Comment by pppoe

1 year ago

This reminds me the trick to make recent text-to-image model generate highly realistic (but amateur) photos by adding "IMG_XXXX" into the prompt. Although these videos have nearly zero views on YouTube, they may be part of the training data behind these models.

It’s also the default naming for every digital SLR, phone camera, etc… lots of which upload with file name as title to Flickr and many other photo sharing services, most of which have also been used in training data.

DSC, IMG, etc etc.

Funny calling that a trick. These models 'generating' stuff is just recovering patterns compressed during the 'training'.