|
internet celebrity posted:Latent3Visions just came out on the creators patreon last night and lets you use multiple prompts, image seeds, and lets you weight all the elements as much or as little as you want. This poo poo is getting crazy now. I keep running out of memory though and I'm not sure how to address that.  It's incredible to see this develop so fast in such short timeframes. There was a fork of l2v that could output generated images directly to google drive and it was incredible, hoping to see that added to l3vI'm not getting errors (yet) on a paid colab account connected to a Tesla, you might want to check what gpu you got and try again if it's not something with 16gb+
|
#
?
Apr 30, 2021 13:59
|
|
|
|
| # ? May 27, 2024 19:19 |
|
|
This might be a dumb question but who owns the art that gets created through these algorithms? Like if I wanted to sell t-shirts with one of the designs on them or something? You can imagine there being museum exhibitions of machine generated art where the human curator has an even more important role, just curious what the future will be like with this stuff.
|
|
#
?
Apr 30, 2021 14:24
|
|
|
A Strange Aeon posted:This might be a dumb question but who owns the art that gets created through these algorithms? Like if I wanted to sell t-shirts with one of the designs on them or something? Why do you even care? Machines are going to enslave us all and it is going to be glorious anyway!!!
|
|
#
?
Apr 30, 2021 14:30
|
|
|
that will most likely have to do with whoever created the code until some big name Artist decides to sell some images at a Christies auction
|
|
#
?
Apr 30, 2021 14:30
|
|
|
Like, entering a prompt doesn't really seem to be doing a lot of artistic heavy lifting, but if you're tweaking lots of parameters and curating the results, if you squint, that's sort of like mixing paints.
|
|
#
?
Apr 30, 2021 14:37
|
|
|
A Strange Aeon posted:This might be a dumb question but who owns the art that gets created through these algorithms? Like if I wanted to sell t-shirts with one of the designs on them or something? There's no legal precedent on it. Personally I look at the AI like a tool - it's a really advanced paintbrush. Generally copyright law will side with someone if their creative input was necessary for the creative work to be created. In this case since the art would not exist without you typing in a prompt (and now with l3v, multiple prompts, weighting them, selecting source images) - and even more, if you then collate several frames into a gif, your creative input was necessary. Toolmakers aren't able to claim ownership of products created with their tools, so I would expect the law to side the person generating the art. That being said it would be entirely within the dev's rights to stop you from continuing to use their tool to generate commercial art. deep dish peat moss fucked around with this message at 14:45 on Apr 30, 2021 |
|
#
?
Apr 30, 2021 14:39
|
|
|
A Strange Aeon posted:This might be a dumb question but who owns the art that gets created through these algorithms? Like if I wanted to sell t-shirts with one of the designs on them or something? the answer is whoever right clicks on the image and hits "save image". until that point is only exists in RAM and it is not a "fixed in permanent form". the act of saving it to your hard drive gives you the copyright (at least in Canada) quote:Fixation in Canadian copyright law is a threshold consideration that must be used in copyright infringement cases by courts to determine if copyright actually exists.
|
|
#
?
Apr 30, 2021 14:40
|
|
|
if you write a song on a keyboard the keyboard company doesn't own the song because their device generated the tones, you own the song
|
|
#
?
Apr 30, 2021 14:43
|
|
|
Also I can say from experience that it still takes certain finessing of prompts and understanding of the way that computers interpret language to create quality work. I've been at this since I first saw this thread and the vast majority of what I've generated has been total garbage. Like it's more nuanced than "type word and get art" Anyway here's an example of how it uses source images in L3V! Based on this drawing I did:  and asked to create an illustrated city skyline it gave me these:   And when asking for an apartment building based on this one:  I got these:   So it seems to not use the image conceptually but for colors/stylistic purposes and a few accents here and there. deep dish peat moss fucked around with this message at 16:05 on Apr 30, 2021 |
|
#
?
Apr 30, 2021 14:46
|
|
|
Rutibex posted:if you write a song on a keyboard the keyboard company doesn't own the song because their device generated the tones, you own the song But what if you write that song only using samples from the auto-play? I think the person typing the phrase has a lot of control and ultimately will be the owner. You can control what happens fairly easily by adding extra words (Haven't tried that new one out yet) but there is a learning curve and you can get the AI to do some impressive stuff. You can even rerun the same prompt for slightly different images sometimes depending on how much source material there is. This also makes me think this tool would be worth a lot or have strings attached such as royalties once it is refined.
|
|
#
?
Apr 30, 2021 15:34
|
|
|
pixaal posted:But what if you write that song only using samples from the auto-play? Wesley Willis still owns the copyright to Rock and Roll McDonald's  (Or, his estate does, I guess.)
|
|
#
?
Apr 30, 2021 15:58
|
|
|
Version 3 is a bit more experimental in later iterations. These are the last 6 frames after generating around 30 frames:   Based on this drawing:  Here's an example of how something can evolve over time with it (1st -> 5th -> 20th generation)  ---->  ---->  It seems to have some stitching issues compared to L2V and loses definition over time but that could be from using flat-shaded illustrated source images deep dish peat moss fucked around with this message at 17:24 on Apr 30, 2021 |
|
#
?
