/ic/ - Artwork/Critique

>>6870339
All compression algorithms get out something that's completely different from the training data at the data level - it's only at the level that we care about that we make sure it retains the structure we're trying to preserve. All a NN does in its training process is try to match the training data. The result, as you might guess, is something that matches the training data, and the only question is 'at which level'. As you've already acknowledged, overfitting is a thing - if an image is in the dataset more than a few times, the NN will learn how to replicate that image specifically. What you don't seem to see is that this isn't actually any different at the 'technical' level from the kind of fitting we do want - the training process works in the exact same way. The difference is that when you have 100000 different images, the thing that's being compressed isn't any one of those images individually, but all of them together. what it gets out isn't something 'completely different' from the training data. You can do interesting things with the resulting NN, like tell it to generate a catgirl in the style of Rembrandt and have it hallucinate something that matches those datapoints, but that's all it is - combining datapoints within the model.
This is not the case with other artists, whose work is never just an average of other peoples work at any point on their journey, and who are figuring out their own process to produce the results *they* want to produce. Like, if you don't understand that this is a fundamental difference I don't know what to say.

>you don't own these things that the AI trains on. none of it. only your style as a whole, but not the parts of it.
The way I see it I absolutely own all of it - every single decision I made on my artistic journey, every experiment and technical exercise is part of the process that culminated in the way I draw now. That is my 'intellectual work' as an artist, not any one illustration.