[a / b / c / d / e / f / g / gif / h / hr / k / m / o / p / r / s / t / u / v / vg / vr / w / wg] [i / ic] [r9k / s4s / vip / qa] [cm / hm / lgbt / y] [3 / aco / adv / an / asp / bant / biz / cgl / ck / co / diy / fa / fit / gd / hc / his / int / jp / lit / mlp / mu / n / news / out / po / pol / qst / sci / soc / sp / tg / toy / trv / tv / vp / wsg / wsr / x] [Settings] [Search] [Home]
Board
Settings Home
/sci/ - Science & Math


Thread archived.
You cannot reply anymore.



File: maxresdefault.jpg (77 KB, 1280x720)
77 KB
77 KB JPG
Not sure if this belongs in /g/ or /sci/ but anyway here it goes...

I'm interested in using artificial intelligence (AI) to generate videos based on a collection of existing videos. The program should be able to receive a playlist of videos as input and generate completely new videos that follow the same style and format as the existing videos.

For example, if you fed the program videos from the Unbox Therapy YouTube channel, it should be able to generate new videos in the style of Unbox Therapy. If you fed the program episodes of Cyanide and Happiness, it should be able to generate new episodes of Cyanide and Happiness based on the style and humor of the existing videos.

I'd imagine the AI would work in a similar way to human dreams in the sense that the videos that it generated wouldn't always make complete sense at first. An AI-generated episode of Unbox Therapy might feature nonsensical products for example. Or an AI-generated episode of Cyanide and Happiness might have an inconsistent plot. This is OK and as the AI continues to learn, I would expect such bugs to be reduced or eliminated.

I have a fairly basic to intermediate grasp on Python and Java programming as well as some knowledge of linear algebra but am completely clueless on things like machine learning and artificial neural networks or any of the advanced math that is (presumably) required for AI. Can anybody here give me some pointers on how I could go about attempting something like this?

Thanks in advance.
>>
>>10712995
>Java programming

Sorry, your brain is mutilated beyond repair
>>
>>10712995
Would require some advanced neural networks. LSTM, Convolutional Neural Nets, and maybe a GAN. I never figured out how to make pictures bigger than 32x32 pixels though.
>>
>>10712995
Not possible.
>>
>>10713003
I've dabbled a bit in C++ programming too.

>>10713006
>advanced neural networks. LSTM, Convolutional Neural Nets, and maybe a GAN
That... Sounds complex. I'm guessing this is all Masters/PhD level stuff?

>>10713039
If an AI can generate faces based on existing faces, music based on existing music, and paintings based on existing paintings then why not videos?

Heck, the human brain is a neural network itself. If you can close your eyes and imagine an episode of Cyanide and Happiness, then that's already an example of a NN generating a brand new video based on prior experiences of watching similar videos.

If biological neurons can do it, then why not artificial neurons?
>>
>>10712995
The cutting edge doesn't even come close to producing full-fledged videos like you describe
>>
This is bordering on the realm of science fiction.
>>
Is this the new low iq thread?
>>
>>10712995
>I would expect such bugs to be reduced or eliminated.
That is not what bug means, retard
>Java programming
You are so fucking stupid.
>>
>>10712995
No, sorry. This is graduate level work. You need a good grasp on computer programming first, then you can do some online machine learning courses
>>
>>10713068
>I'm guessing this is all Masters/PhD level stuff?
You don't need masters or phd to get an intuition about how they work. You also don't need to understand all the theoretical justifications of how they work to make use of them (especially considering theory is pretty far behind practice in AI anyway). But if you want to develop models that do new things having a theoretical background and a solid understanding of the field would probably help.
>>10713078
For something like cyanide and happiness you wouldn't need to generate video, just a frame by frame narrative that can then be animated using templates.

For full fledged arbitrary video you'd need to be able to extract a fuckload of meaning from the visual and audio data to determine what's happening, create a stream of frame descriptions and then have some sort of engine to render each frame by description.

But to have actual human level meaning you'd need to have human level background understanding.
>>
"Nonsensical products" so you want an AI capable of rendering 3D models that appear as if they are being realistically interacted with?

I'm not even sure that would possible.
>>
>>10713068
>an AI can generate faces based on existing faces, music based on existing music, and paintings based on existing paintings
Poorly and with a lot overfitting.


>the human brain is a neural network itself
No. Just no. The "biologically inspired" aspects of neural networks are so pathetically shallow it boggles my mind that people still think they are in any way shape or form similar to the actual brain.
>>
>>10713078
That's unfortunate. One of my favorite YouTubers (I won't name them here) disappeared a few months ago and I miss watching their videos. I managed to save all ~300 videos before they deleted their account and disappeared so I was curious to know if I could use them to generate new videos in the same style using an AI.

