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

[DavidS877]

Any of you guys know what the current best set up for running it on your own hardware?

You mean hardware or software?
4090, 64GB System RAM minimum, fast NVMe drive(s) for loading the checkpoints. You did say best. Significantly more if you want to do text LLMs.

I really like ComfyUI, it's very extensible and I can make it do stupid stuff like easily share queues among multiple systems. It's also very complex and can be unfriendly.

https://github.com/comfyanonymous/ComfyUI

I found SD.Next was easier to get working for SDXL, I think A1111 can do it but it gave me grief.

https://github.com/vladmandic/automatic

 

[Puff]

SDXL is working swimmingly and the larger load on my machine is made up for by much more consistent good generations. It's taking 5+ minutes to load the models though. Is this normal, or is there some way to speed it up?

 

[DavidS877]

SDXL is working swimmingly and the larger load on my machine is made up for by much more consistent good generations. It's taking 5+ minutes to load the models though. Is this normal, or is there some way to speed it up?

SSD SATA or NVMe. But it's only 6 GB so even a normal HD shouldn't be more than a minute. Might check your performance monitor(depending on your OS) and see what is happening during the load.

 

[XL xQgg?QcQCaTYDMjqoDnYpG]

Not mine; happy Thanksgiving

 

[Puff]

SSD SATA or NVMe. But it's only 6 GB so even a normal HD shouldn't be more than a minute. Might check your performance monitor(depending on your OS) and see what is happening during the load.

It's on my SSD, but I'm only seeing like 40MB read speed max on load and most of the time it's 10 or 15. It hangs up right before "Using pytorch attention in VA" shows in the command window, way before the model is loaded if I'm understanding it correctly. I'm seeing some possible fixes on the github issues list. If anyone else is having the problem, I can post if it works.
Edit:

Spoiler: Fix for slow .safetensors load

Some hardware configurations lose their minds at an inefficient load that is in both A1111 and Comfy. For exceedingly long load times try this
Open ComfyUI/comfy/utils.py
and change
sd = safetensors.torch.load_file(ckpt, device=device.type) [/TD]
to
sd = safetensors.torch.load(open(ckpt, 'rb').read())

(should be line 13 in the current version)

Actually, use at own risk. Worked for a while then got a fairly spectacular failure where the resource manager was showing over a gigabyte a second read from the SSD (physically impossible) then my memory usage jumped up to 99% out of the blue in a few seconds, now it's my computer is totally frozen. Neat.

Edit:Edit: My final solution is to convert eveything over to .ckpt files. The issue is only when loading .safetensors, so this should fix it. I wish I had decent internet. No idea if A1111's conversion loses baked in VAEs

 

[Sergeant Major Buzzkill]

Fields:

 

[Puff]

View attachment 5517209View attachment 5517212

Please use thumbnails, you're making dear leader sad.
Also

Spoiler: kiwis

 

[AmpleApricots]

Most of the pieces are there, a bit deformed, but the relationship between the objects is wrong.

SD does a lot better if you can sketch for it what you want, e.g. make a line art and use a control net. There's also a way to fuse different generations together and also tell SD where you want a specific thing from your prompt to generate in the picture. (positional) These things all work best with ComfyUI. You can also set up interesting workflows in ComfyUI like (automatically) detecting a painted picture on the wall of a generated image, and then (automatically) have that picture's motif be inpainted with a custom prompt.

Dall-E is better because it probably has a cleaner dataset but mostly because they obviously strapped much better language processing to it. SDs design is being to be efficent, not to be as good as possible. (Also later datasets kinda suck)

I know no effort most people probably want to invest but I wanted to mention it. With all these techniques SD is pretty much the only thing on the market that can give you consistent results. Then there's the whole finetuning and lora aspect you'll also never get from Dall-E.

 

[NSJAP]

 

[Puff]

View attachment 5518812

Are you looking for this thread? Or this one?

 

[frozen_runner]

After playing around with SD a bunch (and I have a ton to learn still), I'm less concerned than I was about the apocalyptic impact of this tech. No matter how good it gets, you won't be able to get exactly what you want without directly editing the images, because the inexact nature of language will always be a barrier. I'm starting to see all of it as a multipurpose tool for artists and graphic designers.

