Stable Diffusion

[Deleted User]

You probably seen a few of these images make the rounds on everyone's favourite bot infested social media and the seething it has caused.



This was created using StabilityAI's open source deep learning algorithm Stable Diffusion.
Hugging Face has an online demo where you can play around with it but if you find it limiting or doesn't provide you the results you want, don't worry, here's links to help you get it up and running.


Requirements before continuing.

• NVIDIA GPU with CUDA (10, 16, 20, 30 series), AMD works but this works better on NVIDIA, all AI things work better on NVIDIA, blame AMD for not giving a shit.
• 4GB+ VRAM, 10GB is the recommended, please for a love of god don't have less than 4GB, I'm running this on 4GB and it's terrible, don't be poor like me.
• Mid to high end CPU. 6C/12T minimum.
• Time and patience, art takes time, whether by a human or by AI, be prepared to sink hours rerunning different variations to get the result you want.

This is the Stable Diffusion web UI fork, I use linux so installing this was piss easy, I don't know about microshit or faggple, so follow the readme and you'll be fine.
Once you have this setup, you might notice the output isn't much better than the online demo, grab the Waifu Diffusion model from the list, v1.3 is my recommendation and any embeddings you like.

• Waifu Diffusion gives you waifus and attractive looking females compared to the base model, feel free to experiment with others though, don't yiff too hard.
• Embeddings is data trained on a particular art style, as per the readme, I recommend asanagi, henreader and rustle.

Some extra reading material to help you, here's the web UI wiki if you need troubleshooting.

--lowvram --always-batch-cond-uncond --opt-split-attention --force-enable-xformers

This is a savior for <10GB of VRAM, makes it run okay on my measly 4GB.


If you want to run this on the CPU then clone the repo and add the following to the webui-user.

--skip-torch-cuda-test --use-cpu all --no-half

For me on my Ryzen 7 5825U with 8C/16T it takes around 8 minutes for 40 steps, this is compared to 5 minutes with 42 steps on my GTX 1650.


And the halfchan wiki with some helpful prompts depending on what you want to make.
To help you get started, here's a tweaked prompt and negative prompt I use based on the twitter user that generated these images.

FEMININE,((PERFECT FACE)),((SEXY FACE)). (((アニメ))). (((LARGE BREASTS))), (TONED ABS),thicc thighs. EVOCATIVE POSE,looking at camera, small smile. (INTRICATE),(HIGH DETAIL),SHARP, full body

((((ugly)))), (((duplicate))), ((morbid)), ((mutilated)), (((tranny))), (((trans))), (((transsexual))), (hermaphrodite), [[[out of frame]]], ((((extra fingers)))), ((mutated hands)), (((deformed hands))), ((((poorly drawn hands)))), fused fingers, deformed fingers, mutated fingers, ((poorly drawn face)), (((mutation))), (((deformed))), (((blurry))), ((bad anatomy)), (((bad proportions))), ((extra limbs)), cloned face, (((disfigured))). (((more than 2 nipples))), deformed breasts. gross proportions, ((missing arms))), (fused arms), (deformed arms), (((missing legs))), (((extra arms))), (((extra legs))), (((long neck))), deformed vagina, missing vagina. fat, obese

• prompt in () will strengthening the prompt, adding more will strengthening it even further
• prompt in [] will weakening the prompt , adding more will weakening it even further

Don't have a good CPU or GPU? Don't worry, if you have an internet connection and can make a dummy google account then you can also join in with google colab.
I recommend TheLastBen or ddPn08 since you can add your own ckpt instead of being forced to use the official one.


Go nuts and be sure to share your seed alongside images and if you have the hardware capable of training embeddings, give us a hand with more art styles from Japanese artists.

Hope there will be a Kazuma Kaneko embedding one day.

 

[Loded Diaper]

I really don't care that much, i think AI drawings is really it's own thing. but i am desperate to talk about something .I think techfags and artfags are both kinda retarded.