Apr 30, 2021 17:00
|
|
|
Version 3 with just text: "A fat guy feeding a bunch of fat raccoons hot dogs"
|
|
#
?
May 1, 2021 02:18
|
|
|
 the storming of normandy beach in the style of hieronymus bosch
|
|
#
?
May 1, 2021 03:49
|
|
|
Sydney Bottocks posted:"pope innocent x in the style of francis bacon" I think that has to be my favourite one so far, if only because it makes the AI kind of adorable.
|
|
#
?
May 1, 2021 03:56
|
|
|
text input is 'the most beautiful man in the universe' and 'disco' as the second prompt. the image was of ET 1:  2: 
|
|
#
?
May 1, 2021 04:00
|
|
|
Not Duck  Duckish  Duckest  Duck in a tree, i'm very boring Tafferling fucked around with this message at 13:30 on May 1, 2021 |
|
#
?
May 1, 2021 11:32
|
|
|
Jesus as a terminator endoskeleton
|
|
#
?
May 1, 2021 16:11
|
|
|
deep dish peat moss posted:
I joined on patreon and to someone interested at the edge of machine learning this is the rabbit hole to fall into. OpenAI is evaluating what the machine learned and the simple fact they discovered neurons that overlapped “textual name of character/pencil sketch of character/obvious image of character” about spiderman is absolutely nuts. Not that it’s possible, but that there is a recent blog post about figuring out why is amazing. Patreon is $5, but you’ll probably hit a wall if you’re a tech like me but don’t fully understand these collaborative python notebooks. They appear to kick rear end.
|
|
#
?
May 1, 2021 21:30
|
|
|
This thread could make for a nice photoshop phriday, to spruce up the front page after last year constant downer. They were my favourite things about SA.
|
|
#
?
May 2, 2021 11:13
|
|
|
hellotoothpaste posted:I joined on patreon and to someone interested at the edge of machine learning this is the rabbit hole to fall into. OpenAI is evaluating what the machine learned and the simple fact they discovered neurons that overlapped “textual name of character/pencil sketch of character/obvious image of character” about spiderman is absolutely nuts. Not that it’s possible, but that there is a recent blog post about figuring out why is amazing. I still find it hard to believe how good it is at interpreting meaning that's not explicitly defined in the prompt. "The assassination of Jabba the Hutt by the coward Leia Organa" picked up on the fact that the movie title the prompt is based on is a western and it gave Jabba a cowboy hat. It's dark, but "Lincoln's Last Stand" literally split Lincoln's head into two distinct pieces and then stuck a mysterious figure dressed similar to John Wilkes Booth in most photographs of him in between them.
|
|
#
?
May 2, 2021 18:24
|
|
|
There's a new book that asks you for additional input after every 100 iterations (when it spits out a picture). I never got 3 to run always got an error after 30 minutes of running and not even the noise frame. It slows it down but one of the prompts lets you draw to tell it where to focus. This is turning into a tool very rapidly where the person at the keyboard has a ton of input and guidance in the direction the art will take. e: I think the draw command lets you add noise to an area you want to fill it in solid not circle.  yeah this is going to be really fun to play with trying to make something along a single theme and saving frames to upload and use as starting points. tried both circling and filling, not sure it does what I think it does. Need to play around more currently taking about 15 minutes per frame so rather slow to test after the other image I lost the frame it was pretty much the same though  image after fill 
pixaal fucked around with this message at 19:04 on May 5, 2021 |
|
#
?
May 5, 2021 18:39
|
|
|
pixaal posted:There's a new book that asks you for additional input after every 100 iterations (when it spits out a picture). It looks like it's erasing what you draw over and then reconstructing what might be in there, while leaving the other pixels untouched. So maybe do the opposite, and only draw over what you want to change?
|
|
#
?
May 5, 2021 22:14
|
|
|
Fuschia tude posted:It looks like it's erasing what you draw over and then reconstructing what might be in there, while leaving the other pixels untouched. So maybe do the opposite, and only draw over what you want to change? There seems to be arithmetic involved with the neural output, I wonder if it’s ANDing or ORing with with the negative stimulus or something. Do I need to do anything special to use these Colab notebooks on a personal account or does Google just hand out diceroll Tesla GPU time these days? Another crazy thing.
|
|
#
?
May 6, 2021 12:21
|
|
|
Sunset Sleeping kittens  we do not talk about the top middle cat
|
|
#
?
May 17, 2021 21:08
|
|
|
pixaal posted:Sunset that is clearly a fetal sky bison
|
|
#
?
May 17, 2021 21:38
|
|
|
Bula Vinaka posted:Version 3 with just text: "A fat guy feeding a bunch of fat raccoons hot dogs" I feel like I can see that it took inspiration from the buttcrack dude
|
|
#
?
May 17, 2021 22:00
|
|
|
|
| # ? May 27, 2024 19:19 |
|
|
The Bloop posted:fetal sky bison Usernames thread is over that way 👉
|
|
#
?
May 17, 2021 23:09
|
|
