/3/ here. How hard would be to code an application that takes

Wow, you visit another internet board? Crazy.

>>58772150
If you have to ask you can't do it.

>generates a 3D model
>from static 2D images
AHAHAHAHAHA good luck with that.

>>58772181
I just entered college last week.

>>58772150
Harder than it would be to make the model yourself

>>58772150
Pretty hard mate. Less work just to 3D model the loli by hand.

>>58772150
It would be extremely painful

>>58772206
Your a big loli

>>58772201
>>58772203
I do plan on selling the software.

>>58772220
4u

>>58772150
Tbh senpai I have no experience in this area, but I can tell you now that won't be easy to make.

>>58772266
https://www.youtube.com/watch?v=Oie1ZXWceqM

It sounds like a good grad project.

>>58772233
Do you want it to only work for animu grills? Or also for recreating 3D grills/other 2D things?

>>58772297
my interest is anime shit, but I'm sure I could make millions if I make it general.

>>58772150
it'd be hard enough making one for photos of 3DPD, for anime it'd be pretty much impossible, since there's no consistent lighting or even positions of things like facial features

>>58772198
>learning babbys first java / C
>i want to code neural networks
s l o w d o w n

>>58772150
Very,

File: tasks.png (29KB, 267x448px) Image search: [Google]

29KB, 267x448px

>>58772198
>>58772327

this is why 2d > 3d

>>58772289
Neat.

>>58772150
Do cartoons really use that many angles. I'm thinking they might just have 10-20 different angle they can put characters at which cover every scenario. If that's true, it could be a barrier to hd grills

>>58772150
Do you have a clear, unambiguous, objective, way to grade the results?

Can you transfer that grading into code?

Do you have a way to translate the output of the NN into a model?

If you answered yes to all three, then it's easy to get some output. Hard to get usable output, and harder to get good output. Even harder to get consistently good output.

If you didn't answer yes to all three, it's impossible.

>>58772150
It's a great idea OP. But you're gonna have a hard time.

File: ai photos.jpg (119KB, 634x630px) Image search: [Google]

119KB, 634x630px

>>58772384
http://www.sciencedirect.com/science/article/pii/S228843001530049X
https://www.youtube.com/watch?v=Oie1ZXWceqM
https://github.com/andersbll/neural_artistic_style
http://stats.stackexchange.com/questions/214908/is-it-possible-to-train-neural-network-to-draw-picture-in-a-certain-style
http://waifu2x.udp.jp/
http://www.dailymail.co.uk/sciencetech/article-3346683/What-computers-think-HUMANS-look-like-Neural-network-learns-create-bizarre-pictures-people.html

The cost of computing power and time would far far outweigh the cost of getting your own personal 3D anime artist.

>>58772467
>not selling the software and becoming rich

>>58772493
Nobody would buy it. The cost to produce any acceptable output is massive, probably in the $10k to $20k range. Producing not only acceptable but good results every single time would cost you at least over a million. This is without a development team to produce the software into an end user product.

>>58772513
>not making a prototype and then selling the idea to autodesk

>>58772523
I can guarantee you AutoDesk will not even open the email.

"FREE 3D WAIFUS EVERYTIME GAURENTEED XD"

>>58772541
>what is venture capital

>>58772523
They will make 500x what you sold it to them for. You don't like taking it in the ass from corporations, do you?

>>58772551
comunist pls kys

I rather seek a patent than giving it to the looser homelex hobby OS comunity.

>>58772564
>Implying I'm communist
Communism is a plague the only creates hollow counties.

But, I also know when a corporation is scamming me. In the case of selling it, Autodesk is barely using lube.

>>58772150
You're asking if a computer can create a new dimension of detail when given a series of images based in a lower dimension. This would depend on if the images in the lower dimension were all consistent in design (e.g. eye placement relative to mouth is consistent throughout) and the images include pictures from all angles, and some topology. You could map it into a 3D space kind of like how people can map a chart of the earth into a globe.
But by the time you've done all that designing in 2D, you might as well have just made a model from scratch. Though maybe you could try to map shading on characters as topology, but it would be rough and I have a feeling many artists are not seriously thinking about depth when adding shade, and are rather trying to get cel-shading aesthetic or some other aesthetic. And getting the hairstyles mapped would be a nightmare. It wouldn't be reliable for training a NN with.

>>58772593
Or I get money, make a company, hire employees, then I'm fired by the board of CEOS and my company is hired by autodesk in the end.

there's no way to win.

>>58772619
Make sure you are the only one with access to the source code, other than devs who are on a restricted machine that can't connect to the Internet or USBs.

They can't fire you then because that will kill the company.

>>58772150
Photogrammetry is used to take 2d photos to create a 3d model. This is usually okay to create a quick and dirty model that needs clean up when more advance 3d scanning is unavailable.

However, one issue with what you want is that Anime does not use consistent models that are capable of existing in 3d space. So existing algorithms will fail.

Using NNs you could replicate the style of the anime such that it could "draw" what the character would look like at certain angles and poses, BUT learning AI needs an Input to map to an Output. In other words along with the photos of the anime character it needs control values that represent the correct answer. Without something representing whats correct, there's nothing to learn.


One option would be to create a standard bone structure and overlay it on every anime picture, providing the correct answer, and hope the NN can learn from that. BUT this would take a significant amount of work, and considering the volume of styles, constant retraining would be required.

So yeah it's possible to get something, but a human 3d modeling a character is FAR easier, faster and produce more reliable results.

>>58772150
Even hand-modelled anime renders look like shit. The industry has been trying very very hard to move in that direction because it would drastically cut costs, but it just looks like shit.

This is the best explanation I've found so far: https://www.quora.com/Why-does-the-latest-anime-animation-3D-look-so-weird/answer/Martin-Schneider-2?srid=mhun

>>58772446
that video is amazing

>>58772150
Your main problem is that 2d drawing, especially cartoons are often impossible geometries because parts are different shapes depending on the angle you view them. That's why its so hard to get 3d models always look stiff compared to animation. Disney actually came up for a solution to this a while ago I would start there.

File: Screenshot_20170201-194733.png (499KB, 1080x1920px) Image search: [Google]

499KB, 1080x1920px

>>58772187
http://www.wired.co.uk/article/jaw-dropping-3d-model-software

He doesn't have to do it from scratch. I'm sure there are tons of tools if OP could just learn to google

>>58772901
Oh I love that song.

>>58772150
Hey dude, try it out. That sounds like an interesting project, and you might get some neat results.

>>58772150
There is a crude way with a cracked version of a 3DSexvilla called The Klub17. In the game there is a tool that will generate a rough 3D face from 1-2 photos, then it can be further manipulated with sliders to perfect their face. Here are some of my attempts

>>58773016
this is a blue board, m8y

File: Emma-Roberts.png (525KB, 1440x900px) Image search: [Google]

525KB, 1440x900px

>>58773016
Anime is possible, though I haven't tried it myself, you can use the same in-game generator on picture of the characters face head on(should be easy to obtain) then apply it to the Hentai Models, since their head shape would better accommodate the wild proportions.

File: Melissa-George.png (671KB, 1440x900px) Image search: [Google]

671KB, 1440x900px

>>58773028
>>58773039
sorry about that, these two are at least sfw