/ic/ - Artwork/Critique

>>6956585
yes, everything is trained for denoising images. but denoising images is essentially the same as creating an image out of noise. that is the entire trick to diffusion models.

but i genuinely am not sure if i should take you seriously after this. you can't be serious with the linear transformation shit.

yes, a ruler can draw you a straight line more reliably than the AI could most of the time. but the AI isn't using a ruler. even when it makes a straight line it is doing so using patterns and relationships it learned from its training data.
same when making reflections. the model is not using linear transformations when making the reflection. so what is your point even?

and you can't use linear transformations to do anything remotely advanced. not even the reflection in pic related, much less images with mirror reflections that aren't just an upside down flip.
again, are you kidding me? i think i'll stop here.

>>6956610
>>6956609
>>6956601
>>6956593
are the tears coming out? better try harder, otherwise i'll bring in the AI dystopia :)
(this is sarcasm)

smileys is just the simple example.
another example again being reflections.
what reflections show us is ALWAYS dependant on the rest of the image. and AI learned this.
it is not even possible to make this by copying since it's context dependant.
>>6885786 is also an example.