AARD VARKMAN posted:"an ancient civilization constructing an enormous cheeseburger" somebody do all the ancient gods of the past. You know the ones everyone forgot, bronze age idols, 10000 bc, that kinda poo poo. Crom. Humiierc. Bokrug. Yog.
|
|
# ? Jul 29, 2022 06:23 |
|
|
# ? May 29, 2024 20:31 |
Got an invite this morning, already very impressed Impressionist painting of a quokka eating Avocado on a tropical island A llama wizard drinking mate and looking into his crystal ball in the Amazon jungle A red squirrel wizard reading a book and smoking a pipe in the black forest. Yes my wife heavily influenced the prompts E: one more j smoking squirrel wizard Son of Rodney fucked around with this message at 08:25 on Jul 29, 2022 |
|
# ? Jul 29, 2022 07:20 |
|
This one kind of blew me away, even if it's possibly random to some degree, cause the other results were poo poo. I was loving around with horror words trying to see if I could get some ~spooky~ images and one of my prompts was “a vintage, damaged photo of a ten foot tall demon crawling in a dark graveyard” which is about cheesy as it gets but whatever, and one of the results was this. I just think it's wild that if you look at it closely, the AI clearly generated this demon image in center/middle in the shadows. You can see the face, horns, body etc. and then perfectly blended it into the light parts of the image on the right, which form some sort of horrific looking rat-man face. But then it perfectly meshed all of it together to just look like a bunch of trees at a glance.Then completely on it's own accord, Dall-e decided to add in what appears to be a kid/woman impaled on a grave marker. It's like one of those spooky images that people over analyze but the spooky poo poo is actually all there because I told it to be ,and also Dall E threw in some bonus wtf. veni veni veni fucked around with this message at 09:18 on Jul 29, 2022 |
# ? Jul 29, 2022 09:13 |
|
veni veni veni posted:This one kind of blew me away, even if it's possibly random to some degree, cause the other results were poo poo. I was loving around with horror words trying to see if I could get some ~spooky~ images and one of my prompts was “a vintage, damaged photo of a ten foot tall demon crawling in a dark graveyard” which is about cheesy as it gets but whatever, and one of the results was this. im not seeing this demon. are you sure the demon isnt within yourself
|
# ? Jul 29, 2022 10:05 |
|
Yeah turn on your monitor
|
# ? Jul 29, 2022 10:20 |
|
might be a question of (bad?) monitor calibration. when I turned up the gamma I can sort see a demon face or hints of it. there's a reddish glow around the eyes.
|
# ? Jul 29, 2022 10:27 |
|
My tracings make it look ridiculous, but on the right look for the circular area that looks like an eye I feel like this isn't even subtle as the left half of a face The middle part might depend more on your monitor but I can see two eyes, horns, nose, mouth. Less obvious but the shape seems pretty clear and pretty much lines up with what I prompted it with Maybe it's just how my weird brain works but I took note of that poo poo immediately Look at the actual pic tho not my crayon drawing over it.
|
# ? Jul 29, 2022 10:48 |
|
I'm afraid you might have an acute case of pareidolia
|
# ? Jul 29, 2022 10:49 |
|
Atoramos posted:
The best part here is that the top row three are muppets who look like Hayao Miyazaki himself
|
# ? Jul 29, 2022 11:37 |
|
NNs generating an image composition at the original resolution and then losing the plotin the upres/detailing steps is not uncommon. Midjoirney does that hella more often though, which is why you end up with chestfaces on so much of its stuff
|
# ? Jul 29, 2022 12:34 |
|
Craiyon: I developed it myself. Yessiree, this is my ticket upstairs. Everyone: *stares blankly* Craiyon: You know... for kids.
