Stable Diffusion, NovelAI, Machine Learning Art - AI art generation discussion and image dump

[Overly Serious]

Stability AI has opened up the Stable Diffusion 3 API to developers.

https://alternativeto.net/news/2024/4/stabilityai-opens-its-nextgen-text-to-image-generator-stablediffusion-3-api-to-developers/

Hopefully we'll be getting standalone models and software soon.

That crappy reddit forum of Stable Diffusion is just filled with people with early access doing people's prompts for them and an endless cycle of:
"This is disappointing"
"Early ____ was disappointing too. Wait for Fine Tunes".
<repeat ad infinitem>

Seriously - the forum has itself just become an overtrained AI at this point. Just people repeating the same responses back at each other over and over.

 

[AmpleApricots]

Seriously - the forum has itself just become an overtrained AI at this point. Just people repeating the same responses back at each other over and over.

They're almost all coomers. Their reward centers are busted. It'll always be disappointing, no matter what it is. Also for being the "I fucking love science" crowd they've neither respect for, nor understanding of the process. If it doesn't tickle those receptors now now now now! it's just shit.

 

[Dick Mason]

OpenAI winds down AI image generator that blew minds and forged friendships in 2022 (archive)

DALL-E 2 is no more.

Changes to Twitter itself—especially with the transition to Elon Musk's ownership in October 2022—further fractured the AI art community as many people left the platform. Lapine said she has now lost track of most of the people in the community from those early days.

The glory days of 2022. It was so long ago! Who can keep track, where did they all go? Time swallows all, doesn't it. Wtf, they're lamenting like it wasn't just last year.

oh dear:
View attachment 5919786

btw bing has chosen @Null 's dog breed at last [this is the winner because it obeyed my spelling prompt, good boy!]
View attachment 5919787

With the amount of data they'll have on us, by 2030 these things will be able to read many peoples' minds. These last few years of anonymity will be nice, but the stylometric sunset will be brutal.

 

[Overly Serious]

They're almost all coomers. Their reward centers are busted. It'll always be disappointing, no matter what it is. Also for being the "I fucking love science" crowd they've neither respect for, nor understanding of the process. If it doesn't tickle those receptors now now now now! it's just shit.

I agree. But is there a better place I can mainstream discussion on this stuff anymore? Everything seems to have become Discord which is a poor way to have discussions. On the same theme, is there an alternative to Civitai? I have filters on to guard against some of the worse content on there and like somebody else said, those filters are actually pretty good so I appreciate them. But still, is there anywhere else to find the models and LoRA and see convenient galleries of them all in one?

With the amount of data they'll have on us, by 2030 these things will be able to read many peoples' minds. These last few years of anonymity will be nice, but the stylometric sunset will be brutal.

Is "stylometric sunset" your own turn of phrase? Because it deserves some kind of award.

 

[whatever I feel like]

Civitai needs its own community thread. The shit there is as gross/cringy as the DA thread's content and almost as bad as pixiv's.

 

[macrodegenerate]

Civitai needs its own community thread. The shit there is as gross/cringy as the DA thread's content and almost as bad as pixiv's.

Is there a pixiv horrors thread yet? There really isn't anything impressive on Civitai as far as cringe. I guess some of the Loras and Models are pretty disgusting, but based on the Deviantart thread if we start the thread it's just going to be 8 people renting out Podrun instances generating terabytes of fat fetish pornography, and sharing with the class.

 

[whatever I feel like]

Is there a pixiv horrors thread yet? There really isn't anything impressive on Civitai as far as cringe. I guess some of the Loras and Models are pretty disgusting, but based on the Deviantart thread if we start the thread it's just going to be 8 people renting out Podrun instances generating terabytes of fat fetish pornography, and sharing with the class.

There is, not sure if its active currently. I was randomly browsing the civitai gallery yesterday and I saw big buff pepes fucking women, and ladies that were also chairs, thats what sort of put me over the edge.

 

[Roland TB-303]

Is there a pixiv horrors thread yet? There really isn't anything impressive on Civitai as far as cringe. I guess some of the Loras and Models are pretty disgusting, but based on the Deviantart thread if we start the thread it's just going to be 8 people renting out Podrun instances generating terabytes of fat fetish pornography, and sharing with the class.

