Normally that should be reduced or avoided by filtering out duplicate images, but that seems to not be happening or the images slip through due to small changes (e.g. size or crop will be different on different websites).

Note this doesn’t just impact exact duplication, it also impacts remixing, e.g. when you tell it to draw Joker doing some task, you’ll get Joaquin Phoenix’s Arthur Fleck, not some random guy with clown features.

All of this happens with very simple prompts that do not contain all those very specific details.

In AI’s defense: All the examples I have seen so far are from press releases of movie stills. So they naturally end up getting copied all over the place and claiming copyright violation for your own material that you released to be reused by the press wouldn’t fly either. But either way, Midjourney is still misbehaving here and needs to be fixed.

More broadly speaking, I think it would be a good time to move away training those AI almost exclusively on images and start training them on video. Not just to be able to reproduce video, but so that the AI get a more holistic understanding of how the world works. At the moment all its knowledge is based on deliberate photo moments and there are very large gaps in its understanding.