|
Someone whipped up their own variant of the methods used for DALL-E, so you can try it here: https://huggingface.co/spaces/dalle-mini/dalle-mini The results are not as impressive as DALL-E 1 or 2, but it definitely can do some creative stuff
|
#
¿
May 2, 2022 21:04
|
|
|
|
| # ¿ May 14, 2024 05:09 |
|
|
The Alchemist posted:With such AI algorithms the "computer, enhance" meme/trope will be real life and people in the future will see our posts mocking CSI etc. thinking we are dumb as poo poo. And they will be correct. Not really - while AI can be used to enhance pictures so they look good at higher resolution, there is no guarantee that the enhanced picture is anywhere close to the truth of the pictured situation. Like when an AI made a pixelated Obama into a white man
|
|
#
¿
May 8, 2022 17:43
|
|
|
 chill it with the cucumbers, lower right robot
|
|
#
¿
May 16, 2022 02:46
|
|
|
Got inspired by this one: https://twitter.com/hardmaru/status/1525787886284849152    Hmm....that last one didn't work out so well.  Better.
|
|
#
¿
May 16, 2022 03:33
|
|
|
They want to avoid enabling prompts like «Obama giving a blowjob to Soros in the Oval Office» or «Trump loving his daughter» or anything involving children, which is understandable. But since they don’t know if there are ways to bypass the filters, they also filter users do they can manage it better.
|
|
#
¿
May 19, 2022 09:04
|
|
|
A challenger appears A blue jay standing on a large basket of rainbow macarons.  A robot couple fine dining with Eiffel Tower in the background. 
|
|
#
¿
May 24, 2022 00:09
|
|
|
furry artists are soon out of a job https://twitter.com/LitRPGforum/status/1528668105970536448 also browse this guy's feed, he's really churning dalle2 for all that it's worth, including telling it to make game assets
|
|
#
¿
May 24, 2022 02:22
|
|
|
precision posted:IF NOUN1 VERB NOUN2 THEN DO NOT OUTPUT NOUN2 VERB NOUN1 but this isn't a bunch of IF statements, it is putting the sentence into a neural network that then outputs a set of numbers, and then that set of numbers is fed into the image generator if it was that easy, why would both openAI and google fail to fix it
|
|
#
¿
May 24, 2022 19:48
|
|
|
precision posted:imagine if they managed to prove that thoughts occur 0.0000001 seconds after the action we think they refer to, proving forever that we literally don't have any control over our actions and only think about them after they happen they already did that, iirc edit: https://www.nature.com/articles/news.2008.751
|
|
#
¿
May 27, 2022 16:08
|
|
|
LifeSunDeath posted:"It's Morbin' Time Jared Leto looking stupid" Those last two seem to just be the AI mashing up the original image. I guess once you get too specific, it recalls instead of creating
|
|
#
¿
Jun 1, 2022 05:55
|
|
|
nice that it was trained before the movie, so we get the real morb, comic book morb
|
|
#
¿
Jun 2, 2022 18:21
|
|
|
I'm wondering why Dall-E Mini even knows what Grogu is, because as far as I can tell two of the three datasets it uses for training does not contain him, and the third seems to only contain images and captions from before 2014. edit: they might have updated the model after they wrote their paper, I guess - trying to find out exactly what an AI has been trained on is a bit hard, especially when I'm not versed in all the terminology ymgve fucked around with this message at 16:00 on Jun 3, 2022 |
|
#
¿
Jun 3, 2022 15:35
|
|
|
err posted:have any normal people gotten dall-e 2 yet? I assume people within the tech industry are the ones most probable to use social media to show off their creations, but I wouldn't put it past them to use another AI to decide who's invited first I hope we all get unlimited access soon, though it will probably cost money like with GPT3
|
|
#
¿
Jun 6, 2022 19:48
|
|
|
drat it is hard getting dalle mini to work locally on windows edit: im currently installing something called bazel to compile a library for windows to get a bugfix that isnt in the binary release of something called jaxlib, and it seems to be a full framework and has compiled 6600 files so far ymgve fucked around with this message at 20:28 on Jun 8, 2022 |
|
#
¿
Jun 8, 2022 20:14
|
|
|
FINALLY got dall-e mini working on my computer. This is "groverhaus in lego"  "bathtub filled with tiny planets"  
|
|
#
¿
Jun 8, 2022 21:35
|
|
|
One thing I'm really impressed with is that even Dall-E Mini is great at "knowing" reflections. This is "flooded picadilly circus":   On the other hand, it has issues with other things. This is "a blue red panda" - it is a bit sad? Maybe that's how it interpreted "blue"    And then I tried a "a small blue panda" - and just got red pandas again.   