1. It feels if an artist got angry at cameras, because some peolpe stop commissioning him to draw portraits.
2. Like yeah your kinda retarded if you think a Photograph is better just because it's more "realistic".
3. But your aslo retarded if you think Photographs don't count as thier own art.

Anyway enjoy your art.

 

[Saibrah]

As an artist myself I kinda wanna seethe and I kinda like it. I don't know what to think. There's a lot of valid points as to why it won't replace art but they all seem to apply to right now (ie. critiques of how it sucks at anatomy, all looks the same) when i'm more scared of how fast it's improving, with the potential to throw that criticism out the window in probably the near-ish future. There's a feeling of dread I kind of can't shake off.

At the same time, I noticed that NovelAI's stable diffusion can generate big breasted anime girls so now I'm interested. Thanks for the info I'm gonna go play around with this.

 

[Deleted User]

As an artist myself I kinda wanna seethe and I kinda like it.

Actually the seething was more wokies angry people said AI perfected the female form and the character is white
I don't think any artist really cares about this.

https://twitter.com/i/web/status/1576591378259529728

I noticed that NovelAI's stable diffusion can generate big breasted anime girls so now I'm interested. Thanks for the info I'm gonna go play around with this.

Please post your results, been playing with it for the past few days and I finally have a few that looks okay-ish to share, results are interesting.

 

[ジエントＰ]

There is 12 others that didn't turn out good but I'm glad some of it did turn out okay.



anotherhiki, what's your process of outputting the AI, what prompt, steps, resolution, hypernetworks and such did you use?

 

[Deleted User]

Nice
Been making these for another project.
This took probably a week for a result I'm satisfied with.

 

[Mandrake]

Nice idea for a thread, I use both NAI and local SD (with NAI model, mixed model with SD etc).
I make them for my own enjoyment, erotic or just nice, or both.
I'll post some of the good results I got later here.

 

[Deleted User]

anotherhiki, what's your process of outputting the AI, what prompt, steps, resolution, hypernetworks and such did you use?

I run it locally on my machine with the fork in the OP.
It runs on a GTX 1650 so pretty low spec.

FEMININE,((PERFECT FACE)),((SEXY FACE)). (((アニメ))). [erect nipples], vagina,(((LARGE BREASTS))), slim, (TONED ABS),thicc thighs. EVOCATIVE POSE,looking at camera, small smile. (INTRICATE),(HIGH DETAIL),SHARP. full body, simple white background, standing straight

((((ugly)))), (((duplicate))), ((morbid)), ((mutilated)), (((tranny))), (((trans))), (((transsexual))), (hermaphrodite), [[[out of frame]]], ((((extra fingers)))), ((mutated hands)), (((deformed hands))), ((((poorly drawn hands)))), fused fingers, deformed fingers, mutated fingers, ((poorly drawn face)), (((mutation))), (((deformed))), (((blurry))), ((bad anatomy)), (((bad proportions))), ((extra limbs)), cloned face, (((disfigured))). (((more than 2 nipples))), deformed breasts. gross proportions, ((missing arms))), (fused arms), (deformed arms), (((missing legs))), (((extra arms))), (((extra legs))), (((long neck))), deformed vagina, missing vagina. fat, obese

Steps: 42, Sampler: DDIM, CFG scale: 10, Seed: 3842841314, Size: 768x1088, Model hash: 84692140, Denoising strength: 0.7, First pass size: 448x640

If you choose to run it locally, download my images, then open it in PNG Info and you can copy all my settings to text2img.

Nice idea for a thread, I use both NAI and local SD (with NAI model, mixed model with SD etc).
I make them for my own enjoyment, erotic or just nice, or both.
I'll post some of the good results I got later here.

Nice
Always happy to see more on varis play with AI art.
Same here, for me this is something fun to play with, just letting it generate new images while I play games.

How is NAI compared to WD? I purposefully didn't choose any existing style and let SD run wild to see what I can get and I think the results are pretty good.
Still can't get the hang of outpainting though
There's honestly too many features and new ones added frequently, I don't think I'll ever be able to try all of them unless I buy a better machine.

