How hard would be to code an application that takes thousands of anime pictures, train a NN on it and generates a 3D model based on it?
sound pretty hard
http://www.sciencedirect.com/science/article/pii/S228843001530049X[YouTube] 3-Sweep: Extracting Editable Objects from a Single Photo, SIGGRAPH ASIA 2013 (embed)https://github.com/andersbll/neural_artistic_stylehttp://stats.stackexchange.com/questions/214908/is-it-possible-to-train-neural-network-to-draw-picture-in-a-certain-stylehttp://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
real photos, sure whatever, but anime? the model is all over the place with those
>>550022You can't train a neural network without giving it a solution. Meaning you can't feed it "thousands of anime pictures" unless you have thousands of completed models that correspond to the pictures.
>>550022Easy, but doing it with any sort of efficiency would be really hard.
>>550024This is fucking with my head
>>550024>2013>Nothing in 2017Seems like some corporation bought the tech and failed to use it in public software.
>>550022>tfw absolutely no Zone tan model avaiable for public so i can use it for porn
>>550159The one in OP was made by leslyzerosix to be 3D printed, it ain't gonna animate no good.Quick look reveals that this is the best Zone-Tan model available http://anfrien.deviantart.com/art/Zone-Tan-Original-3D-Download-611351390Though I think she still has a weird skull nose, but you could fix that.
>>550159It'd take like two or three hours to cobble shit together from SFMlab and model the hair.
Am I the only one who isn't a retard?All you have to do is look at the background of the "picture created models". They are different images.
>>550217You mean the 3 Sweep vid? We've had content-aware fill in Photoshop and GIMP for years, the ability to separate an object that a human selects isn't unfeasible, or even difficult with simple backgrounds like the faucet.And if they were different images, it's still making a model from an image.
>>550024I could see this speeding up the workflow for making basic meshes.
>>550022training convolutional deep networks takes a lot of processing power, one of the reasons only brands like google, ibm, and microsoft (plus unis) do largescale image recognition researchi'd say the hard part would be making sure a combination of heuristics (pro tips from human artists) and machine learning (the nn) can work together to make sure a normal human model does not violate certain rules like having only one head, maintaining human proportion while also creating sensible variety in the models etc.an extra layer of complexity would be emulating specific artistic styles (anime, blizzard's heroic scale, disney infinity's style, etc.)
>>550022to give you an idea, no one on earth has managed to do it yet.
>>550541the chinese government got secret 3d anime technology they waiting to distribute to topple the american economy once they're self-sufficient
>>550022I do image recognition stuff with pycaffe if you want to kik me kazzy1 or skype kaz_coyote. I could help with in general stuff, but the more specific 3d stuff, not too much
>>550530When you say "blizzard's heroic scale" is that something that is documented or formalized in any way? Googling that phase dose not seem to reveal anything.
>>550024Is this real?Seems way too advanced to not be fake.There's way too many assumptions that would have to be made by the software to make this possible, which are all being made correctly in this video.
lets just say it would be cheaper to hire an artist
>>551326anon it traces boxes and cylinders by following the cursor and snapping to detected edges, it isn't making any sort of complex greebled or even beveled shapepick up photoshop or gimp, we've had edge detection and edge selection tech for yearswe have technology that will detect faces and morph them into smiles using an approximation of a human brain and it's given out for free in a chinese app for iphones and you think we can't trace cylinders over an image?
>>551326SIGGRAPH videos are supposed to look insanely cool and not tell you the flaws. (All academic research works this way.) You should take a skeptical eye. "Show me the money." I want to see a Maya plugin or it doesn't actually exist.So here's what a skeptical perspective means, just as an example of how to respond to a video like that. Yes the video is real. But it came out in 2013 and there seems to be no commercial implementation. They probably couldn't get it to do much more than produce simple boxes and faucets.
>>550022This is something I've been wondering about, too. I think it may be in some parts possible. But you won't get a perfect 3D model. It will be some kind of tool that speeds up your workflow in some way. A human will always be needed to fine tune the result.Deep learning in its current form is rather new and there is a lot of research going on what the limits are. Especially generative models seldom produce acceptable output at the moment. And that is for images. 3D models might be even harder to produce.I am still astounded that something like http://paintschainer.preferred.tech/ works that good even if not comparable to a real artist's work. But things like this that don't require much creativity may be automatable. I think for 3D there are some boring tasks that may become easier with deep learning, too. >>550530I hope GPU accelerated cloud instances (I think Microsoft Azure and Google offer these by now) will make it possible to develop such applications even if you don't have a farm of high end computers in your basement.