AI images
- Thread starter pwe
- Start date
Tuco
I got Tuco'd!
- 45,446
- 73,524
No real idea. Most neural network stuff operates on 2d grids ( matrices, tensors etc) that make it amenable to image-based algorithms. A lot of the work with non-image algorithms involves first putting the data into an image so it can run in a convolutional neural network of some kind. Like Kharzette says the audio can be put into waveform images.
One problem is how sensitive humans are to rhythm and harmonics, so I'm guessing you'll have a problem where offbeat, offtune sounds are really irritating to human listeners but totally fine for whatever optimization process is used to create them. This is similar to how even really good NN portraits of humans kind of make you nauseous and their eyes look weird, but hand-drawn portraits from hyper realistic to even very stylized ones, convey human emotions without looking gross. Meanwhile, landscape pictures look very pleasant and interesting because we don't have this evolved trait of wanting to avoid genetically divergent people.
- 2
Edaw
Parody
- 12,335
- 79,155
I guess 7 notes vs 3 primary colors makes the math much harder to build algorithms for. Makes sense. I can wait.
Tuco
I got Tuco'd!
- 45,446
- 73,524
I skimmed this article on it and enjoyed reading about the current state of the art for AI music.
Generating Music using Deep Learning
Introducing a new VAE-based architecture to generate novel musical samples
towardsdatascience.com
No idea if this is actually AI-generated or total bullshit, but it's enjoyable so I'm posting it. It sounds too good to be "real" AI-generated music.
- 4
Edaw
Parody
- 12,335
- 79,155
Good read. It is interesting that they point out one of the challenges is that visual AI represents a snapshot in time, where music has a temporal component that must be accounted for. We will have procedural VR porn within 20 years. Buy stock in Viagra.I skimmed this article on it and enjoyed reading about the current state of the art for AI music.
Generating Music using Deep Learning
Introducing a new VAE-based architecture to generate novel musical samples
No idea if this is actually AI-generated or total bullshit, but it's enjoyable so I'm posting it. It sounds too good to be "real" AI-generated music.
- 31,998
- 125,504
Bandwagon
Kolohe
- 22,803
- 59,822
P
pwe
or anyone else that doesn't have limited tries on this....would you mind trying a few things out for me? feel free to add whatever modifiers/descriptors/whatever you want, I'm just wondering how well this thing works with map-related stuff.
Swiss style hillshade relief map, mountainous
Nautical map with sea monsters, Jo Mora
Nautical map showing leviathan and ship wreck
Swiss style hillshade relief map, mountainous
Nautical map with sea monsters, Jo Mora
Nautical map showing leviathan and ship wreck
- 1
Cybsled
Avatar of War Slayer
- 16,500
- 12,156
I want to try this with something really obscure to see what it pulls - I’m presuming this is pulling various artworks and redoing them/combining them. Would be easier to test this theory with something you know very little art exists of it or if it does, it will be easy to identify
I didn't find anything good looking for the first one, but the second one looks pretty badass (even if probably not the style you wanted, Jo Mora).
- 10
A slight tweaking of keywords (Nautical map sea monsters in the style of Jo Mora colorful simple fantasy 4k) and it goes full LSD.
- 8
