AI Art: It is criminal to not post your prompt

The Something Awful Forums > Main > General Bullshit > AI Art: It is criminal to not post your prompt

«‹›435 »

Squatch Ambassador: Nov 12, 2008; What? Never seen a shaved Squatch before?

img2img, prompt: best quality, They Live, Junji Ito, Geiger

https://i.imgur.com/4jXsGPO.mp4

# ? Oct 18, 2022 03:35

Adbot: ADBOT LOVES YOU

# ? May 30, 2024 20:19

TheGoonspiracist: Jul 24, 2002; The terrible secret of space... the Mods, they knew!

What if Norman Rockwell did animorphs covers?

# ? Oct 18, 2022 04:04

Rutibex: Sep 9, 2001; by Fluffdaddy

TheGoonspiracist posted:

thank you for your service

# ? Oct 18, 2022 04:22

TheGoonspiracist: Jul 24, 2002; The terrible secret of space... the Mods, they knew!

# ? Oct 18, 2022 04:57

LifeSunDeath: Jan 4, 2007; still gay rights and smoke weed every day

TheGoonspiracist posted:

OH poo poo this is legit terrifying

# ? Oct 18, 2022 05:29

Commie Lasorda: May 15, 2009; IT'S CLOBBERIN' TIME!

Check out these kickass babies

# ? Oct 18, 2022 05:51

Mindblast: Jun 28, 2006; Moving at the speed of death.

My automatic1111 upscaler stopped working. Idk if its just my install getting borked by a github update or this is something bugged in the latest version.

It gives a warning regarding an upscaler 2 option in the ui config json being incorrect and on the extra tab the second upscaler isn't visible. Trying to use it will immediately error out.

Not the biggest issue, I mean we still got all the other upscaler programs, but I did like the ease of use.

# ? Oct 18, 2022 07:18

Rutibex: Sep 9, 2001; by Fluffdaddy

Drive-Thru Salad Bar posted:

Check out these kickass babies

lol nice

# ? Oct 18, 2022 12:42

AARD VARKMAN: May 17, 1993

Drive-Thru Salad Bar posted:

Check out these kickass babies

bad rear end

# ? Oct 18, 2022 15:11

EVIL Gibson: Mar 23, 2001; Internet of Things is just someone else's computer that people can't help attaching cameras and door locks to!; Switchblade Switcharoo

quote:

don�t try to suffocate the baby Jesus, put him back in the box book_cover

# ? Oct 18, 2022 16:34

mobby_6kl: Aug 9, 2009; by Fluffdaddy

So I've been fighting with SD about applying a particular style to a subject which obviously wasn't trained in that style. I thought this wouldn't be really a problem, it clearly can combine stuff that's never been seen before, like drawings of spaceship by Leonardo Da Vinci, or anime Henry Kissinger, etc. But somehow it won't do an impressionist GT-R:

I've mad an off-hand post earlier with a similar example with a Miata before, and it's basically the same issue with it looking photorealistic and standing out too much.

I tried lowering the weight on the GTR:

And doing that while increasing the weight on Monet:

(left->right is more GTR, top->bottom is more Monet)

The issue is basically the same, it becomes a not-GTR before it's at all impressionist. Yet somehow it can do passable 50s car, which is definitely something Monet could've have painted. Any ideas? I mean other than a photoshop filter

# ? Oct 18, 2022 19:33

Attack on Princess: Dec 15, 2008; _{To yolo rolls! The cause and solution to all problems!}

Maybe try adding different artist prompts into the mix?

There's a thing with SD where it'll occasionally pull the most generic, poorly shaded, tin can robot when you prompt robots and you aren't referencing anything photographic, CGI or artist related.

Similarly, you get weird, bland Super Mario's without support prompts for Mario. It's a reach, but I'm thinking maybe that's happening here.

# ? Oct 18, 2022 19:43

Sedgr: Sep 16, 2007; Neat!

Add photo heavily weighted in the negative prompts? Worth a shot.

# ? Oct 18, 2022 19:48

WhiteHowler: Apr 3, 2001; I'M HUGE!

I decided to get some more experience with fixing a good but imperfect image using only the AI.

