I don’t think it comes from these Youtube videos – Flickr and other photo upload services are a more likely source of training images with default file names.

Maybe its a combination of both.

It seems exceedingly unlikely to me that frames from random YouTube videos would have been used to train image generation models. First off, they're difficult to extract and second, the quality of individual video frames is very low, especially if we're talking about 15 year old phone videos at what, 480p at the very best!

You are probably right. I approached it from a high-value dataset perspective but would agree that fuzzy frames probably don't help much.

Its not a rumor, you really do and you can try it out yourself. Unfortunately its very finnicky and you cant really leverage it to produce a realistic image of what you want since any further prompting seems to override it.

Its like a ghost in the machine prompt.

Makes you wonder if its possible with the right seed, scheduler, and prompt to complete recreate the original.