- Joined
- Jun 23, 2016
Stable Diffusion, NovelAI, Machine Learning Art - AI art generation discussion and image dump
- Thread starter Near
- Start date
-
🐕 I am attempting to get the site runnning as fast as possible. If you are experiencing slow page load times, please report it.
- Joined
- Jan 28, 2018
That option basically just runs a different, face restoration "AI". It does work but it will make the faces look a lot different and quite same-y.
- Joined
- Dec 12, 2015
From what I've generally seen, if the character is not in the foreground and the main focus of the image you're generating isn't their face, odds are decent their face may get messed up. Here's a few examples from before my recent re-install, if it's working never ever update.
In this particular case, I'm using prompts like "close-up, extreme closeup" and "portrait, face focus". Basically, the face is the main focus and I think I've very rarely seen one of these with a heavily distorted face. At worst it's usually a miscolored eye which isn't hard to fix with 30 seconds in an image editor.
So, as you can see, this particular face didn't turn out so well and I believe running face correction didn't help at all in this particular case. As a general guideline, I tend to think of it in terms of "the farther the character is from the point of view, the more likely it is their face will get distorted". Inpainting/img2img on the face and photoshopping may have been able to fix it, but decided not to bother in this case.
Pretty sure these two images had the same prompt as the one prior; as you can see, much closer to the point of view, vastly superior faces.
Another general rule; in my opinion most models really don't perform well when the character is oriented/facing away from the point of view. Straight on to the back of the head seems fine, but any other angle facing away often seems to have issues.
Overall, this seems to be more of an issue when using realistic/painting styles than anime styles, but I may be either remembering incorrectly or more fortunate in my experimentation.
TL;DR -
1. Characters should face you and if possible be oriented towards you.
2. Characters (or their faces) should dominate the work and take up the majority of it for best results.
- Joined
- Dec 17, 2022
- Joined
- Sep 28, 2022
"Adolf Hitler with a catgirl"
Done using everything 4.5
Done using everything 4.5
Attachments
- Joined
- Jul 19, 2022
This is cool shit, but fuck if it hasn't made me depressed about learning drawing. Whats the point if a machine can do it better?
- Joined
- Aug 28, 2019
Learn survival skills instead. Like what to do after nuclear war, or where to shoot drones and killer robots.
vladimirovanova
kiwifarms.net
- Joined
- Mar 11, 2019
I mean, to me it's like playing chess. AI crushed humans in chess forever and I don't think a human will ever beat a chess machine ever again except if the computer is nerfed to shit. That doesn't make playing chess any less fun.
Though I guess it's different since it's not like chess-playing employs millions. But honestly, if you really love drawing, a machine doing it better shouldn't demoralize you any more than a human master doing it better.
On that note, a lot of the anti-AI bros on Twitter are already throwing in the towel and asking to just have AI art be segregated (snigger) from them now. I bet anything they'll be angry as soon as another big studio uses AI, but for Twitter, it's surprisingly not as vitriolic as I was expecting.
- Joined
- Oct 6, 2016
Better a machine than another human. It's a common beginner's trap to get intimidated by the skills of other people. This just evens the playing field.
- Joined
- Jan 28, 2018
Even with all the improvements in this technology we have seen in the last months it's really difficult to impossible to make it do detailing in a certain way or really just draw a specific motive. With some work you can get roughly what you wanted, but never exactly besides sheer luck. As long as that's the case, I personally don't see the direct competition with human artists.
- Joined
- Dec 17, 2019
Exactly. AI art is good for churning out pictures quickly, and can produce some surprisingly good results at times, given the right combo of model, prompt, and seed. However, when it comes to specifics, that's where you'll either need to try and wrangle it further with inpainting and Photoshop and such, or just accept that you're not going to get it exactly how you want and leave it at "good enough."
That said, you can use it as a springboard to get a rough concept of your idea, and then take the output to a human artist to refine and redraw with the details you want. Instead of ideas being trapped in your head and being unable to show them without spending months or years learning, or having to hope that an artist can understand your description and translate correctly, you can have the computer get you most of the way and the artist the rest.
So to @alpha889boba and any other artists-in-training, don't see AI as a threat. It's a tool like any other, one that you'll probably be able to get a lot of use out of too.
- Joined
- Jun 17, 2021
I just powered up my old laptop, and I found some images generated in April 2022 using Looking Glass. I'd completely forgotten about this.
