Spaces:

stabilityai

/

stable-diffusion

like 10.8k

Running on CPU Upgrade

For Yoshitaka Amano's sake, HOW??!!

For Yoshitaka Amano's sake, HOW??!!

I used SD via https://playgroundai.com/ You do have to register with Google to use it though.

I know the online demo is super buggy atm, it probably has something to do with the transition to 2.0

https://www.engadget.com/stable-diffusion-version-2-update-artist-styles-nsfw-work-124513511.html

very cool. Once I get my new GPU I think it will be easy to achieve with the automatic1111 web ui.

For Yoshitaka Amano's sake, HOW??!!

I used SD via https://playgroundai.com/ You do have to register with Google to use it though.

I know the online demo is super buggy atm, it probably has something to do with the transition to 2.0

https://www.engadget.com/stable-diffusion-version-2-update-artist-styles-nsfw-work-124513511.html

And theeeeeeen?

For Yoshitaka Amano's sake, HOW??!!

I used SD via https://playgroundai.com/ You do have to register with Google to use it though.

I know the online demo is super buggy atm, it probably has something to do with the transition to 2.0

https://www.engadget.com/stable-diffusion-version-2-update-artist-styles-nsfw-work-124513511.html

And theeeeeeen?

Well, then you're gonna have to craft and evolve, lol. Probably quite a few times, in all honesty. It took about 15 - 20 prompts on average for me to get to where I was at least mostly satisfied with the final result.

It's far from a perfect system, too. You're gonna get a lot of 'something went wrong' and 'safety filter' errors. It's often difficult to pinpoint just what went wrong regarding those messages (often simply trying it again a few times will break through), but especially if trying to upgrade simple pixels, it may not have enough to work with. So you have to tweak the image strength, prompt strength, and putting something extra like 'making heroic pose', or ' casting spell' as it will give the AI more to originate. Another solution is copy/paste (in this example) Terra onto a background. Even a real photo can work. Copy/paste can give you more specification on how you want your overall image. It's mandatory to get rid of signatures or watermarks sometimes, too.

I noticed from a regular prompt result that it doesn't know Terra. I've noticed in general if you put most characters from final fantasy, save for maybe a moogle or chocobo or something, it goes towards Cloud. it seems to do alright with classes, however. So for. my results I just put female 'green-haired mage'

The 'warm box' preset filter can often help with anime-oriented images. Same with the 'flat-pallete'. They're tonally pretty different, just depends on what direction you want to go.

Also, be sure to start your prompt with something like 'digital anime illustration' or something, and ending with an actual artist often helps. Make sure your detail is always at least 50.

It's a good idea if trying to upscale a relatively small or simple image, to start in the

Don't forget to upscale. Depending on the complexity of your image you may want/need to upscale and then use the upscaled image as a seed instead of your original image. In general, upscaling (or perhaps facial correction) is a must for your final product.

The sampler you choose matters. It's under advanced options. You'll have to tinker around with them, but in general I can safely say heun gives the most consistent high-quality results. It's just personal preference, but I'd recommend trying it and euler.

Anyway, hope none of this seemed patronizing or anything, and I am far from an expert. I've only been making AI art for roughly 4 months, and 2 of those was only with dall-e mini. Here's my evolving Terra's though :)

oh yeah, the prompts may help, sorry

'digital anime illustration of green-haired female mage from final fantasy VI making defiant pose standing in dimly lit narshe cavern, Akihiko Yoshida, Yoshitaka Amano, miyazaki, complicated background'

Then I just evolved, and moved the image strength up by 10 each time

oh yeah, the prompts may help, sorry

'digital anime illustration of green-haired female mage from final fantasy VI making defiant pose standing in dimly lit narshe cavern, Akihiko Yoshida, Yoshitaka Amano, miyazaki, complicated background'

Then I just evolved, and moved the image strength up by 10 each time

Thank you so much. I'll give it a couple (of thousands) tries.

You bet, good luck! And do post when you're done if you don't mind, I'd love to see how they turn out

You bet, good luck! And do post when you're done if you don't mind, I'd love to see how they turn out

First try. Going to chibi side. I changed "mage" to "magitek warrior/soldier" to see if it tries to make the armor.

I forgot the parameters for this one. Shame on me.

Update. Mixing playground

with https://huggingface.co/Linaqruf/anything-v3.0

I'll stay here from now. I'll update again when I get more interesting results.

I'll stay here from now. I'll update again when I get more interesting results.

They turned out great! And yeah, the armor really works in the last one. Thanks a lot for sharing :)

A couple Lockes I tried :) His arm is pretty janky, and had a lot of trouble getting him to have 'bohemian rag' sort of clothing. Whatever that style is called, lol

https://cdn-uploads.huggingface.co/production/uploads/1669842086973-62f0b1190e5e5ec49861c420.png
I like this one. The proportions on the other one are too messy.
I have another idea. A previous step with resgan. This are the results for direct pixel and resgan:

I'll use those results as seeds to see what can I get.