>>10713085
Self-driving cars used to be science-fiction at one point.

>>10713115
>That is not what bug means, retard
By that I meant that since NNs are trainable and self-learning, I'd expect any inconsistencies to be ironed out over time as the program learns and adapts from its mistakes.

>>10713366
That's what I suspected. My current plan is to learn all of this through websites like edX.

>>10713388
> But if you want to develop models that do new things having a theoretical background and a solid understanding of the field would probably help.
From the responses in this thread, it seems like what I'm trying to do has never really been attempted before. So it looks like I will probably have to study more. :(

>>10714338
>so you want an AI capable of rendering 3D models that appear as if they are being realistically interacted with?
Ideally, yes. In the example of Unbox Therapy videos, he should be unboxing them, turning them on, testing them, using the camera, etc. all the while making humorous comments throughout.

>>10714397
So far people have used AI to generate faces:

https://thispersondoesnotexist.com/

bodies:

https://in.mashable.com/tech/3226/say-what-ai-can-now-generate-whole-real-looking-bodies-of-people

voices:

https://www.theverge.com/2019/6/10/18659897/ai-voice-clone-bill-gates-facebook-melnet-speech-generation

scenes:

https://venturebeat.com/2019/04/07/promethean-ai-automatically-generates-game-scenes-like-a-bedroom-for-human-artists/

and screenplays:

https://en.wikipedia.org/wiki/Sunspring

But it seems that an AI that generates complete videos which incorporate all of the above elements hasn't been attempted yet.
>>
>>10715385
>That's what I suspected. My current plan is to learn all of this through websites like edX.
This is research level work. If it is ever possible, it would take you years to understand the technology behind it and to become efficient with it. After that it will take you years to develop/train your model. It will also cost you lots of money to train your overly complex model on a supercomputer(if you ever have access to one). (You probably need a team of scientists + HPC programmers.)
Either you are really interested in this topic and you can dedicate your life to this or just move on.
>>
>>10715550
edX has graduate level courses in AI for people who already have computer science degrees. Example:

https://www.edx.org/micromasters/columbiax-artificial-intelligence

The courses are 12 weeks long, free to audit, and count toward 7.5 of the 30 credits for a master's degree in computer science.
>>
Since you want to model time dependent statistical events you would want to use a probabilistic graphical model like an HMM, probably fully Bayesian. Sounds like a pipe dream anon
>>
>>10712995
I'm just an mechanical engineer with minimal programming experience but I've been watching this guys channel https://www.youtube.com/user/keeroyz/videos

I think you'd need to use multiple algorithms, like there's AI that generates scripts compiled from all the scripts of sitcoms, you use that first then use an algorithm that can synthesize the voice to read the script, then make generated scene from unbox therapy and finally some algorithm that can lip sync the script effectively and superimpose deepfake style over the generated videos lips/face.
>>
File: Somera-chan.jpg (46 KB, 320x180)
46 KB
46 KB JPG
>>10712995
And if you fed it 4koma anime shorts it'd be able to generate full-length anime episodes?
>>
File: 1560318556621.gif (319 KB, 600x500)
319 KB
319 KB GIF
Artificial intelligence is quite powerful nowadays

I wish you well on your projects OP!
>>
File: IMG_20170511_34534.jpg (50 KB, 492x621)
50 KB
50 KB JPG
>>10712995
This thread is so fucking retarded, and you know nothing about AI

do you think that AI is some kind of ACTUAL intelligent machine that can create faces and phrases ...

supposedly you want to create a program that uses AI to generate a video, you need at least 1000 of the best programmers in the world and hundreds of terrabites of informations and 30 years of work
>>
File: face.jpg (231 KB, 1317x796)
231 KB
231 KB JPG
>>10718409
>do you think that AI is some kind of ACTUAL intelligent machine that can create faces and phrases
That has actually been done.
https://www.youtube.com/watch?v=-cOYwZ2XcAc
>>
>>10713068
Biological neural nets != artificial neural nets

ANN's are poor approximations of real neurons

Brain works on fuzzy logic all the time, gans and all those algorithms to generate content are human inventions, not discoveries. No one really understands how the brain process things.
>>
>>10715385
You want to study neural networks because a YouTuber channel went down?