 

[Mandrake]

Nice
Always happy to see more on varis play with AI art.

Happy to share, more shared art is good!

Same here, for me this is something fun to play with, just letting it generate new images while I play games.

It's honestly just so good to be able to think of anything and be like "let's try it", then being able to try it.
Sure it's not as advanced as having literally an actual artist doing the job, but it's good enough to test concepts or go in a libido fueled rabbit hole lol.

How is NAI compared to WD? I purposefully didn't choose any existing style and let SD run wild to see what I can get and I think the results are pretty good.

It's in another level, the novel ai people really made something crazy good with that model. I recommend getting the NAI leaked model instead of using WD, then using it as is or/and mixing it with SD, WD, or even R34/Yiffie models.
The online version of NAI is extremely efficient to make good looking results almost every time, but kinda is limited in options, while the leaked local one gives incredible results when you tweak it around and mix with other models.

Still can't get the hang of outpainting though
There's honestly too many features and new ones added frequently, I don't think I'll ever be able to try all of them unless I buy a better machine.

That's the only problem for me, it's very difficult to find up to date wikis or explanations on what each configuration does, so I have to look around, play with stuff, until I get more or less what each does.
Also, so many ways to do the same thing, and since it's very new, no one has any opinion on which is best.

 

[Deleted User]

Happy to share, more shared art is good!

It's honestly just so good to be able to think of anything and be like "let's try it", then being able to try it.
Sure it's not as advanced as having literally an actual artist doing the job, but it's good enough to test concepts or go in a libido fueled rabbit hole lol.

It really is. When I found out about it, I spent a whole week obsessively generating as much as I can just to see what results I get, I had so much fun.

It's in another level, the novel ai people really made something crazy good with that model. I recommend getting the NAI leaked model instead of using WD, then using it as is or/and mixing it with SD, WD, or even R34/Yiffie models.
The online version of NAI is extremely efficient to make good looking results almost every time, but kinda is limited in options, while the leaked local one gives incredible results when you tweak it around and mix with other models.

Downloading it now. I originally didn't download it because I didn't know how good it was. Got to say though, having used both v1.2 and v1.3, WD has improved significantly, the creator said he/she was going to remake the model so I'm assuming v1.4 will be a massive improvement. Shame NAI didn't make their model public since they want you to pay for to use it.

That's the only problem for me, it's very difficult to find up to date wikis or explanations on what each configuration does, so I have to look around, play with stuff, until I get more or less what each does.
Also, so many ways to do the same thing, and since it's very new, no one has any opinion on which is best.

That's why I made the thread, we can share and document how we made things, it might never reach the wider web but at least we know what to do.

 

[Mandrake]

Downloading it now. I originally didn't download it because I didn't know how good it was. Got to say though, having used both v1.2 and v1.3, WD has improved significantly, the creator said he/she was going to remake the model so I'm assuming v1.4 will be a massive improvement.

You don't need to download the full 50GB+ btw.
What I recommend is to either use it as is, or mix it with 10-20% of WD1.3 just to get some stuff back like artist styles, or mix with SD to get a more hybrid style.
Also yeah WD1.3 is ok but trust me, NAI model is insanely good at anime style(s).
WD1.4 I'm sure will use inspiration from NAI to finetune something good.

Shame NAI didn't make their model public since they want you to pay for to use it.

I'm fine with them doing their own thing as it's basically simplified SD when you don't have a gpu at home or want quick simple but good results. I wish they made their work public at some point though, to keep pushing the community but hey, we at least got a boost of quality through the leak.
I use both online NAI and local DB anyway, I like their work overall and even after the leak, I'm sure they will keep improving.

 

[Mandrake]

Here are some results I got over time with playing around with stuff I found here and there. I tried to only have the less explicit stuff here but it's still NSFW.
Overall it's either the NAI model (locally installed), or NAI + another one, usually a mix of WD/SD or maybe Yiffie.

This one I still do because the results are always stunning, using a mix of NAI + SD + WD gets me that when inputting Elsa (from disney) + Taylor Swift, which gives the image very good symmetry and a beautiful face.

