|
experiments with: Generate 512x512 SD image -> use an image upscaler/enhancer ai (currently testing vanceai.com - mixed-ish results, these are among the better examples) It handles things like wood planks or other basic shapes extremely well in upscaling but detailed things or abstract sort of brushstrokes just turn into mush
|
# ? Aug 29, 2022 21:37 |
|
|
# ? May 28, 2024 22:38 |
|
Goddamn, Dall-E has just introduced outpainting and an eraser tool. This is dangerous.
|
# ? Aug 31, 2022 19:32 |
|
deep dish peat moss posted:experiments with: Generate 512x512 SD image -> use an image upscaler/enhancer ai (currently testing vanceai.com - mixed-ish results, these are among the better examples)
|
# ? Aug 31, 2022 23:39 |
|
Finally, the mystery is solved!! ...now I have more questions... Well how about we discover some of history's most secretive identities? Hm. Maybe they were better left unknown... feedmyleg has a new favorite as of 00:20 on Sep 1, 2022 |
# ? Aug 31, 2022 23:58 |
|
I started throwing a bunch of AI generated art on t-shirts because drat if I could recoup some of the cost I've put into various AI tools that would be cool, I can't afford it anymore https://bleaksploitation.threadless.com/ and when I can afford it I'm gonna buy this forums banner ad Also I've posted several of these individual images but: And here are some scifi panoramas courtesy of SD, upscaled to giant sizes. Unfortunately I don't know where to post the even bigger ones (30+mb 8192x4096 .pngs) SD seems to make very desaturated and low-contrast images overall but I think that's a good thing compared to dall-e's tendency to go all-in on saturation. At least when the source image is desaturated you can just do something simple like duplicate it and set the top layer as a Soft Light layer to do most of the work for you: deep dish peat moss has a new favorite as of 01:40 on Sep 1, 2022 |
# ? Sep 1, 2022 01:35 |
|
feedmyleg posted:Finally, the mystery is solved!! Do Dr Claw
|
# ? Sep 1, 2022 10:13 |
|
https://twitter.com/ai_curio_bot/status/1565262137857032193
|
# ? Sep 1, 2022 15:22 |
|
|
# ? Sep 1, 2022 17:59 |
|
first two are "Donald Trump with too many limbs, big and wet, National Geographic photo, high definition", second two are "Ilhan Omar with her Gundam, Socialist Realist, high definition"
|
# ? Sep 3, 2022 02:03 |
|
Ugh. Keep the strawman arguments in CSPAM.
|
# ? Sep 3, 2022 02:08 |
|
that second trump is clearly boris johnson
|
# ? Sep 3, 2022 03:39 |
|
Ron perlman jellicle cat
|
# ? Sep 3, 2022 17:11 |
|
https://twitter.com/talyarkoni/status/1565894421094322176
|
# ? Sep 4, 2022 04:47 |
|
Neons Generated with Stable Diffusion running on my own dang computer
|
# ? Sep 4, 2022 05:03 |
|
Lmfao https://i.imgur.com/IrDWgof.png
|
# ? Sep 4, 2022 08:52 |
|
Apparently this is what robots looked like in Ancient Rome (via Stable Diffusion)
|
# ? Sep 5, 2022 22:56 |
|
My wife's been playing with Midjourney for a while. Anyway, here's baby Gandalf from the new Lord of the Rings, I assume.
|
# ? Sep 6, 2022 08:33 |
|
hosed up of eru iluvatar to intentionally create a baby that already has wrinkles
|
# ? Sep 6, 2022 09:15 |
|
I made some entirely normal Garfields with stable diffusion. For the whole set (40 in total) see here: https://imgur.com/a/qNhbAp7
|
# ? Sep 6, 2022 12:30 |
|
END CHEMTRAILS NOW posted:Apparently this is what robots looked like in Ancient Rome
|
# ? Sep 6, 2022 12:40 |
|
moonmazed posted:hosed up of eru iluvatar to intentionally create a baby that already has wrinkles
|
# ? Sep 6, 2022 12:57 |
|
Tiggum posted:I think I've seen this Doctor Who episode. Oh God, in three years the doctor who episode about AI artists is going to be very off-putting
|
# ? Sep 6, 2022 13:29 |
|
https://twitter.com/supercomposite/status/1567162288087470081
|
# ? Sep 7, 2022 17:23 |
|
That’s fascinating. Also fascinating is this likely explanation for why that phenomenon occurs: https://twitter.com/mattskala/status/1567300206969982979 Tl;dr Loab is something that the AI keeps coming back to as the result for negative prompts, meaning that it considers her to be the conceptual opposite or at least the furthest thing conceptually from a lot of things in its model. In any given Euclidian space, there is a relatively small number of “furthest points.” The tweet author used the example of a polyhedron with 12 corners. No matter where you are on that polyhedron, the furthest point from your current location will always be one of the 12 corners. Furthermore, you can usually whittle that list down even further to a list of points of which one will always be a reasonable approximation of the furthest point. The image of Loab, then, is probably representive of a concept that is far away from most concepts in the AI’s conceptual model and is therefore approximated as being the opposite of many concepts.
