AI [AI Art] - Show Us Your AI Skill *NO TEENS*

4.80 star(s) 6 Votes

Jimwalrus

Well-Known Member
Sep 15, 2021
1,091
4,168
For a laugh (and primarily because everyone says it's straight-up impossible) I tried running Stable Diffusion on my elderly GTX 960 with just 2GB of VRAM. Everywhere you look says that 4GB is the absolute minimum, with 8GB being the 'true' minimum if you don't want to sit around all day.

Well, I enabled the 'Low VRAM' setting (which is intended for those with <8GB) and gave it a go. It won't take negative prompts, it won't run GFPGAN for improving faces & it won't upscale using ESRGAN, but it would allow me to use F222 as the model. How it runs a 4GB model in 2GB I do not know.

The prompt was "Taylor Swift posing nude on a couch, naked, blonde". 100 steps, CFG of 12

After 6 and a half minutes(!) it gave me this:
-1-9199777399-scale15.00-ddim-f222.png
Face is all over the place (which GFPGAN would sort, were it available), body has a few issues (fixable if negative prompts were available), but yes - SD DOES run with only 2GB of VRAM!!

Just barely...
 
Last edited:

fr34ky

Active Member
Oct 29, 2017
812
2,200
For a laugh (and primarily because everyone says it's straight-up impossible) I tried running Stable Diffusion on my elderly GTX 960 with just 2GB of VRAM. Everywhere you look says that 4GB is the absolute minimum, with 8GB being the 'true' minimum if you don't want to sit around all day.

Well, I enabled the 'Low VRAM' setting (which is intended for those with <8GB) and gave it a go. It won't take negative prompts, it won't run GFPGAN for improving faces & it won't upscale using ESRGAN, but it would allow me to use F222 as the model. How it runs a 4GB model in 2GB I do not know.

The prompt was "Taylor Swift posing nude on a couch, naked, blonde". 100 steps, CFG of 12

After 6 and a half minutes(!) it gave me this:
View attachment 2302583
Face is all over the place (which GFPGAN would sort, were it available), body has a few issues (fixable if negative prompts were available), but yes - SD DOES run with only 2GB of VRAM!!

Just barely...
Sometimes more steps is worst, I stick with euler-a 20 steps. I don't feel like I understand the software so much to be playing with that stuff yet.

I agree that SD is a nightmare for video cards and it has too many bugs, yesterday I've been the whole day trying to train a model with dreambooth lora, reading and watching videos, just to get to the conclusion that the last version in automatic1111 has a problem with it apparently.

So in the end I just ended creating more girls ((huge breasts and asses)) (that I'd soon post here) while I wait that bug to be fixed :ROFLMAO:
 
  • Haha
Reactions: Jimwalrus and Elefy
Sep 10, 2022
227
267
Sorry to bother but anyone here has recomendations of models for anime ai art? I want to make MHA and DBS ones but the models that I'm currentlt using are not so good, or maybe my prompts are bad but I wound not know, I'm using Invoker AI for anyone curious
 

fr34ky

Active Member
Oct 29, 2017
812
2,200
Sorry to bother but anyone here has recomendations of models for anime ai art? I want to make MHA and DBS ones but the models that I'm currentlt using are not so good, or maybe my prompts are bad but I wound not know, I'm using Invoker AI for anyone curious
I couldn't get very nice results with dragon ball girls to be honest, but I got some hot anime-ish girls with Berrymix model, KriboMix-NSTAL model and I think Healy's Anime Blend is good too.
 

DreamingAway

Member
Aug 24, 2022
313
743
Sorry to bother but anyone here has recomendations of models for anime ai art? I want to make MHA and DBS ones but the models that I'm currentlt using are not so good, or maybe my prompts are bad but I wound not know, I'm using Invoker AI for anyone curious
I couldn't get very nice results with dragon ball girls to be honest, but I got some hot anime-ish girls with Berrymix model, KriboMix-NSTAL model and I think Healy's Anime Blend is good too.
Is "DBS" Dragon Ball?

 
Sep 10, 2022
227
267
I couldn't get very nice results with dragon ball girls to be honest, but I got some hot anime-ish girls with Berrymix model, KriboMix-NSTAL model and I think Healy's Anime Blend is good too.
000014.a08edaf1.489334935.png Getting some decent results using the EimisAnimeDiffussion, apart from the face, although I cant make sex art since is not a nsfw model, gonna try download one and see the results to compare
 

DreamingAway

Member
Aug 24, 2022
313
743
yup, a took a look at this model and I don't think is a nsfw model? but is there anyway that I can merge two different models? or retrain a model with extra data? kinda wish I could train one specifically for the style of MHA and DBS
Yes, you can merge models. See "Checkpoint Merge" tab / documentation. It's actually pretty easy. Getting the right ratios to produce something good might take trial and error though.