I created this starting image with txt2img and a basic prompt containing styles of artists that I enjoy:

Oof. I like the composition, but as you can see, there are several major issues. Deformed faces, weird anatomy, awful hands. The person in the middle kind of resembles Klinger from M*A*S*H.

Using only inpainting -- with lots of configuration tweaks and some prompt fiddling for certain stages -- I was able to create this:

So much better! It's still not perfect -- I had a heck of a time with the right-side lady's hand, and the gentleman's face could use another pass. There are also a few disparities between the sharpness of faces and objects at various distances. Then again, I've seen worse in real, professional paintings, so that doesn't bother me quite as much.

The important part is that a learned a ton about how to do quality image manipulation using only the AI. Sure, I could have loaded up Photoshop or GIMP and probably had a better result in less time, but learning to do this through the AI feels like an accomplishment, and I'm already looking forward to the next project.

# ? Oct 18, 2022 19:55

IShallRiseAgain: Sep 12, 2008; Well ain't that precious?

You can probably use img2img inpainting and adding noise to get what you want. I have found that adding a light amount of noise to something will make Stable Diffusion better able to change something.

I made these 2 images using that technique without much effort:

I've also been experimenting with generating really small source images, enlarging them, and then using img2img to fix them.
Warhammer Inquisitor in movie still of sci-fi film (original prompt used others for img2img)

elmo in Se7en by David Fincher

It seems to fix the problem of character cloning that happens with high res fix, and reduce the extra limbs and other weirdness

Here is a bonus image of pinhead from Hellraiser in a Ghibli Studio film

# ? Oct 18, 2022 20:09

mobby_6kl: Aug 9, 2009; by Fluffdaddy

Donnerberg posted:

Maybe try adding different artist prompts into the mix?

There's a thing with SD where it'll occasionally pull the most generic, poorly shaded, tin can robot when you prompt robots and you aren't referencing anything photographic, CGI or artist related.

Similarly, you get weird, bland Super Mario's without support prompts for Mario. It's a reach, but I'm thinking maybe that's happening here.

Thanks, I tried adding all the impressionist artists I knew, plus stuff like brush stroke and painting but it didn't fundamentally change the situation

Sedgr posted:

Add photo heavily weighted in the negative prompts? Worth a shot.

Unfortunately that didn't change much. But the negative prompt was the right track it seems! I had the usual stuff there like "watermark, signature, deformed, ugly" there, and getting rid of it entirely completely changed the nature of the results.

With negative prompts

Without:

With:

Without:

I do kind of prefer the previous, more vibrant and detailed backgrounds, but at least it's clear what was making the difference! I'll try to see if the prompt matrix works for negative prompt to hopefully figure out what effect each token has.

Thanks y'all.

WhiteHowler posted:

I decided to get some more experience with fixing a good but imperfect image using only the AI.

I created this starting image with txt2img and a basic prompt containing styles of artists that I enjoy:

Oof. I like the composition, but as you can see, there are several major issues. Deformed faces, weird anatomy, awful hands. The person in the middle kind of resembles Klinger from M*A*S*H.

Using only inpainting -- with lots of configuration tweaks and some prompt fiddling for certain stages -- I was able to create this:

So much better! It's still not perfect -- I had a heck of a time with the right-side lady's hand, and the gentleman's face could use another pass. There are also a few disparities between the sharpness of faces and objects at various distances. Then again, I've seen worse in real, professional paintings, so that doesn't bother me quite as much.

The important part is that a learned a ton about how to do quality image manipulation using only the AI. Sure, I could have loaded up Photoshop or GIMP and probably had a better result in less time, but learning to do this through the AI feels like an accomplishment, and I'm already looking forward to the next project.

So how exactly does that work? Do you mask off a small area and tell it to draw a hand? I've never really bothered with inpainting so far.

# ? Oct 18, 2022 20:40

WhiteHowler: Apr 3, 2001; I'M HUGE!

mobby_6kl posted:

So how exactly does that work? Do you mask off a small area and tell it to draw a hand? I've never really bothered with inpainting so far.

I mask off an area and usually run with the original prompt. I will sometimes change the prompt slightly if I'm trying for a specific thing -- for example, I added "holding a large perfume bottle" to the main prompt for the inpaint of the front-left person's hand.