I think i got this one playing around a furry hypernetwork :

Random Chinese input gives me this cool output

1980s style output can be super nice :

Playing with hypernetworks again, this time Aini I think

Was playing with the womb tattoo stuff and I stumbled upon this cool looking tattoo.

Of course, these are the results of many failed attempts and cursed images output.
I'll post about the tips and tricks I understood over time to get a good image later.

 

[Deleted User]

You don't need to download the full 50GB+ btw.
What I recommend is to either use it as is, or mix it with 10-20% of WD1.3 just to get some stuff back like artist styles, or mix with SD to get a more hybrid style.
Also yeah WD1.3 is ok but trust me, NAI model is insanely good at anime style(s).
WD1.4 I'm sure will use inspiration from NAI to finetune something good.

After playing with it for a day, NAI leak is amazing, you don't need to emphasise your prompts much and baby it unlike WD and especially SD. Still doesn't have the artist I like though but I'm sure that will be added in time.

I'm fine with them doing their own thing as it's basically simplified SD when you don't have a gpu at home or want quick simple but good results. I wish they made their work public at some point though, to keep pushing the community but hey, we at least got a boost of quality through the leak.
I use both online NAI and local DB anyway, I like their work overall and even after the leak, I'm sure they will keep improving.

It goes against the spirit of stable diffusion and what stability AI wanted by making their ckpt not accessible to the public. The project is to democratise AI and have it accessible by everyone not just those that pay money for it. If it's not ready and novel AI will release it one day then fine but it doesn't look like they will.

 

[Mandrake]

After playing with it for a day, NAI leak is amazing, you don't need to emphasise your prompts much and baby it unlike WD and especially SD.

IKR!
It's amazing either by itself or by mixing it as some kind of MSG in all other models.

Still doesn't have the artist I like though but I'm sure that will be added in time.

Many artists names got purposefully weighted to 0 by NAI to avoid shit. Shit that we still got anyway with so many western artists going insane in twitter the first few days.
You can play around adding them back using SD0.8+WD0.2 for example.
I had some fun playing with purely SFW artists but in NSFW situations, something they'd never do IRL, lol.

It goes against the spirit of stable diffusion and what stability AI wanted by making their ckpt not accessible to the public. The project is to democratise AI and have it accessible by everyone not just those that pay money for it. If it's not ready and novel AI will release it one day then fine but it doesn't look like they will.

Sure, I was talking about my opinion on it, not what they should have done ;)

 

[Deleted User]

IKR!
It's amazing either by itself or by mixing it as some kind of MSG in all other models.

It really is, NAI is how I expected the AI to work, I don't know what training NAI did but it's significantly better than SD and an improvement over WD.

IKR!
Many artists names got purposefully weighted to 0 by NAI to avoid shit. Shit that we still got anyway with so many western artists going insane in twitter the first few days.
You can play around adding them back using SD0.8+WD0.2 for example.
I had some fun playing with purely SFW artists but in NSFW situations, something they'd never do IRL, lol.

I think we all expected western artists to sperg out, they always do.
My comment was more, my favorite artist that doesn't design anything anymore is too niche therefore isn't in WD or NAI, would love to recreate Kazuma Kaneko's style but barely anyone knows who he is.

Sure, I was talking about my opinion on it, not what they should have done ;)

No problem, I like the goals that Stability AI has and wish other companies using it would contribute back so the community has more things to play with.
Reading the changelog for WD and seeing him/her spend about 3.5K just to give it to everyone for free is the type of community effort Stability AI wanted.
Unlike someone on varis, I don't think AI should be exclusive to the rich but accessible for everyone.

I'll post some of my newer creations once I find something I'm satisfied with.

 

[Mandrake]

It really is, NAI is how I expected the AI to work, I don't know what training NAI did but it's significantly better than SD and an improvement over WD.

AFAIK they trained their model in a giant cleaned and properly organized dataset, way more than WD did, for longer, and with actual researchers behind the scene to tweak things out. No wonder the result is that good.