This model is based on SD 1.5 (according to the documentation) so it should support NSFW just fine.

You can always baseline your image in this model then img2img in another if you want (For the lewd touch ups).. using your first generated image as a denoise seed will keep the art style consistent.


Does anyone knows what the VAE is?
A VAE is sort of like a guidance for the model, they are usually used to clean up hands or improve color palettes.
They are 100% optional but you should use them if you can.
 

fr34ky

Active Member
Oct 29, 2017
812
2,200
Is "DBS" Dragon Ball?

I'm not very convinced about that model, the sexy pictures seems to have errors. IMO the most important part when doing this kind of characters, is not that it captures exactly the original artist style, but that it has an anime style and also handles all the sexy stuff in a versatile way.

Getting some decent results using the EimisAnimeDiffussion, apart from the face, although I cant make sex art since is not a nsfw model, gonna try download one and see the results to compare
For some reason I think most of anime models were not trained with many DB pictures, since it's hard to get a good looks for DB girls. I'm don't know much about how all the background stuff works, may be blending that toriyama model with another one does some good trick.
 
Sep 10, 2022
227
267
Yes, you can merge models. See "Checkpoint Merge" tab / documentation. It's actually pretty easy. Getting the right ratios to produce something good might take trial and error though.

This model is based on SD 1.5 (according to the documentation) so it should support NSFW just fine.

You can always baseline your image in this model then img2img in another if you want (For the lewd touch ups).. using your first generated image as a denoise seed will keep the art style consistent.




A VAE is sort of like a guidance for the model, they are usually used to clean up hands or improve color palettes.
They are 100% optional but you should use them if you can.
thanks for the info, gonna take a look at it later!
 

Elefy

Member
Jan 4, 2022
241
947
0001-05.jpg

This imput image with img2img Anything3 without post edits

result:
00277.png

Im not so good in prompting , but use this for img2img here:
"sexy woman siting on chair , ((milf face)) ,seductive look,green eyes ,dark red hair,big breasts,hard nipples,leather long boots"
 

DreamingAway

Member
Aug 24, 2022
313
743
sorry to bother again, but the Toryama model is not cpkt but safetensors and Invoke AI cannot read it, did I got something wrong?
safetensors is the future.

PyTorch Checkpoints (ckpt) can be pickled and execute malicious code so the community is moving in the direction of using safetensors because they can't do malicious imports or RCE.

I don't use Invoke AI - but a lot of models are shifting to safetensors so I would hope Invoke AI would add that support soon.

TL;DR - Use Automatic1111
 

Jimwalrus

Well-Known Member
Sep 15, 2021
1,091
4,168
As a bit of a fan of Holiday Island I decided to see what would happen if I tried to make one of the girl's renders a bit more photorealistic.
Starting image (render from game by Darkhound):
Amy_pool_distraction_angry3_swim3.png
Amy has an angry face here, but it's such a good picture I wanted to try it out.
Img2Img, prompts etc are as follows:
You don't have permission to view the spoiler content. Log in or register now.
Result (best of 16 after much finetuning):
-11-2128887760-scale5.00-k_euler_a-f222.png

SD doesn't seem to have been trained on unusual swimsuits, hence it's a bit wonky.
 
Sep 10, 2022
227
267
safetensors is the future.

PyTorch Checkpoints (ckpt) can be pickled and execute malicious code so the community is moving in the direction of using safetensors because they can't do malicious imports or RCE.

I don't use Invoke AI - but a lot of models are shifting to safetensors so I would hope Invoke AI would add that support soon.

TL;DR - Use Automatic1111
Glad that it's safer, shame that Invoke AI don't support them yet, I will take a look at Automatic1111
 

Whitekink

Member
Apr 17, 2022
461
857
As a bit of a fan of Holiday Island I decided to see what would happen if I tried to make one of the girl's renders a bit more photorealistic.
Starting image (render from game by Darkhound):
View attachment 2305366
Amy has an angry face here, but it's such a good picture I wanted to try it out.
Img2Img, prompts etc are as follows:
You don't have permission to view the spoiler content. Log in or register now.
Result (best of 16 after much finetuning):
View attachment 2305371

SD doesn't seem to have been trained on unusual swimsuits, hence it's a bit wonky.
The face is spot on...photo-realistic. (y)
Obviously the over-sized breasts and, as you say, 'wonky' swimsuit, don't fit, but...
 
  • Like
Reactions: Jimwalrus
4.80 star(s) 6 Votes