Register a SA Forums Account here!
JOINING THE SA FORUMS WILL REMOVE THIS BIG AD, THE ANNOYING UNDERLINED ADS, AND STUPID INTERSTITIAL ADS!!!

You can: log in, read the tech support FAQ, or request your lost password. This dumb message (and those ads) will appear on every screen until you register! Get rid of this crap by registering your own SA Forums Account and joining roughly 150,000 Goons, for the one-time price of $9.95! We charge money because it costs us money per month for bills, and since we don't believe in showing ads to our users, we try to make the money back through forum registrations.
 
  • Post
  • Reply
AARD VARKMAN
May 17, 1993
I've been doing it on Google Colab which seems like by far the best option available right now for free StableDiffusion generation and Img2Img.

Followed this guide

https://sassexplorer.notion.site/sassexplorer/Guide-to-run-Stable-Diffusion-on-Google-Colab-fa803a5cffc04de491eb4a7c036ec15b

but there are a lot of other guides out there

can generate 4 images at 150 steps in under 3 minutes with Colab Pro ($10 a month, but the speed increase is not particularly large over free tier), or a huge number of them if you want, way quicker than the Huggingface web demo and doesn't require you to own a thousand dollar graphics card

Adbot
ADBOT LOVES YOU

AARD VARKMAN
May 17, 1993
it's so nice to just always be able to do 150 steps again for the type of horror stuff i like doing most, it really benefits from extra steps :frogbon:

lunar detritus
May 6, 2009


IShallRiseAgain posted:

I used img2img on Johnny Five Aces:



I hope the next training model is better at dealing with hands. I think it was mentioned that was going to be the case.

That's a fun experiment.

This was my best result without masking

mcbexx
Jul 4, 2004

British dentistry is
not on trial here!



BARONS CYBER SKULL posted:

close up painting, vampire, sci-fi armor, purple smoke hair, beautiful, stern face, gorgeous, face visible, scars, eyeliner, cinematic lighting, eva suit, wallpaper, extremely detailed, sharp focus, by Greg Rutkowski and [[[WLOP]]], intricate, beautiful, award winning, Trending on artstation, pixiv

Which got me



Which tweaking and keeping the seed also got me



Weird how SD assumes the subject's gender. Problematic.

Adding "male" to your prompt.





Bottom Liner
Feb 15, 2006


a specific vein of lasagna

mcbexx posted:

Weird how SD assumes the subject's gender. Problematic.


the more pressing issue is that literally all of these default to white unless you use certain language that is uh, not great.

Moongrave
Jun 19, 2004

Finally Living Rent Free

Bottom Liner posted:

the more pressing issue is that literally all of these default to white unless you use certain language that is uh, not great.

that's based on the artists used more than anything else from my testing

the ai is dumb and can only make stuff from what it's been trained on and the tags it has, most of the tags are english, most of the artists we'd know to use are english, so our, and it's, inherent biases end up with that



i found the the secret incantation to manifest actual blades instead of just metal shivs, and it was steak or chef knife



Moongrave fucked around with this message at 00:31 on Aug 31, 2022

mcbexx
Jul 4, 2004

British dentistry is
not on trial here!



Bottom Liner posted:

the more pressing issue is that literally all of these default to white unless you use certain language that is uh, not great.

Eeeeh, the last one could be interpreted as being asian at least, but yeah, you're not wrong.

But phew, at least it is not actively ignoring color of skin when properly prompted, so ... yaaaaay?

And it looks the coolest, so maybe the AI is trying to make up for its inherent caucasian bias.

As someone who can't draw/paint at all, this poo poo is blowing my mind.
No one is ever going to hire someone to do an album cover ever again, are they?


mcbexx fucked around with this message at 00:42 on Aug 31, 2022

SniperWoreConverse
Mar 20, 2010



Gun Saliva

Vlaphor posted:

I done made robot bears.



You already did maturin?

mcbexx
Jul 4, 2004

British dentistry is
not on trial here!



Any idea why the local SD installation sometimes overwrites the existing output file when you go back and do an iteration with slight differences?

Sometimes it does, sometimes it doesn't and I can't figure out what triggers it.
The output files are numbered, 00000.png counting up and sometimes it just hops back a couple of digits and starts overwriting.

mcbexx fucked around with this message at 01:00 on Aug 31, 2022

Longpig Bard
Dec 29, 2004




AARD VARKMAN
May 17, 1993


"a place that should not exist, colorful photograph by Michal Karcz and Gregory Crewdson"


"a place that should not exist, colorful photograph by Gregory Crewdson and Michal Karcz"

Rahu
Feb 14, 2009


let me just check my figures real quick here
Grimey Drawer

AARD VARKMAN posted:

I've been doing it on Google Colab which seems like by far the best option available right now for free StableDiffusion generation and Img2Img.