I think we all expected western artists to sperg out, they always do.

It was really sad to see the difference between the western artists having a total meltdown trying to get the new tech banned by every mean possible, from behaving like the music industry by dmcaing their own works to shaming people using it, and the eastern ones trying to see how they could use the tool, and stop the stupid shit like AI art not being properly tagged in pixiv etc.
The contrast couldn't have been bigger, it's amazing in a way.
Though I guess the negative IQ reactions I saw were just the usual vocal people doing their usual performative antics.

My comment was more, my favorite artist that doesn't design anything anymore is too niche therefore isn't in WD or NAI, would love to recreate Kazuma Kaneko's style but barely anyone knows who he is.

You can probably look into hypernetworks and embeddings for that.

 

[Deleted User]

AFAIK they trained their model in a giant cleaned and properly organized dataset, way more than WD did, for longer, and with actual researchers behind the scene to tweak things out. No wonder the result is that good.

Ah, makes sense. WD v1.3 only ran for 10 epochs and it's not bad to be honest, it's not as good as NAI but compared to vanilla SD, it's an improvement.

It was really sad to see the difference between the western artists having a total meltdown trying to get the new tech banned by every mean possible, from behaving like the music industry by dmcaing their own works to shaming people using it, and the eastern ones trying to see how they could use the tool, and stop the stupid shit like AI art not being properly tagged in pixiv etc.
The contrast couldn't have been bigger, it's amazing in a way.
Though I guess the negative IQ reactions I saw were just the usual vocal people doing their usual performative antics.

The contrast really couldn't. I honestly would love to understand why western artists reacted like this but then again western artists in general don't like their eastern counterparts, look at how many of them hate how manga and anime is popular and keep trying to alter language or obfuscate their sales just to make western entertainment look good or the constant attempts to tear down manga and anime.

You can probably look into hypernetworks and embeddings for that.

I really should, I'm stuck with a shit 1650 and I haven't used google colab before.
I ironically already have a lot of his art saved so properly organizing and cleaning it for the embedding is easy.

 

[Mandrake]

Ah, makes sense. WD v1.3 only ran for 10 epochs and it's not bad to be honest, it's not as good as NAI but compared to vanilla SD, it's an improvement.

I have high hopes for WD1.4.

The contrast really couldn't. I honestly would love to understand why western artists reacted like this but then again western artists in general don't like their eastern counterparts, look at how many of them hate how manga and anime is popular and keep trying to alter language or obfuscate their sales just to make western entertainment look good or the constant attempts to tear down manga and anime.

From what I've seen it was mostly mediocre artists chasing clout or being their usual mediocre selves. I'm sure many good artists are actually low key working on harnessing the power of this tool instead of twitter crying.

I really should, I'm stuck with a shit 1650 and I haven't used google colab before.
I ironically already have a lot of his art saved so properly organizing and cleaning it for the embedding is easy.

Ah yeah a 1650 is on the verge of unusable for this.

 

[Deleted User]

I have high hopes for WD1.4.

Me too, hope he/she opens up something so people can pool in money to help pay for it.
$3.5K is a lot of money and I'm sure many would love to chip in and help out.

From what I've seen it was mostly mediocre artists chasing clout or being their usual mediocre selves. I'm sure many good artists are actually low key working on harnessing the power of this tool instead of twitter crying.

I'm sure the good artists are looking at the best way for he/she to use AI, well the eastern ones are at least.
It's always clout chasing, they can't help themselves and twitter is a shit platform because it supports their ideology.

Ah yeah a 1650 is on the verge of unusable for this.

I really wish I can get better hardware, currently looking at building a cheapo machine with a 2060 Super to run SD better.

 

[Mandrake]

I really wish I can get better hardware, currently looking at building a cheapo machine with a 2060 Super to run SD better.

What about used 3000s cards? They are flooding the market, though it depends on your budget of course, but imo it's worth keeping an eye on them.
Though the vram is probably the most important to get the best high quality results.