|
|
#
¿
Jun 8, 2022 23:19
|
|
|
because I'm a sensible adult I'm now trying to make it make porn, which is pretty hard - maybe some filtering in how it interprets words https://imgur.com/a/9sCMs7T https://imgur.com/a/QtZbz8r - some of these are NSFW, I guess
|
|
#
¿
Jun 9, 2022 00:09
|
|
|
not good at context had to test the online version to see if it was better tuned than the one I run locally, but it's really hard to make it actually make something that's not just an image of the full planet
|
|
#
¿
Jun 9, 2022 01:38
|
|
|
Dall-E Mini completely forgets some parts of the input if it feels like it It managed "blue apple" on its own, but only gave normal apples for "magenta apple" - I feel like it's a lot more focused when it had something close to it in the original training data (lots of google image hits for blue apples, but not that many for magenta) Also compare and contrast "very shiny glass rabbit" (lots of google image hits, so it probably exists a few times in the training data)     vs "very shiny glass capybara"     Reformulating the query made it a bit better though - "capybara figure made of glass" gave these, and there is like one glass figurine on google images and it doesn't look like these at all     Bonus: "underwater phantom of the opera" 
|
|
#
¿
Jun 9, 2022 02:41
|
|
|
KakerMix posted:MidJourney is a ton better than dall-e mini, though I like that dall-e mini seems to handle certain things better anyway plus as far as I can tell, no text restrictions. I can't do anything with "Al Gore" because "gore" is filtered >:| Yeah, but dall-e mini is free and open, which is a big plus, also no censoring I managed to make some porn-like images, but it struggles with the concept of verbs so I haven't managed to make any images with "action" happening (yes I know there is porn on the internet shut up) Also tried really hard to make any gore photos after you mentioned the censoring, but (thankfully) it seems the training data hasn't had much of that. I managed to make some dead pigs, but they were all fully intact and just lying down, with some slightly bloody spots on the skin. I did manage to make it cough up these brilliant pieces though:   "human, damien hirst installation"
|
|
#
¿
Jun 9, 2022 03:18
|
|
|
When the AI misunderstands brilliantly "tower of babel, professional camera"
|
|
#
¿
Jun 9, 2022 03:39
|
|
|
I have a new favourite thing                (most of these are variants of "octahedron of [material] on [place]", most often a coffee table) edit: now we're getting more bold        
ymgve fucked around with this message at 04:37 on Jun 9, 2022 |
|
#
¿
Jun 9, 2022 04:17
|
|
|
A Strange Aeon posted:Can someone do a still from an unreleased live action Calvin and Hobbes movie by a director of your choice? ok. "live action calvin and hobbes movie by wes anderson" - these were the live action ones, most were just comic book looking      "live action calvin and hobbes movie by david fincher"    "cgi calvin and hobbes movie by disney"  "cgi calvin and hobbes movie by pixar"  "live action calvin and hobbes movie by quentin tarantino"    
ymgve fucked around with this message at 05:04 on Jun 9, 2022 |
|
#
¿
Jun 9, 2022 04:56
|
|
|
    I cheated, "boy in red shirt with giant tiger [in place]"     boy in red shirt with giant tiger, movie directed by wes anderson ymgve fucked around with this message at 05:25 on Jun 9, 2022 |
|
#
¿
Jun 9, 2022 05:21
|
|
|
I've been trying to make a "just download and run" package for Dall-E Mini - I haven't tested in on a fresh Windows install, but it should in theory only need the Cuda 11.7 toolkit, then just run the .bat file: https://mega.nz/file/oUty2STD#GYSku4Mm1y5k36vGiXdZcer3VWedj1fyde_2_uJZb2E - package with the small 1.7GB model, might work on older nvidia cards https://mega.nz/file/FZkHlJga#_91XXGD3RGpihl4iRA3RzBo1UlaHBAB9qMMSCQfCobk - addon with the large 5GB model, definitely requires more GPU RAM I got a 3080ti and it takes 10 seconds for each picture, so be prepared to have some patience
|
|
#
¿
Jun 9, 2022 20:49
|
|
|
Atoramos posted:
You need both packs unzipped in the same place, the mega model download is only the model ymgve fucked around with this message at 05:34 on Jun 10, 2022 |
|
#
¿
Jun 10, 2022 05:28
|
|
(I do have Python and the Cuda toolkit installed)
|
Yeah, the image reconstruction part of DALL-E Mini (vqgan imagenet) has issues with faces. Ironically, the best faces I got was with the phrase "brightly lit scene from a porno movie" (which were not brightly lit at all) Technical details - the DALL-E Mini algorithm doesn't actually output an image, it outputs a sequence of 16384 numbers which you could call a "summary" of an image, not pixel by pixel but small and large features. Then it is fed into the VQGAN Imagenet network, which transforms those numbers into a 256x256 image. ymgve fucked around with this message at 06:05 on Jun 10, 2022 |