Followed this guide

https://sassexplorer.notion.site/sassexplorer/Guide-to-run-Stable-Diffusion-on-Google-Colab-fa803a5cffc04de491eb4a7c036ec15b

but there are a lot of other guides out there

can generate 4 images at 150 steps in under 3 minutes with Colab Pro ($10 a month, but the speed increase is not particularly large over free tier), or a huge number of them if you want, way quicker than the Huggingface web demo and doesn't require you to own a thousand dollar graphics card

Google just announced today that collab pro is moving to a credit-based system instead of unlimited use for $10. Probably directly because of all the people running stable diffusion now.

lol

AARD VARKMAN
May 17, 1993

Rahu posted:

Google just announced today that collab pro is moving to a credit-based system instead of unlimited use for $10. Probably directly because of all the people running stable diffusion now.

lol

I got that e-mail less than 30 minutes after my confirmation on paying for Pro and loled

The Butcher
Apr 20, 2005

Well, at least we tried.
Nap Ghost

AARD VARKMAN posted:

"a place that should not exist, colorful photograph by Michal Karcz and Gregory Crewdson"


Nice one.

"So uh, what's with the hovering orb?"

"We don't talk about the orb."

strange feelings re Daisy
Aug 2, 2000

Rahu posted:

Google just announced today that collab pro is moving to a credit-based system instead of unlimited use for $10. Probably directly because of all the people running stable diffusion now.

lol
Booooo, that sucks poo poo

A Strange Aeon
Mar 26, 2010

You are now a slimy little toad
The Great Twist
Is Virgil Finlay known by the models enough to be used for bad rear end black and white fantasy and horror stuff?

MartingaleJack
Aug 26, 2004

I'll split you open and I don't even like coconuts.
The easiest free version, assuming you have a capable rtx card, is Visions of Chaos. No python or Ubuntu cmd prompts after the initial install. It doesn't have built in img2img but I'm going to guess it's coming in am update. I've been using Gradio for that. Outpainting currently sucks in SD. I hope it gets as good as Dalle-2 eventually.

i am harry
Oct 14, 2003

Rinkles posted:

It's maybe not the best for fine tuning, because minor changes can completely change the output, but k_euler_a can spit out images very efficiently.

These were made in under 4 seconds each (14-18 steps, 3060ti)



the prompt varied but it was always some variation of
    fallout new reno concept art, old truck in a dry wasteland, artstation, thick outlining, subtle lighting, muted colors, symmetrical composition, it's dusk, blue-gray, john berkey, syd mead, Ralph McQuarrie

e:and img2img batches are lighting fast (2s per image here)



i...cant really express how...lucky i feel about abandoning my desires to be a concept artist.

Rutibex
Sep 9, 2001

by Fluffdaddy

i am harry posted:

i...cant really express how...lucky i feel about abandoning my desires to be a concept artist.

don't worry friend the technological singularity moves faster than you would expect once it gets going. Everyone on earth is gonna be out of a job by 2025

The Butcher
Apr 20, 2005

Well, at least we tried.
Nap Ghost

Rutibex posted:

Everyone on earth is gonna be out of a job by 2025

I've already got most of a women's romance novel, airport kiosk quality, targeted for white women into white men written for free for me. Need to work over a few parts and then give it the human editing touch after.

Self publishing stuff is easy as hell. If that one first basically works, I can use the same core bones of it to switch up genders, race combos, sexual preferences, etc. Slap a new AI generated cover on whatever the theme is, use a good pen name that fits the genre... could be money in them hills.

And if that works, I can start getting really weird with it and do BDSM/fetish stuff. That could be even more profitable.

I'd be a little bit nervous to be a creative type right now.

The Butcher
Apr 20, 2005

Well, at least we tried.
Nap Ghost
Some of the weirder ones for books will need a lot of prompt work. Trying dragon and car romance is difficult... "Ford Bronco" is apparently very confusing:

The dragon princess Diana fell deeply in love with her suitor, the Ford Bronco.

But her father, the owner of the Chevrolet Kingdom, forbid their love.

As she took a trip to a small, private zoo to experience the animals' love, Diana began to hate the animal kingdom's business.

It was only a couple of months before the Bronco became a symbol of Diana's true love to the Ford Bronco.

Diana is a dragon princess and, like many other dragons, only has one sister - Herp. [ed: why is this a necessary detail]

The Ford Bronco is not her main car and, like other dragons, Diana can only play on the carriageway.

---

So it's love, but you can't even commit to one car. I dunno Ford Bronco, she might break your transmission.

Doggles
Apr 22, 2007

The Butcher posted:

I'd be a little bit nervous to be a creative type right now.

AI animation is coming soon and it's coming fast. :stare:

https://twitter.com/karenxcheng/status/1564626773001719813

https://twitter.com/karenxcheng/status/1564632137482899460

https://twitter.com/karenxcheng/status/1552348027834118144

Rinkles
Oct 24, 2010

What I'm getting at is...
Do you feel the same way?
Does anyone know how to use the weight system? This is from the hlky fork.



