It is that time of year again to see just what we can make with AI generation. Unfortunately who knows which anons originally made these threads back in the days of the very first ancestors of things like disco diffusion. The attached image was made with midjourney. You do this and that and then you get such and such finally resulting in whatever. I was writing a tutorial but ran out of steam. Perhaps I will post it in a million years.
Essentially I want you to go use the Midjourney free trial, or the Nvidia gaugan, or the Dallemini, or the Comic Face Diffusion Model variant within the recent Disco Diffusion or Centipede Diffusion colap jupyter notebooks, ... or I'm sure there's a million others available online, and I want you to try to make not just cruddy anime art but actually awe-inspiring stuff that makes you think you personally could put Tsutomu Nihei out of work by NEET effort alone.
Obviously this is not yet true but the results are rather fascinating, aren't they? I'll post a few more and then I'll just sit and wait for this thread to die.
It's possible to generate a Nihei-style composition as seen in the recent color renditions of Blame!.
Ah, these pictures are big because the nn in this case really does not lend itself to its pngs being minified. It does make full use of the png, even if not necessarily to create coherent art.
You can recreate this, if you can use midjourney using this prompt: - s.mj.run/kxOWDdbqrNM side-by-side final results, nihei-esque
(That first URI is dereferenced to some cloud location where midjourney stores the previous image generated as you progressively refine your images. It is interpreted as an "image prompt" in the same vein as many other contemporary image-generating nns. The original images in the OP post were made by hunting through the midjourney discord for a nihei-looking starter image someone else probably labored over for hours to make just right -- generally, a disco diffusion or midjourney input to midjourney results in excellent output.)
A very similar prompt yielded this result. Originally, the two halves were indisernible, they were perfectly identical. After an single upscaling pass, where the image is converted from 128x128 to 1024x1024 (the second upscaling pass, which yielded the previous image, results in 1664x1664, though supposedly custom dimensions can be specified as well), the figures became very distinct. The more cartoonish a human looks, the worse the upscaling performs. There is nothing stopping you from using krita or gimp to overlap images, erase the undesirable parts of new images, and replace them with the desired parts of old images, only to then use that new custom composite as the image prompt to a new prompt entirely (though your results may certainly vary; presumably you want to control the random seed value as well when doing such work).
We needn't restrict outselves to complicated prompts. These were generated after a couple of passes with only the text prompt "Nihei Sameface" and neither image prompt nor command-line arguments. The quality is much lower but that's a small price to pay for a potentially good starting point for more elaborate prompts using these outputs as subsequent inputs (and there is no reason why you can't combine outputs to make a single input).
Such linear combination of niheifaces can suddenly, out of nowhere, for no reason at all, yield actual nihei art! Kinda. The previous images were both used as image prompts and the text prompt was changed to "sketch in the style of tsutomu nihei of blame!".
Finally, we have "meta text tags" wherein we indirectly bias our prompt in favor of certain pieces of training data and against other pieces. For example, western art is usually published in a taller aspect ratio than eastern art. On the left we have a 1:2 AR and on the right the typical manga 2:3 AR. The text prompt is again "nihei samefaces" yet we get very clearly western art on the left with wildly different facial features. This is not only due to more realistic art favoring perhaps "taller" faces but also due to the nn prefering features extracted from its training data fitting into a taller imaginary rectangle. This fancy AR customization, used such as with "--ar 2:3", is necessary to avoid creating squished characters lacking torsos (midjourney really likes drawing legs for some reason, and they're usually rather good legs, too).
We can also get "low-end niheifaces" by simply interrupting the upscaling process and extracting an intermediate face. For some reason, this consistently gives very WIDE eyes, consistent with early nihei art. This image was such an interrupted result corresponding to the process yielding the OP image.
That's just about it. Here is one final image demonstrating some horrifying results, and sometimes some interesting results. The text prompts are again very simple - I like very simple text prompts, but complicated ones are also acceptable. For Tsutomu Nihei, his name along is acceptable, but, as danbooru 2021 seems to have been used in the training data of midjourney, to get really good results we need to use verbatim danbooru tags, for example, "nihei_tsutomu" (with the underscore). This is kinda nice. The following were made with an interesting collection of horror, such as these tags: - dead_eyes - glazed_eyes (as in "glazed-over blind eyes") - asanagi
Well. Ahem. For some reason the most horrifying danbooru tags also render the best. I won't try to think about why that might be.
Is this the same AI people were messing some time ago or a different ai?
Matthew Adams
There are a few different categories. This is a very different AI.
Quite a few years ago, the very first diffusion-style AIs were advertised on Any Forums. These created blurry images with bright features. The best folks could get out of them were things like highly-detailed oak leaves or zucchini flowers, though you could see little elves or gnomes marching across a forest or stone-bordered river occasionally.
From that AI came the many diffusion AIs which, for whatever reason, lended themselves well to integration with state-of-the-art AI methods. Disco Diffusion was most popular for a couple of years. Today, it has further differentiated into things like Centipede Diffusion. These systems can also produce video animations; you can zoom into and rotate around a noneuclidean/nonrectilinear scene, and is kinda makes sense.
Dall-e and its descendants, Imagen, Parti and this one, Midjourney, are all related. They behave totally differently.
Nvidia's gaugan is the easiest way to make photorealistic images. Scroll down in nvidia.com/en-us/research/ai-playground/ until you find it. It is a fun little webapp you can play with in a phone, unlike the heavyweight colap jupyter notebooks usually used to interact with the *-diffusion AIs.
The above uses a very old version of the *-diffusion family. It is somewhat challenging to find the most up-to-date of that family of notebooks.
Confusingly, midjourney produces results that can sometimes resemble the *-diffusion models. Though unreleated technically, they are trained similarly.
Alexander Cook
>the very first ancestors of things like disco diffusion Excuse me?
CLIP-guided diffusion wasn't even close to the first AI art tools we were using (this was two years ago at most mind you). You skipped over the VQGAN+CLIP era where everything exploded. Learn some history.
Brandon Sanders
bleh, i know nothing of ai history. I only know I can click funny buttons and make funny pictures appear, and that vaguely these aggregations of jupyter notebooks with buttons or the equivalent fall under a few families that get passed around the internet and incrementally improved. If I can't play with it then I don't know about it, and even if I can I hardly know anything about it.
nvidia gaugan examples attached, these take about 1-2 seconds to create.
But what AI do you use to make waifus as the thread theme says? Because most seems to have a hard time grasping what humans are besides being meatbags.
OP where do I read more about this stuff? What about automatic colorizers? Can you tell me a good reference or index for multiple AI or projects like these?
Hunter Bennett
Automatic colorization can probably be done with this prompt:
You can imagine doing this with automatically with a browser extension. $50/mo rents an A100 80gb part-time on some services so that's somewhat cheap.
Jeremiah Phillips
>It's possible to generate a Nihei-style composition The composition and colors in this image are really fucking cool. I don't know a lot about ai; did you feed it something similar actually by Nihei? Or did the ai just come up with this shit by combining patterns. I guess what I'm asking is that is there an original Nihei with even relatively similar composition.
Jose Fisher
Oh, nevermind, mangadex deleted direct links to images.
Colton Parker
>waifus Kill yourself normalfag scum.
Jace Morris
Recently generated these using a Dall-e bot with the keywords: blame lain