Jesus Christ what the fuck this is some top level cancer
>>
The absolute state of /sci/
at least its better than the iq threads
>>
>>10712995
How I would do it:

Video -> normalized format -> split into frames -> split into pixels -> into GAN network. The tricky part would be to make it sequential, to generate a movie. Not sure how to do that, unsure if it's even been done. And even if it's possible, the amount of hyperparameters would be insane, to get it to work you'd need a lot of experience in AI.
>>
>>10718432
You know that feeling when someone post a reply so fucking stupid you don't even know how to start?

well I'm having this feeling now, how many faces that AI created? and how many frames of it moving? and how many phrases it created ??

it's like we're talking about a human teleportation and you post source about a teleportation of a photon and say "That has actually been done"
>>
>>10718972
Aww cute, a little Dunning trying to appear like a big Kruger, hello little wubbly do, what you wanna do? wanna go out for a walk? Good boy!

They're already starting to use algorithms in games to render scenes, ohh and the end of this video has a completely generated video, pic related, it's a still of a generated video on the right and the scene it was using as a framework on the left.

https://www.youtube.com/watch?v=ayPqjPekn7g

Seems like your photons are really bosons all along and you're too limited to know the difference.

I don't think you grasp how this technology works, like... at all. *sad face*

Talking to you is fun, like playing with a down syndrome child because you need some fun in your retarded life, only occasionally you shit yourself and everyone laughs.
>>
File: tenor.gif (82 KB, 220x205)
82 KB
82 KB GIF
>>10718972
Anon's right you totally shat yourself, what a faggot.
>>
>go to write out list of every single ridiculously complex thing this fucking retard would either have to implement or literally invent from scratch and do better than multi-billion dollar companies
>get bored after the 5th item
>>
>>10719083
Hey you should probably watch videos before you use them as an argument. This isn't generating video, it's filling out a coloring book (poorly, I would add).

>the end of this video has a completely generated video
You don't know that. They don't say how that video was generated.

Deep learning is smoke and mirrors, get over yourself.
>>
>This thread is so fucking retarded, and you know nothing about AI

>you think that AI is some kind of ACTUAL intelligent machine that can create faces and phrases
Yes, you are wrong. (not actual intelligent though)
>you want to create a program that uses AI to generate a video
>30 years of work
Yes, you are wrong
>ny faces that AI created? and how many frames of it moving? and how many phrases it created ??
There's hundreds of videos and papers on this subject, constantly being updated with more progress, do you know how to google?
You need me to feed you information you pretend to know like a big mommy bird?
>when someone post a reply so fucking stupid you don't even know how to start?
Totally know that feel dude, real talk
>>10719633
>Hey you should probably watch videos before you use them as an argument.it's filling out a coloring book
And you should thoroughly read posts before posting, I said they rendered scenes, but that part wasn't the important part and you know that.
>You don't know that. They don't say how that video was generated
They trained an neural net on people dancing gangnam style and generated the same moves on a scanned avatar of one of the researchers and placed him dancing in a random scene, that's exactly how this technology works and the video was pretty realistic.

>get over yourself.
That's a cute, have a nice day anonchan.
>>
File: pose.webm (2.92 MB, 2048x485)
2.92 MB
2.92 MB WEBM
>>10719813
>They trained an neural net on people dancing gangnam style and generated the same moves on a scanned avatar of one of the researchers and placed him dancing in a random scene, that's exactly how this technology works and the video was pretty realistic.

a.) You're just making this up since the video didn't say how it was made. This isn't good practice, especially when you know as little about this technology as apparently you do.

b.) That's wrong. The actual details are in the paper https://arxiv.org/pdf/1808.06601.pdf and it's computed in virtually the same way as the driving footage except that the "high-level layout" in this case is pose information extracted by other off-the-shelf tools and additional smoke and mirrors are employed by just outputting the optical flow of some initial input frame, as well as a few other tricks that guarantee overfitting but make for a semi-convincing tech demo. Then it's run on a giant supercomputer wasting god know's how much electricity and the results are mediocre and totally useless. What a waste of time and effort.
>>
>>10719083
>>10719118
fucking CS cunts
>>
File: 1479120633914.jpg (554 KB, 1500x1000)
554 KB
554 KB JPG
>>10719083
This has to be the all-time most angry --and the most autistic-- post on 4chan. I've visited this site daily for almost 8 years and never seen anything like this.
>>
>>10719083
that doesn’t make it an intelligent machine
>>
>>10721817
AI is by definition intelligent
>>
>>10719083
>I don't think you grasp how this technology works, like... at all. *sad face*
cringe, you came from reddit, didn't you?



Delete Post: [File Only] Style:
[Disable Mobile View / Use Desktop Site]

[Enable Mobile View / Use Mobile Site]

All trademarks and copyrights on this page are owned by their respective parties. Images uploaded are the responsibility of the Poster. Comments are owned by the Poster.