|
# ? Sep 7, 2022 18:20 |
|
Due to the curse of dimensionality, as dimension space increases the number of farthest points increases exponentially (and any normal distribution of data shifts so virtually all points are at an extreme)
|
# ? Sep 7, 2022 18:26 |
|
Tunicate posted:Due to the curse of dimensionality, as dimension space increases the number of farthest points increases exponentially (and any normal distribution of data shifts so virtually all points are at an extreme) The tweet thread does touch on how and why the number of dimensions at play here is likely quite low. I didn’t include it in the tl;dr because frankly it was a bit beyond me.
|
# ? Sep 7, 2022 18:32 |
|
I'm not sure if that's the right explanation. If you lay down a set of random* points in higher dimensions, the ratio of the distance to the farthest point to the distance to the nearest point approaches one pretty quickly. Intuition about distances and geometry in general based on our experiences in low-dimensional spaces generally fails pretty badly in high dimensions. Here's my favorite example of how bad things can get even with as few as ten dimensions: https://twitter.com/CihanPostsThms/status/1566659355084292096 *Here "random" isn't completely general but involves much weaker assumptions than iidness.
|
# ? Sep 7, 2022 18:35 |
|
Ariong posted:The tweet thread does touch on how and why the number of dimensions at play here is likely quite low. I didn’t include it in the tl;dr because frankly it was a bit beyond me. I didn't see this when I wrote my last post. Based on what the author describes I think my objections are addressed.
|
# ? Sep 7, 2022 18:39 |
|
so she's a demon
|
# ? Sep 7, 2022 18:57 |
|
let's all love loab
|
# ? Sep 7, 2022 19:04 |
|
Anyone tried to cross-pollinate Loab with Crungus? Would the result be too horrifying to even show?
|
# ? Sep 7, 2022 19:41 |
|
Here are some things I've made or have been making recently using Midjourney-generated art as a base, because I'm mostly interested in 'functional' use of AI art tools I guess. A font of alien pictoglyphs: Made by using Midjourney to generate prompts like "An alien pictoglyph meaning "idea"" or whatever and then tracing over the generations. An 80s retrowave magazine about a dream-based videogame console/internet with a cosmic horror bent I used midjourney to generate the background stationery (I actually did a huge, huge set of cyberpunk and 80s retrowave inspired backgrounds that look like magazine pages or data pads or whatever), and all of the images are things I've generated previously or generated for this, then the layouts/text/lettering/screen effects/etc. all done manually. anatomical diagrams of mutants and monsters and neohumans to e.g. go in an encylopedic world-building text generating bulk character/monster designs for e.g. a JRPG deep dish peat moss has a new favorite as of 04:07 on Sep 13, 2022 |
# ? Sep 7, 2022 19:57 |
|
Lisa Loab: Surreality Bites
|
# ? Sep 7, 2022 20:01 |
|
Huh, it looks like the Brungus shaved.
|
# ? Sep 7, 2022 21:12 |
|
deep dish peat moss posted:generating bulk character/monster designs for e.g. a JRPG If you are interested in pixel art for a game Dall-e 2 does a much better job than Midjourney: Dall-E 2
|
# ? Sep 7, 2022 21:28 |
|
Loabs Georg who is secretly in every procedurally generated image ever made is an outlier and should not have been counted.
|
# ? Sep 7, 2022 21:36 |
|
|
# ? Sep 7, 2022 21:36 |
|
At last, thru cybernetics, dog has become Cat
|
# ? Sep 7, 2022 21:44 |
|
Rutibex posted:
I don't think any of them handle pixel art well enough to do actual pixel art with. You can mimic it pretty well in dall-e by prompting for a "pixel mosaic" but for it to work right it needs to make each 'pixel' a 16x16 pixel square so you're limited to a 64x64 canvas (per image) and it's just a big pain IMO. But doing it the old fashioned way of shrinking and retouching concept art works really well (though obviously shrink and trace would work better) (these definitely aren't even close to finished, just uploading in-progress things from my ipad) Midjourney seems to be better at fine detail than Dall-E (because of dall-e's 16x16 chunk size limitation mostly) so I've actually found it more useful for generating this kind of stuff because you can have it cram a bunch in a single image without losing much detail
|
# ? Sep 7, 2022 22:35 |
|
|
# ? May 28, 2024 22:38 |
|
I got into Dall-e 2 a while ago and it's so much fun to play around with I'm sure nothing of mine is too groundbreaking here but it's still a good time and I've already pumped some money into getting more credits One thing I want to use it for is to generate character portraits next time I play Baldur's Gate or something (or even to print out on a physical character sheet!) It took plenty of tinkering but I eventually got some favorable portraits: It's also great at making spooky animals: and really cute animals: as a suggestion I put in "the most horrible secret in the universe" and got a pretty chill phone wallpaper that I imagine is the equivalent of : also I like trying to get botanical illustrations out of it: It's been a long time since I did any pixel art but seeing the examples above makes me want to try making some environmental pixel art based off of something generated in dalle! e: i realized this isn't the AI generated art thread, oh well Telltolin has a new favorite as of 02:49 on Sep 8, 2022 |
# ? Sep 8, 2022 02:13 |