[ 3 / biz / cgl / ck / diy / fa / ic / jp / lit / sci / vr / vt ] [ index / top / reports ] [ become a patron ] [ status ]
2023-11: Warosu is now out of extended maintenance.

/ic/ - Artwork/Critique

Search:


View post   

>> No.6904350 [View]
File: 2.92 MB, 512x512, 1669006879174790.gif [View same] [iqdb] [saucenao] [google]
6904350

>>6904294
>"from its training data" that does not belong to the creator of the AI. if it did, no problem. why can't there be consent for how data is used?
i don't need your consent to use your images for photobashing
i do not need your consent to take your image, put it through a filter, trace over it digitally and doodle a moustache and some other stuff on it and posting that on instagram.

you know what these cases have in common? their need for consent depends on how transformative they are.
but in both cases i used data that didn't belong to me and fed it into photoshop.

now if these cases can work, how do you think your argument has any legs at all with AI?

>The fact that you can prompt AI and see some of the training data means we know it's not always fundamentally different.
that's overfitting. and it is neither desired nor the goal of the AI model. and they are tiny outliers that happen due to errors.
that right image in particular is actually a stock ad-photo where they would keep the same background and only change the phone skin that is being sold, so it's no wonder the AI overtrains on it.

>Building huh. maybe using the dna data it could one day?
one day? is already doing that. it is trying to see an image of your prompt inside the noise, and then gruadually builds on that with more detail and clarity.

>I don't really have an issue for dogs. natural objects with no human/creator are not really the concern here.
jesus anon...
you understand that it creates paintings of dogs in the same way that it creates photos of dogs, right?
in neither case is the AI using a real image as reference. it is building them from its own "knowledge" base

>> No.6870286 [View]
File: 2.92 MB, 512x512, 1691199168588584.gif [View same] [iqdb] [saucenao] [google]
6870286

>>6870177
i explained it already. countless, countless times.

it saves features and relationships of the images across the entire network. and those are tied to the tags we use.
i repeat: no images are tied to the tags, only the broken down elements are still there in the form of weights across the neural net.
and ON TOP OF THAT, those weights are further adjusted by other images. especially when using the same tags.
so how do you think the original image is still there? it simply isn't

and NO i don't mean
>this neuron is responsible for the cat's eye so it just steals one eye from the training image that is somehow still in the NN (it isn't)
instead it's
>a section/collection of neurons are doing the work TOGETHER to form a cat's eye, one that fits with the rest of the image, which is also created by collections of neurons, i.e. nodes.
there is a process to all this, and it has nothing to do with recalling pixels and stiching them together.
(that wouldn't even work because if you can paint you'd know that colors and values are conditional.)

instead what is happening is that an image is gradually shaped. you can even see this in action with every gif and webm i posted.

Navigation
View posts[+24][+48][+96]