|
#
¿
Jun 10, 2022 05:58
|
|
|
Vlaphor posted:Speaking of faces...why did this only bring up faces? try "butthole" small snake in a snowglobe  samus aran painted by vincent van gogh (still very focused on starry night)  samus aran statue made by jeff koons  donald trump statue made by jeff koons  joe biden statue made by jeff koons  a piece of grilled chicken fillet with ketchup and pasta on a green dinner plate  star wars drawn by boris vallejo 
|
|
#
¿
Jun 10, 2022 06:14
|
|
|
Some more "X statue made by jeff koons"        edit: some more        
ymgve fucked around with this message at 07:08 on Jun 10, 2022 |
|
#
¿
Jun 10, 2022 06:49
|
|
|
Atoramos posted:Edit: of course. Yup, works perfectly. Great! Could you tell me your hardware and your performance? It takes about 10 seconds to generate an image on a 3080ti here
|
|
#
¿
Jun 10, 2022 08:47
|
|
|
I managed to get Disco Diffusion working locally too, but it's much slower than dall-e mini (takes a few minutes to generate each image) and the results have a certain fever dream "look"
|
|
#
¿
Jun 10, 2022 10:20
|
|
|
Yeah, the neural net library seems to be like "grab ALL the GPU RAM" which makes sense for the standard use case where it runs on a linux server with no display
|
|
#
¿
Jun 11, 2022 09:06
|
|
|
An Ounce of Gold posted:I have been using Disco Diffusion for a week and I am finally starting to crank out some good cartoon style concept art: Are you using the default settings from the workbook, or are you tweaking something? What are the exact prompts?
|
|
#
¿
Jun 11, 2022 14:31
|
|
|
GABA ghoul posted:Hmm, I quickly tried to set up dall-e mini on my machine and the MEGA and MEGA_full models don't fit into my 6GB card. Is there a quick way to force it to fall back to CPU? Or do I have to actually try to understand what I'm doing to do that? Open generator.py and place this line right after "import jax": code:You might have better luck going through the https://colab.research.google.com/github/borisdayma/dalle-mini/blob/main/tools/inference/inference_pipeline.ipynb notebook since that runs on google's servers
|
|
#
¿
Jun 11, 2022 14:52
|
|
|
Justa Dandelion posted:Do you think it would make a difference if the training sets for agi were presented in a way where the concepts of safety, love, and kindness were heavily reinforced before exposing the ai to the information of the world? Sort of like an ideal ai childhood. Worth a try, if it fails we can just wipe the AI and start again (but don’t tell the AI that)
|
|
#
¿
Jun 12, 2022 17:47
|
|
|
Yeah, Disco Diffusion is pretty slow, partially because the defaults generate much larger images than Dall-E Mini. It also seems to have a much more artistic eye, but much less understanding of objects and language than Dall-E Mini. Some of the better ones I made:  abandoned mall photo by edward burtynsky  landfill filled with floppy disks, award winning photography by edward burtynsky  spaceship graveyard photo by edward burtynsky But it often fails in pretty interesting ways:  landfill filled with funko pops, award winning photography by edward burtynsky (it did get the basic idea, but just one giant pop)  a man sitting on the floor trying to fix a computer, painted by Jacques-Louis David, trending on artstation (anatomy? never heard of him)  mewtwo statue made by jeff koons (oh god) Here's Dall-E Mini's attempts at some of the same prompts: (the landfill is too much for the AI to handle and it just devolves into a noisy background)    
ymgve fucked around with this message at 17:14 on Jun 15, 2022 |
|
#
¿
Jun 15, 2022 17:09
|
|
|
You should note that the OpenAI terms of service forbids sharing pictures with realistic faces, so be careful with where you share those
|
|
#
¿
Jun 19, 2022 15:48
|
|
|
Also, if anyone has access to Dall-E 2, could you try the prompt "Landfill filled with funko pop figurines, award winning photography by Edward Burtynsky"? edit: Also you could try some AI prompts that neither midjourney nor dall-e mini nor disco diffusion has gotten yet: "a red panda with blue fur, wildlife photography" or "a blue apple cut in half" double edit: Thanks! ymgve fucked around with this message at 17:07 on Jul 1, 2022 |
|
#
¿
Jul 1, 2022 17:00
|
|
|
External Organs posted:Just a random thought - I wonder if you'd ever see Adobe license this technology and the results would be exported into separate layers in a .psd. Wouldn't work without a complete retooling of how the AI creates images Right now it's all a "package deal" - the AI generates the full picture all at once, there are no layers to separate
|
|
#
¿
Jul 1, 2022 19:21
|
|
|
|
| # ¿ May 14, 2024 05:09 |
|
|
I think the drug dogs were made by fiddling with «neurons» in the model directly, to reveal what a neuron «sees»
|
|
#
¿
Jul 9, 2022 01:32
|
|