|
# ? Jul 29, 2022 16:23 |
|
Atoramos posted:Funny, the original one was actually 20% larger but imgur won't host it. Honestly, I disagree with the comment for two reasons. First off, yea you could just render this at a different scale (and use different prompts. Mine were specifically targeted at architecture and not little creatures and such). Secondly, while this is a weakness of Dall E, Disco Diffusion doesn't share this issue. Consider the 'infinite-resolution' images that allow for the animation I made here: I've been playing with DISCO this morning. Shows promise. Better than Mini making it the best free project (you do need a free Google Notebook and roughly 8GB of free space for cache on your Google Drive) 0: ["GIlbert Gottfried as The Joker, eating tacos, 4K, HDR"], Missed the joker theming mostly but there's hints of it I have no idea what is going on with that tiny face by the styrofoam cup. And if this looks remotely familiar I always use Gilbert Gottfried as the joker doing something to test an AI been doing it since before he died. Just really wanted to see him play the Joker with that laugh. Can't wait to see later iterations, I'll keep playing with it but it's not MJ or DALLE2 yet not a hell of a lot worth sharing in my few samples. I got assigned a T4 today so it's 30 minutes per 500 step render.
|
# ? Jul 29, 2022 16:27 |
|
I just got the invite to the big boy Dall-E Siouxsie Sioux in the style of Yoshitaka Amano Peewee Herman playing the role of the T-1000 Jason Statham eating 4 giant cookies at a Chili's, style of 1950s print ad
|
# ? Jul 29, 2022 16:42 |
pixaal posted:I've been playing with DISCO this morning. Shows promise. Better than Mini making it the best free project (you do need a free Google Notebook and roughly 8GB of free space for cache on your Google Drive) For $10 a month you can get Collab Pro which will basically never kick you off for overusage (only for long idle times). Honestly, I've found unique benefits to each of the models I've played with. I have access to Midjourney but haven't tried it out. Dall-E 2 is good at: -Interpreting a single-weighted input and providing a solid result -Taking in portions of an image and reinterpreting the missing areas (the edit functionality is real good) -Interpreting an input image and coming up with variations on that image -Incredibly fast, at roughly 13cents per token DallE Mini is good at: -Rougher and more abstract interpretations of single-weighted input (example, "Flower Fork" came up with abstract forks turning into flowers, DallE 2 came up with forks and flowers. Similar results shown throughout my experiments) -Less filtered results -Free, can be run locally so speed can be improved Disco Diffusion is good at: -Multiple, complex inputs with varying weights -Fine control over multiple aspects of the generation process -Seed control, ability to generate similar results including in-art camera movement -Reinterpretation of image and video input -Free, or $10/month for heavy use. Slow and demanding, running it locally requires some real hardware you're not gonna have and getting good speeds is neigh-impossible. In terms of getting back a solid results with the least amount of work? Dall-E 2 hands down. But I think I'm more interested in the potential of mastering Disco. Consider this animation: https://www.youtube.com/watch?v=HA3kL57Lf3c The frame count is making slight changes to the prompt while the seed is kept static, resulting in a slowly morphing image. They've played with the model to improve things like how the colors change over time. The DallE 2 API might allow for similar results (I haven't played with it yet) but the sheer cost to make that animation makes it pretty prohibitive. Atoramos fucked around with this message at 17:19 on Jul 29, 2022 |
|
# ? Jul 29, 2022 17:14 |
|
I wanted a Golden Corral on the moon with the earth in the background but it really wanted to give me The Piss Moon instead
|
# ? Jul 29, 2022 17:17 |
|
|
# ? Jul 29, 2022 17:51 |
|
SniperWoreConverse posted:somebody do all the ancient gods of the past. You know the ones everyone forgot, bronze age idols, 10000 bc, that kinda poo poo. Crom. Humiierc. Bokrug. Yog. "Depiction of the construction of the enormous statue of Bokrug, forgotten ancient God, by his worshippers"
|
# ? Jul 29, 2022 18:30 |
|
I love when it tries to use words. It's never just random letters, it seems like it understands English in a very distorted way.