How stupid am I?

 

[DavidS877]

After playing around with SD a bunch (and I have a ton to learn still), I'm less concerned than I was about the apocalyptic impact of this tech. No matter how good it gets, you won't be able to get exactly what you want without directly editing the images, because the inexact nature of language will always be a barrier. I'm starting to see all of it as a multipurpose tool for artists and graphic designers.

How stupid am I?

It depends on "exactly what you want"... "Waifu with huge tits" it can do pretty well. "Scene as depicted in Robert Frost's Stopping by Woods on a Snowy Evening" it's going to need some human help... for now.

 

[AmpleApricots]

Biggest problem for SD is the language processing. It's not that you couldn't make it better, it's just that it wouldn't run on normal consumer hardware anymore. I'm pretty sure that will change though as there's a lot of room for improvement in the efficiency department, as last year has shown.

 

[Susanna]

AI also has the advantage that it can produce ten thousand pictures in the time it takes an actual artist to draw a single one. One of those ten thousand is probably "good enough".

 

[Puff]

It depends on "exactly what you want"... "Waifu with huge tits" it can do pretty well. "Scene as depicted in Robert Frost's Stopping by Woods on a Snowy Evening" it's going to need some human help... for now.

FYI:

Spoiler: That prompt

Spoiler: The Text of the Poem

 

[TheDarknessGrows]

How good is this stuff at making porno? Is it there yet?

 

[Catgirls are Love]

How good is this stuff at making porno? Is it there yet?

I'd imagine it depends on the model you're using and how much you care about the parts of anatomy that aren't tits.

 

[Puff]

FYI:

Spoiler: That prompt

View attachment 5521733View attachment 5521736View attachment 5521738View attachment 5521743View attachment 5521745

Spoiler: The Text of the Poem

View attachment 5521750View attachment 5521757View attachment 5521762View attachment 5521766View attachment 5521767

And with "waifu with huge tits" I was unable to get any waifus with huge tits actually (barring one generation which I stopped early to add negative prompt "topless" to remove nipples for kiwifarms posting. Tried all of the XL models I have downloaded. Default color for all of my models seems to be purple, oddly enough.

Spoiler: Notably not Waifus with huge tits

Tried some variations, but it's real hit or miss. The XL models really need a few specific prompts to do anything useful, huh? They really like complex prompts.

Spoiler: Woman with huge tits got 1.5 out of 5 generations being correct.

Waifu with huge breasts got 0
Waifu with huge boobs got 0

Edit:
The models I'm using are surprisingly unhorny.
"Woman with huge boobs, 1girl" got a 60% hit rate, but one was a nigger and another had it's tits and ass on the same side of it's body (clothed if you're worried about that)

Spoiler: Humanity has clearly gone too far

"Woman with huge breasts, 1girl" got an 100% hit rate in my small sample but my anti-nudity prompts were not enough to contain it. The only notable thing is that they were all wearing purple if anything.

Whatever did the tagging for the base model is pretty anti-slang, it seems.

 

[DavidS877]

How good is this stuff at making porno? Is it there yet?

You can go to https://civitai.com and create an account(to unlock the 18+ option) and see what the samples look like.
You can also go to a Discord called "Unstable Diffusion" and see samples there. No link because Discord is cancer, just search for that name.
There's probably a ton of other ways to see what is possible.

 

[DavidS877]

I have a wide range of NSFW and VERY-NSFW(generate NSFW content even on benign prompts) checkpoints, and of the 42 I currently have in rotation: 100% Gave me a human looking waifu with big tits, some asian, some western, no negative prompting for nsfw. A few were anatomically incorrect, as usual. 12 were 'clothed' enough that nipples weren't visible. I'm not going to share any as they all run the risk of the ban-hammer.

I added the simple drawing LORA and got one sharable one:

Spoiler: Waifu With Huge Tits

The results from just asking "scene as depicted..." are all boring, except one, which missed the prompt totally, but was fun. Edit: added a second fun one that did match the prompt.

Spoiler: Scene

The text of the poem also had some fun ones. Almost all had horses.

Spoiler: Poem

None were purple.