A Strange Aeon posted:

Is Virgil Finlay known by the models enough to be used for bad rear end black and white fantasy and horror stuff?

it recognizes him, but results will vary, and as usual the further away the subject the more likely the ai will screw something up



You can probably get decent results with a bit of work

i am harry
Oct 14, 2003

Rutibex posted:

don't worry friend the technological singularity moves faster than you would expect once it gets going. Everyone on earth is gonna be out of a job by 2025

I’m not worried as much as I am relieved that I didn’t do incur the debt and stress only to have it vaporize one morning

i am harry
Oct 14, 2003


Jesus Christ all the work put into things like Scanner Darkly only to, again, watch it vaporize into dust…and this fashion stuff……..

Ima Computer
Oct 28, 2007

Stop all the downloading!

Help computer.

Rinkles posted:

Does anyone know how to use the weight system? This is from the hlky fork.

(This is hlky-fork specific, as far as I know, but) you can enter multiple prompts separated by colons (:) with optional weights values after them, similar to MidJourney and Disco Diffusion.

Here is two prompts, one with a weight of 1 and one with a weight of 2:

code:
prompt with weight of one:1 prompt with weight of two:2


The weight numbers are optional (each prompt gets a default weight of 1). So this is the same thing:

code:
prompt with weight of one: prompt with weight of two:2


If this checkbox is toggled, it will add up all the weights, then average them and redistribute the weights as decimal values that add up to 1:

code:
prompt with weight of one:0.33 prompt with weight of two:0.67
Decimal values work and negative weights are a thing too as long as the numbers don't add up to less than or equal to 0.

Ima Computer fucked around with this message at 07:11 on Aug 31, 2022

Rinkles
Oct 24, 2010

What I'm getting at is...
Do you feel the same way?
Yup that works, thanks

close up portrait of nomad: Hiroya Oku: moebius:


than with moebius:2, than :4


than instead with Hiroya Oku:2, than :4

Rinkles
Oct 24, 2010

What I'm getting at is...
Do you feel the same way?
Anybody figure out how to get good airplane shots? Cars and trucks it gets right most of the time (to spectacular results), but planes almost always come up as a mishmash of aircraft parts. Sometimes it gets the profile vaguely right, but often it can't even tell the front apart from the back.

Bottom Liner
Feb 15, 2006


a specific vein of lasagna
Same with dragons. I haven't seen a good normal dragon that didn't have like 7 legs or 3 heads in weird places.

AARD VARKMAN
May 17, 1993
If you can't get a good representation of a basic concept then you usually need to change your style prompts. Try running it with no or very few prompts and see what kind of style comes back with a realistic aircraft.

AARD VARKMAN
May 17, 1993
if you could find an artist or photographer that frequently does aircraft and use them early in the "by" it will likely improve things a lot

Gromit
Aug 15, 2000

I am an oppressed White Male, Asian women wont serve me! Save me Campbell Newman!!!!!!!
Try Robert Taylor?

mobby_6kl
Aug 9, 2009

by Fluffdaddy

Rinkles posted:

Anybody figure out how to get good airplane shots? Cars and trucks it gets right most of the time (to spectacular results), but planes almost always come up as a mishmash of aircraft parts. Sometimes it gets the profile vaguely right, but often it can't even tell the front apart from the back.



Try generating 512x512 images first. It often gets confused when it's supposed to have a single primary subject and creates a mess when it's bigger than that. It only kind of works for landscapes and more abstract stuff.

The Butcher
Apr 20, 2005

Well, at least we tried.
Nap Ghost
Yeah airplanes do seem a bit hard, but I'm sure right prompts can get it.

On my first try for the SR-71 Blackbird I got a cool alternate universe one. Not quite anatomically as correct as ours, but still looks pretty dope.

The Butcher
Apr 20, 2005

Well, at least we tried.
Nap Ghost
Got some engines attached now at least, looks like something Batman would be into:

mobby_6kl
Aug 9, 2009

by Fluffdaddy
Hmm, planes seem to be a challenge but it's possible to get something reasonable out of it occasionally.

tote up a bags
Jun 8, 2006

die stoats die

Messed around with this a bit last night. Still getting used to it but had some fun making cupcakes.



and cool satan

A Strange Aeon
Mar 26, 2010

You are now a slimy little toad
The Great Twist

Rinkles posted:

Does anyone know how to use the weight system? This is from the hlky fork.



it recognizes him, but results will vary, and as usual the further away the subject the more likely the ai will screw something up



You can probably get decent results with a bit of work

Those look really cool! I really need to try to get this running locally.

Necrothatcher
Mar 26, 2005




Made some British politicians:









Adbot
ADBOT LOVES YOU

Rutibex
Sep 9, 2001

by Fluffdaddy

Necrothatcher posted:

Made some British politicians:











:eyepop:
was that the prompt, just "British politicians"

  • 1
  • 2
  • 3
  • 4
  • 5
  • Post
  • Reply