|
# ? Jul 29, 2022 18:44 |
|
It's the Ho-Hod GI... Ho-Ho D GI? That GI is totally gastrointestinal
|
# ? Jul 29, 2022 18:48 |
|
veni veni veni posted:Anybody have one(s) that came out so good you almost don't want to post it online in case you want to use it for something later? lol
|
# ? Jul 29, 2022 18:54 |
|
sporklift posted:Once they updated the TOS and allowed for commercial use I generated 20 or so in the same style that I will use for promoting my mixcloud show over the next year or so. How does that work? I saw you're supposed to indicate that it is AI generated anywhere it's used. I want to run a thread making people avatars but I don't want to risk my use if they get strict about that. "Disclose the role of AI. You must clearly indicate that images are AI-generated - or which portions of them are - by attributing to OpenAI when sharing, whether in public or private." "Indicate that the content is AI-generated in a way no user could reasonably miss or misunderstand."
|
# ? Jul 29, 2022 19:00 |
|
axolotl farmer posted:This prompt worked pretty well. Someone care to try it on Dall-e or Midjourney?
|
# ? Jul 29, 2022 19:01 |
Ok, did a little bit of a dive into Dall-E's ability to generate variations from an image input: -Very good at filling in transparent space between objects. -Very good at taking some initial image with big transparent areas, and applying a new prompt to those areas -Very bad at taking a full initial image and generating a new theme to it without explicit, transparent areas. I uploaded some of my own artwork into it surrounded by transparency, and used that to generate new themes which worked pretty well: My image inputs: Was able to generate:
|
|
# ? Jul 29, 2022 19:03 |
|
AARD VARKMAN posted:How does that work? I saw you're supposed to indicate that it is AI generated anywhere it's used. I want to run a thread making people avatars but I don't want to risk my use if they get strict about that. the copyright of AI images isn't exactly decided by the courts or laws yet. they clearly trained these things on copyrighted material, so gently caress em
|
# ? Jul 29, 2022 19:34 |
|
Rutibex posted:the copyright of AI images isn't exactly decided by the courts or laws yet. they clearly trained these things on copyrighted material, so gently caress em good enough for me Please enjoy this horrifying example of what happens when it tries to generate too many people "historical photograph of haters and clout chasers twisting in anger"
|
# ? Jul 29, 2022 19:37 |
|
Rutibex posted:the copyright of AI images isn't exactly decided by the courts or laws yet. they clearly trained these things on copyrighted material, so gently caress em I think they are more worried about covering their rear end so if you get a convincing image of say Trump's piss tape they can say you were supposed to disclose that instead of pass it off as real, clearing them. Hopefully.
|
# ? Jul 29, 2022 19:38 |
|
Stubb Dogg posted:Just got access Dall-E 2 and tried this one, I'm not sure if it got the style quite right Cool! I like that there’s a tiny ox in one them for some reason.
|
# ? Jul 29, 2022 19:57 |
|
https://www.midjourney.com/app/ Midjourney is very good if you don't have access to Dall-E-2 yet. cowgirl shrek in the style of a 70s sci-fi book cover bone-chilling skull creature gliding through a darkened city lost factory interior It also has a community feed, where people share what it's made for them and some of them are amazing:
|
# ? Jul 29, 2022 20:12 |
|
MJ is $10/month for 200 GPU minutes after you use up your trial ones. If you haven't paid yet, you are using a limited number and wont be told any of this until you have like 2 images left. Every image is about 1 minute on average. Upscaling and variants have slightly different costs but they have most actions tuned to take about a minute.
|
# ? Jul 29, 2022 20:26 |
|
AARD VARKMAN posted:How does that work? I saw you're supposed to indicate that it is AI generated anywhere it's used. I want to run a thread making people avatars but I don't want to risk my use if they get strict about that. Check the discord, the official company rep has lifted some restrictions but says TOS isn't updated yet.
