it's all getting quite odd now - image synthesis using adversarial networks

simply crazy developments

If this is a feature-request, keep dreaming.

There is no way a computer can accurately determine what image the person has in mind (even if the subject is correct, it will likely not be a fully accurate depiction of his imagination). It would get even worse if the genre was in something such as fantasy (as the computer might have no idea what it might look like, even with a detailed description).

Then there’s the fact that the resulting images tend to have artifacts and areas that lack definition (something that might require years of research yet), and don’t even try to mention the ideas of movies being made that way (how would Star Wars; Rogue One be replicated using just text input if a number of things do not even have a real-world reference for the computer to use?)

I put not spin as to useful or not, simply that it is.

I would like to see what image synthesis using adversarial networks algorithm for one of my fantasy creatures.Even if it were not perfect.I think it would be exciting.Maybe this could be used to make animated creatures from a person’s descriptions one day.I would love to see that even if it were not perfect.

The page itself seems a very interesting place

I don´t think that really matters. A person (artist) can not accurately determine what image another person (art director) has in mind either. The art directors ussually go with what they get or let the artist iterate towards the goal they imagine as long as the budget permits.
I dont´t think it is absurd to speculate that 3D artists get replaced by AI in the future. It´s going to take a while, though.

I see this as amazing and terrible,

we all will be able to do anything - and it will make us all worth nothing.

Unless scientists actually make algorithms that give computers an imagination, anything that creates an image based on text will appear as nothing more than the clever synthesis of multiple pieces of pre-existing imagery.

On a high level, I agree with Ace. You aren’t going to be able to write out a paragraph and get a beautiful scene in front of you.

on a lower level, I can see this working pretty well. Imagine more along the lines of “rusty metal with scratches and peeling blue paint”. The quality of automatic infill painting right now is astounding, so it isn’t hard to see how using similar synthesis techniques could be feasible for asset production.

if I can get this:

from this:

…with one button, that speaks a lot to the strength of imaging synthesis algorithms. Sure, there are some artifacts, but if there was a big action gunfight going in the the foreground, and this was just a background plate it’d be pretty hard to notice.

“David will I dream ?”…

(jusr the record so far no HAL7000 computer has ever failed)

I think using random seeds, and example geometry, one could procedurally texture an object by texturing a sphere, and painting a few places later to mix in other maps etc.

or a sphere that transitions into a cube etc*

have it observe patterns per ‘facet’ of the geometry, and the faces, transitions and overall etc, and use sliders to drive it?

for example. one could texture 3 buildings, and then randomly generate 10,000

But that is not the same thing that is in the youtube video.

fire burns in candles, and also in firepits and car engines and jet engines,

learning algorithms can be focused on gather low, med, high level details.

(see style transfer)

a learning algorithm can be paired with a adversarial network and a style generator to produce buildings that look like real ones,

one could further guide it with something like the text system

person (here is a thing I textured and modeled)

ai (here is geometry) other Ai(yep looks like a house)

other ai textures house, different ai judges it, yep looks housish

they make 200 you choose 5, and it makes 200 real looking models with realistic variations.

This is probably an inevitable extension of photogrammetry. All you need is a library to draw from. It’s not hard to imagine a future where everything except the curation of assets is automated. Not sure how I feel about that exactly.

I am unsure if you can objectively conclude that ‘imagination’ is anything but clever synthesis of multiple pre-existing imagery. The difference is that we have access to imagery outside of the internet.

GAN now generate images and then decide if they are ‘passable’

it’s not drawing on data sets other than to say, is that close enough to reality?

If you’re going to make a regular habit of reviving your own threads, it would help if it was by a post that was more substantial than a line or two of text.

Perhaps citing your source with a link or a video for instance.

“You can think of a GAN as the opposition of a counterfeiter and a cop in a game of cat and mouse, where the counterfeiter is learning to pass false notes, and the cop is learning to detect them. Both are dynamic; i.e. the cop is in training, too (to extend the analogy, maybe the central bank is flagging bills that slipped through), and each side comes to learn the other’s methods in a constant escalation.”