I’ve found one or two creators who make interesting stuff but it’s few and far between.

 

[Overly Serious]

Civitai needs its own community thread. The shit there is as gross/cringy as the DA thread's content and almost as bad as pixiv's.

On the one hand, it deserves it. But on the other do we really want a thread that is 90% pasted screenshots of weirdly sexualized foxes or Kim Possible naked?

(To be clear, I just made those examples up as the sort of degeneracy I imagine I would find if I turned my filters off).

 

[inception_state]

On the one hand, it deserves it. But on the other do we really want a thread that is 90% pasted screenshots of weirdly sexualized foxes or Kim Possible naked?

(To be clear, I just made those examples up as the sort of degeneracy I imagine I would find if I turned my filters off).

I'm doing my part!

Spoiler: NSFW, Definitely not Kim Possible as a weirdly sexualized fox

Agreed though, a thread needs to be funny to be successful, at least horrifying in a weirdly captivating way (eg. the tranny surgery thread). Weird fetish pornography is just gross, not usually funny.

 

[Rembrandt]

Weird fetish pornography is just gross, not usually funny.

There is clearly room for it on KF though, since the Furry Art Freakshow thread is 300 pages strong

 

[Overly Serious]

I'm doing my part!

Spoiler: NSFW, Definitely not Kim Possible as a weirdly sexualized fox

View attachment 5930550

Dammit!

BTW, I was curious if KiwiFarms preserves the metadata in images or if it strips it like Reddit and seems in his benevolence that Null permits it. I dragged your image into my own ComfyUI set up as I'm always curious to see other people's workflows. This is what my system thinks of it, fwiw:

Spoiler

As you can see, something went very wrong here. But then given the subject matter, maybe I don't want it to be right.

 

[Involuntary Celebrity]

It's one of those things where you can't tell when they're sincere. Like how nobody's actually meant to jerk off to shitting dicknipples (we will ignore the generation of fucked-up zoomers who didn't realise it was a joke for the purposes of this argument), a lot of the really bizarre shit on civit is probably just intended for genning shit that will freak people out.

I mean, you could focus on the more mundane weird stuff there instead but that's all kinda tame in comparison.

 

[Susanna]

It's one of those things where you can't tell when they're sincere. Like how nobody's actually meant to jerk off to shitting dicknipples (we will ignore the generation of fucked-up zoomers who didn't realise it was a joke for the purposes of this argument), a lot of the really bizarre shit on civit is probably just intended for genning shit that will freak people out.

I mean, you could focus on the more mundane weird stuff there instead but that's all kinda tame in comparison.

I think maybe you’re underestimating just how degenerate men can get. I’d accompany this with a photo of a diaper tranny polycule, but then I’d have to look at one. :horrifying:

 

[Involuntary Celebrity]

I'm certain that the only reason someone would cut their penis off and then swaddle the wound in a shit-filled diaper is because they possess a true commitment to comedy.

 

[Slav Power]

Speaking of degenerate furry shit, this talk inspired me to suffer with my 1060 and ComfyUI again to try and make something half-decent for once.

After waiting 90 seconds per every parameter change, dozens of rejected generations and attempts, here's the final result:
(WARNING: DEGENERACY)

I know, I should be burned on a stake for using SD for shit like this.

It's funny how with ML art, the real limiter to what you can do is your hardware. Any basic generation will take minutes instead of seconds, and you'll have to grind through multiple of them, so you get badly dissuaded from playing with it. And by the end of it, you're limited to screwing around with proompts, parameters, LoRA's, and models limited to SD 1.5 because you can just barely run those, and you'll avoid doing mass batches of random seeds to find a good base. And SD upscaling? Forget about it, gotta settle for some ESRGAN.

I wish I could generate higher quality stuff, and faster, but this one image took me like two hours of fiddling around to get a decent result, so an imperfect blurry low res picture with ESRGAN upscaling is the best I can do.

