Thread replies: 11
Thread images: 2
Thread images: 2
File: Vp3g9pYM.jpg (390KB, 1536x1536px) Image search:
[Google]
390KB, 1536x1536px
Hey - can i get some feedback on this?
https://github.com/robbiebarrat/art-DCGAN
I just wrote it - It's a modified implementation of deep convolutional generative adversarial networks with a focus on creating artwork.
Basically, I've written scripts that get huge datasets of artwork, and then train a GAN on it. The GAN is a collection of 2 neural nets (a generator and discriminator) that use each other to get better and better at
>generating things that look like they're from the dataset
and
>discriminating between real dataset images and generated ones
over time, it'll learn to paint whatever type of art you feed it, and if you scroll down in the readme you can see results. (pic related is after it looked at 14,000 landscape paintings).
Please help me out with what I could do to make it more interesting - and just feedback in general - literally anything is appreciated. I wanted to write some stuff to accompany it (maybe some stuff about machine creativity or whether or not what the network generates is "art", but i dont want to sound like a goddamn loon...
Thoughts?
>>
>>61911720
>its an anon makes art libs even more worthless episode
god speed
>>
>>61911720
which dataset are you using?
>>
>>61911720
try cubism, or a specific artist like david hockney
>>
>>61911776
in the repo i put the script that i wrote - it scrapes the "paintings by genre" page of wikiart
>>
>>61911782
i'm working on writing a scraper for individual artists - but i have to work on this style-transfer-esque thing because each artist doesn't have nearly enough paintings to train a GAN...
but thank you I'll try it out in the next few days.
>>
>>61911732
this is what i'm excited about :^)
jk
i am looking forward to not only the new types of art that results from advances in this area of machine learning - but also the new types of artists. if you write machine learning models - you're an artist. a new type of artist, like a sculptor almost. I'd argue that making ML models is much closer to some sort of art than it is to engineering.
>>
This is really cool..
Do you think you can make this work for videos as well?
>>
>>61912728
Thanks! If you like it, please star it on github.
due to the nature of GANs, it probably wouldn't work for what you're thinking.
I'll add support for animated gifs of stepping through the space of samples later (basically it'll just pick two random points in the space of landscapes, and then animate a smooth transition of all of the landscapes between these points - so you might see a desert morph into a forest, but first turn into scrub-land or a field or something - just smoothly.
>>
it looks like a photo taken by an old 0.3Mp dumbphone camera, "oil painting" filter applied and reencoded as highly compressed jpeg
>>
>>61911720
So, you just slapped to a DCGAN some art pieces? How much did you change the implementation?
Thread posts: 11
Thread images: 2
Thread images: 2