[ 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.7041875 [View]
File: 169 KB, 1121x725, Screenshot 2023-10-04 173415a_proc.jpg [View same] [iqdb] [saucenao] [google]
7041875

>>7041782
>So AI basically just remembers the color and position of every pixel of the image then creates pattern out of it and "hallucinates" images through noise?
>So it's not a copy and paste of an image. It's a trillion copy and pastes of trillion different images with a little amount of random seed generation on top so brainless retards like you can do mental gymnastics?
also have you ever though this through? next time you spout this dumb shit, think it through, yeah?

i'll help you through some of it:
>So AI basically just remembers the color and position of every pixel of the image then creates pattern out of it
what do you call this? what do you think a "PATTERN" is?
a shape is a pattern. a line is a pattern, angles are patterns.
colors have patterns, values are patterns.
the way light interacts with things, that's patterns
and these are just the simple things that we can put to words.

you know what these are not? PIXELS. because they are concepts that go beyond just pixels. they are things that can be adjusted. any curve can have a different arch.
the other thing here is: do you think any artist can take claim on these kind of things? these are things we take from each other all the time. we literally study each other to gain new insights into how things can be done.

what the fuck do you think
>It's a trillion copy and pastes of trillion different images
actually is supposed to work? have you EVER truly thought this through?
maybe try to. or maybe all of these are beyond you. i can see that being the case if you are a permabeg. in which case, shut the fuck up and let the adults talk.

>> No.6956261 [View]
File: 169 KB, 1121x725, 1677548010715349.jpg [View same] [iqdb] [saucenao] [google]
6956261

>>6956226
>What's your point? That machine learning is.... Learning? Mind blown.
no, that learning is not stealing. yes. mind blown, apparently.

>The reason why I mentioned AGI is that the things you claim image models are "totally doing" like understanding the concepts and things they generate like humans do would actually require AGI.
no. i call it "understanding", but what it actually has are just representations of tokens. it's not wrong to call those "understanding". in an LLM you'll find these tokens include relationships with other tokens and in image gen a token will include the generalized representation of what it has learned about that token in the training data. in short: a sense of understanding of sorts.
if the token is "dog" the the token will have a representation of the features and relationships that make up "a dog" in general.
and this is something it has learned from seeing patterns across a large sets of dog images. not from individual dog images.

researchers will use words like understanding all the time. because that's essentially what it is, yes, even without a shred of sentience. because that's just what these tokens are. a representation of a concept or idea.

cont.

>> No.6935338 [View]
File: 169 KB, 1121x725, 1693259618107690.jpg [View same] [iqdb] [saucenao] [google]
6935338

>>6935323
it's called overfitting and it's an undesireable error.
that's not what AI is supposed to be doing nor is it what AI is normally doing in everyday use.
a ton of duplicates (among other conditions) in the training data can lead to that. that's why memes and more famous photos that get shared around in articles often get overfit.

pic related still applies. imagine instead of learning on a bunch of different smileys, it just trains on the exact same smiley over and over.
it's akin to learning the entire composition as a whole, almost like a symbol.

>> No.6897261 [View]
File: 169 KB, 1121x725, 1694020237675801.jpg [View same] [iqdb] [saucenao] [google]
6897261

>>6897246
>but it doesn't work without scraping your data
you don't understand the process at all. and this is the only way you can describe it.
because you're a fucking idiot. :)

it looks at the data. it tries to copy it, if it gets it wrong it adjusts itself.
it does that for every image.
it learns to do it more effectively the more images it sees, so it learns what all the images with the same token have in common.
it learns features and relationships.

that is what it takes from its training data. this is how AI works.

and this is again just a simplified explanation for you toddlers. writing any more is futile anyway.

ask me and i'll explain more in the evening if anyone is really curious.

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