Compare that to some of the recent images I've generated with Stable Diffusion
It's fucking ridiculous seeing how far the tech has come in less than a year.
Compare that to some of the recent images I've generated with Stable Diffusion
It's fucking ridiculous seeing how far the tech has come in less than a year.
Last edited:
- Joined
- Jan 10, 2020
Looking Glass was Google's and through their servers only, right? Now you can generate something far better not even for a premium on someone's enterprise hardware but on a consumer pc built for gaming from like 3 years back. I hope we see this kind of advancement in text ai.
- Joined
- Jun 17, 2021
Looking glass was a prototype of Dreambooth that used ruDALL-E as the base model. In this case, I fed Harold McNeill's Circle of Protection paintings and some Fullmetal Alchemist fan alchemy circles as training data.
I believe you were able to run the code locally since the project was hosted on colab, but it's been so long I can't remember.
And I fully agree. I think OpenAI not sharing ChatGPT is highly unethical, since it's stifling the tech's evolution and going against the company's original mission.
- Joined
- Jan 10, 2020
It goes against the company's own name. Why did they keep that when they decided to sell out?
- Joined
- Aug 28, 2019
I've been critical of OpenAI for years and I don't remember noticing the convoluted for-profit transformation until recently:
OpenAI - Wikipedia
They have been whining about the dangerousness of their own tools since at least GPT-2. Partly because of Musk killer robot rhetoric when he was a bigger influence there, the usual fake news and deepfake reasons, and now marketing and protection of the golden eggs they are laying. This shit is so good we can't even release it to the general public, yo.
There was a ton of press behind the OpenAI name even before the GPT-3/ChatGPT craze. OpenAI sounds nice and... open. I don't think they are going to change the name anytime soon, no matter how closed they get.
Girl Named Sandoz
kiwifarms.net
- Joined
- Aug 11, 2022
by far the best way i've found to "fix" faces is just using hires fix. if you aren't aware, almost every one of these models is trained on images at a 512x512 resolution. the more you stray from that, the more likely it is to start introducing weirdness to your images. basically once it gets to 512 it "restarts" the generation, and anything beyond that will be duplicated or tiled. this is what leads to those images you see where it looks like it gave up halfway through making a character and started generating a new body where their neck should be or things like background objects and characters repeating. ideally you would generate the initial image as close to 512 or less as you can, but because of the low resolution it can struggle with finer details in things like faces or hands. hires fix works by first generating a smaller image at the resolution you specify, and then essentially doing img2img passes over that image at a higher resolution. it works very well both for fixing imperfections and for adding more details to your image, depending on the upscaler you're using.
a couple examples:
not terrible but kind of derpy
much better!
it doesn't even have to be a huge difference in image size, 1.25x should be plenty just to get better faces. play around with different upscalers for different results. latent tends to add a lot of detail but can only really be used at higher denoise values and it sometimes changes the image a lot. ESRGAN works well at lower denoise, around 0.3 should be fine.
really, try upscaling your images in general. you might be surprised how much different the final results can be!
initial image:
1.25x hires fix:
2x img2img:
this was actually an upscaler meant for anime images but i liked how the contrast and colors turned out
1.25x hires fix:
2x img2img:
this was actually an upscaler meant for anime images but i liked how the contrast and colors turned out
same settings as the first one
Last edited:
- Joined
- Dec 28, 2014
I think it will probably be soon at generating more visually compelling procedurally generated content for games. I bet we see another tantrum when that starts happening.
- Joined
- Jun 17, 2021
Friend just finished fulfilling her kickstarter after she allocated the art budget to a pair of A4000s and generated all the art with a custom dreambooth model. Lowered her total budget from $10k for commissioning artists to $2k for the GPUs and $4k on printing books and game pieces.
Nobody knows the art was AI generated, especially since a lot of the imperfections are lost in printing.
- Joined
- Jan 28, 2018
I think first we'll see about a metric fuckton of low effort shovelware using this to add graphics, probably with poor and disjointed results.
Not saying it can't be used in a good way, just the kind of people who want to go the low-effort road don't lend themselves to well, effort.
Still, from voice generation to graphics generation, this is going to be a boon for video game devs. I wish I had that shit 20 years ago when I wrote my early version of stellaris/elite crossover. Nobody wants to play that stuff text-only.