 

[Involuntary Celebrity]

Speaking of degenerate furry shit, this talk inspired me to suffer with my 1060 and ComfyUI again to try and make something half-decent for once.

After waiting 90 seconds per every parameter change, dozens of rejected generations and attempts, here's the final result:
(WARNING: DEGENERACY)
View attachment 5932414
I know, I should be burned on a stake for using SD for shit like this.

It's funny how with ML art, the real limiter to what you can do is your hardware. Any basic generation will take minutes instead of seconds, and you'll have to grind through multiple of them, so you get badly dissuaded from playing with it. And by the end of it, you're limited to screwing around with proompts, parameters, LoRA's, and models limited to SD 1.5 because you can just barely run those, and you'll avoid doing mass batches of random seeds to find a good base. And SD upscaling? Forget about it, gotta settle for some ESRGAN.

I wish I could generate higher quality stuff, and faster, but this one image took me like two hours of fiddling around to get a decent result, so an imperfect blurry low res picture with ESRGAN upscaling is the best I can do.

Try SDXL Lighting if you just wanna play around. There's several versions optimised for working with super low sample steps (2, 4, 8, and 10 iirc, and last I checked they were experimenting with a 1-step edition). I have decent hardware but it can do XL gens with the 4-step model that would normally take me tens of seconds in like 1 second.

It's somewhat limited due to aggressive pruning but does surprisingly well with photoreal stuff. And a lot of XL LoRAs still work with it, though they'll add gen time and I've found there's a lot of variability in how much influence they'll actually have at low step counts (but like 10 steps might turn out fine, so it's still a big time save overall).

 

[macrodegenerate]

Speaking of degenerate furry shit, this talk inspired me to suffer with my 1060 and ComfyUI again to try and make something half-decent for once.

After waiting 90 seconds per every parameter change, dozens of rejected generations and attempts, here's the final result:
(WARNING: DEGENERACY)
View attachment 5932414
I know, I should be burned on a stake for using SD for shit like this.

It's funny how with ML art, the real limiter to what you can do is your hardware. Any basic generation will take minutes instead of seconds, and you'll have to grind through multiple of them, so you get badly dissuaded from playing with it. And by the end of it, you're limited to screwing around with proompts, parameters, LoRA's, and models limited to SD 1.5 because you can just barely run those, and you'll avoid doing mass batches of random seeds to find a good base. And SD upscaling? Forget about it, gotta settle for some ESRGAN.

I wish I could generate higher quality stuff, and faster, but this one image took me like two hours of fiddling around to get a decent result, so an imperfect blurry low res picture with ESRGAN upscaling is the best I can do.

It's possible you are using your card incorrectly. 1060s are awful at 16fp, try disabling 16fp and force 32fp. This is an issue with tesla p40s as well which are designed to work with 32fp.

Source:https://www.techpowerup.com/gpu-specs/geforce-gtx-1060-6-gb.c2862

 

[Slav Power]

It's possible you are using your card incorrectly. 1060s are awful at 16fp, try disabling 16fp and force 32fp. This is an issue with tesla p40s as well which are designed to work with 32fp.

Source:https://www.techpowerup.com/gpu-specs/geforce-gtx-1060-6-gb.c2862

No, it's 100% on the GPU being over 7 years old. I added --force-fp32 to the launch parameters and it took ~84 seconds to output an image at 2.7s/it. Yes, second per iteration, not iterations per second. Without it it's roughly the same.

I can't be gaslighted into this cope that my hardware is fine and it's the software that's at fault. So many performance issues that I was having with video games or other software got improved by me moving on from Haswell to Alder Lake, instantaneous improvement. And I can tell that the 1060 is no longer cut for modern games and software, bottlenecking in some titles at 100% utilization while the CPU barely breaks a sweat, where I can't hold stable 60FPS at 1080p.

At this point if I want to pursue ML image gen further I just have to upgrade to something with a stronger core and more VRAM. There is no free workaround to it.

 

[Susanna]

At this point if I want to pursue ML image gen further I just have to upgrade to something with a stronger core and more VRAM. There is no free workaround to it.

The 4060Ti is a reasonable budget option. It’s not super fast, but it does have 16GB of memory.