|
# ? Jul 29, 2022 20:55 |
|
I decided to give Hulk Hogan meat shoes a try on Dall-e 2 but getting the meat drawer was nearly impossible. I did manage to produce this one that has an amazing innovation in meat storage though, the entire sole can be filled! "Yellow and red reebok shoes with a heated drawer full of beef jerky in the sole, product photo, 1992" I also got a laugh out of one generation that just produced a few strawberries I also tried craiyon, and while I didn't get a jerky drawer I was shocked at how good it is at shoes now so I generated the weirdest footwear I could think of craiyon is getting pretty drat good
|
# ? Jul 29, 2022 21:19 |
|
You made me curious about a more literal interpretation "Air Jordan shoes made out of deli meat" I think this one is almost successful? assuming that's rotting meat of some kind
|
# ? Jul 29, 2022 21:28 |
|
craiyon is pretty good at shoes made of meat also someone requested I make the shrek boots in a kermit style so have that too
|
# ? Jul 29, 2022 21:40 |
|
If it came down to it, I'm curious how the DALLE people would even be able to prove that an image was generated using their AI, especially when the whole point of it is to produce something that looks like it was created by a human. I don't see any watermarks other than the little colored squares at the bottom right, which could be easily cropped out.
|
# ? Jul 29, 2022 21:43 |
|
Devils Affricate posted:If it came down to it, I'm curious how the DALLE people would even be able to prove that an image was generated using their AI, especially when the whole point of it is to produce something that looks like it was created by a human. I don't see any watermarks other than the little colored squares at the bottom right, which could be easily cropped out. They currently keep all the output images, if they also have the random seed they can recreate the same thing each time.
|
# ? Jul 29, 2022 23:34 |
|
Tunicate posted:They currently keep all the output images, if they also have the random seed they can recreate the same thing each time. You’ve shown a lot of knowledge about these models in this thread. Could you tell me a little bit more about the random seed? I know if I put the exact same prompt into DALLE2, it’ll give me a different series of results every time.
|
# ? Jul 30, 2022 00:17 |
|
King Carnivore posted:You’ve shown a lot of knowledge about these models in this thread. Could you tell me a little bit more about the random seed? I know if I put the exact same prompt into DALLE2, it’ll give me a different series of results every time. basically the dalle model is trained by taking real images and adding progressively large amount of random noise to them, and making the model try to restore the lost detail with its best guesses of what's being depicted (with some guidance from the prompt and CLIP, an already trained image-to-text model). As the noise progressively gets worse and worse , the model has to hallucinate to invent more and more plausibly accurate information, until you can feed it in random noise and tell it to 'recover the chicken that I noised up', and it will create a chicken even if no chicken existed. (a lot like the puppyslug deepdream thing) When you generate an image, it starts with a 64x64 grid of noise*, and works from there. If they're at all sensible about debugging, that grid of noise is actually produced pseudorandomly from a seed value, which gets stored - that way if something goes wrong in generation, they can retrace every step as things go through the model and find if there's a bug, and see how altering different parts of the model alter outputs. Midjourney actually lets you set the seed when generating images, which makes it possible to make incremental changes by tweaking different parameters *Incidentally, this is why the 'erase and inpaint' feature operates on a 64x64 grid (of 16 pixel by 16 pixel square blocks) - it takes the areas that are erased, rerandomizes them, and goes back to the first step with the new prompt as guidance. Tunicate fucked around with this message at 00:52 on Jul 30, 2022 |
# ? Jul 30, 2022 00:49 |
|
AARD VARKMAN posted:How does that work? I saw you're supposed to indicate that it is AI generated anywhere it's used. I want to run a thread making people avatars but I don't want to risk my use if they get strict about that. I will probably put a note on the tracklist part of my mixcloud page. But if they expect me to plaster a disclaimer on a 500x500 image for a radio show that gets 50 listeners on average then I really don't see the point of using it and I guess they can revoke me . If you can't actually use the images generated then it is just a dumb toy.
|
# ? Jul 30, 2022 00:58 |
|
tried a lot of variations on "beksinski painting of a gigantic imposing <animal>" lots of them just didn't work very well but some good stuff Pig Guinea pig Sheep definitely worked best This one is goat but I added "towering over a forest" and these two are with the addition of "flying"
|
# ? Jul 30, 2022 02:23 |
|
|
# ? May 29, 2024 20:31 |
|
|
# ? Jul 30, 2022 02:44 |