Telling it to just draw a hand ends up weird. With the denoising low enough and setting the mask content to "original", the AI will take what's already there and do its best to redraw. If there was already a hand, it will usually try to draw another hand. Often a better one. Sometimes a worse one. Occasionally a goddamned abomination that should not be seen by human eyes. Set expectations low and batch count high, and you'll get something good eventually.

The key to inpainting faces is using the "Inpaint at full resolution" checkbox. Stable Diffusion is terrible at drawing faces at medium range and beyond, but it's very good at drawing close-up faces. The "full resolution" setting makes it scale up your masked area to 512x512, do the rendering, then shrink it back down and paste it into the original masked area. Without this setting, it just tries to re-draw a face at medium range, and you get more awful faces.

# ? Oct 18, 2022 20:56

deep dish peat moss: Jul 27, 2006

using AI as concept art to make the drawing process a million times faster: day 4 or 5 or whatever

timelapse:

https://i.imgur.com/se3STUE.mp4

# ? Oct 18, 2022 21:24

mobby_6kl: Aug 9, 2009; by Fluffdaddy

So the prompt matrix doesn't work for negative prompts, which is a bummer. I did remove the tokens one by one (so no combinations) and... it's interesting. Basically they can have huge, unpredictable impact on the result.

So the first image has all the usual parts, "ugly", "deformed", etc. Each each image I then removed the prompt that's in the filename. 01630 removed the last negative prompt, which was "watermark". I then added "poorly drawn hands" back, and got a completely different image.

Whether or not these negative prompts actually helped in the first place is questionable of course, I just copied them from others and assumed they did something

Overall it just seems to have a pretty random effect, in a way that the normal prompt doesn't. I added "rain" and even "snow" and that changed the image less, and in predictable fashion, than adding "watermark" to the negative prompt.

In the end it probably doesn't matter too much since it's all a very random process and you're selecting the best image from dozen randomized versions. But at least in this particular case, it seemed to consistently ruin the styling on the cars which was a problem.

WhiteHowler posted:

I mask off an area and usually run with the original prompt. I will sometimes change the prompt slightly if I'm trying for a specific thing -- for example, I added "holding a large perfume bottle" to the main prompt for the inpaint of the front-left person's hand.

Telling it to just draw a hand ends up weird. With the denoising low enough and setting the mask content to "original", the AI will take what's already there and do its best to redraw. If there was already a hand, it will usually try to draw another hand. Often a better one. Sometimes a worse one. Occasionally a goddamned abomination that should not be seen by human eyes. Set expectations low and batch count high, and you'll get something good eventually.

The key to inpainting faces is using the "Inpaint at full resolution" checkbox. Stable Diffusion is terrible at drawing faces at medium range and beyond, but it's very good at drawing close-up faces. The "full resolution" setting makes it scale up your masked area to 512x512, do the rendering, then shrink it back down and paste it into the original masked area. Without this setting, it just tries to re-draw a face at medium range, and you get more awful faces.

Oooh, I see, thanks. Unfucking small faces in particular could be helpful.

# ? Oct 18, 2022 21:29

SubG: Aug 19, 2004; It's a hard world for little things.

mobby_6kl posted:

So I've been fighting with SD about applying a particular style to a subject which obviously wasn't trained in that style. I thought this wouldn't be really a problem, it clearly can combine stuff that's never been seen before, like drawings of spaceship by Leonardo Da Vinci, or anime Henry Kissinger, etc. But somehow it won't do an impressionist GT-R:

I've mad an off-hand post earlier with a similar example with a Miata before, and it's basically the same issue with it looking photorealistic and standing out too much.

I tried lowering the weight on the GTR:

And doing that while increasing the weight on Monet:

(left->right is more GTR, top->bottom is more Monet)

The issue is basically the same, it becomes a not-GTR before it's at all impressionist. Yet somehow it can do passable 50s car, which is definitely something Monet could've have painted. Any ideas? I mean other than a photoshop filter

In addition to decreasing attention on the car, try increasing attention of the style, e.g. by emphasising a specific impressionist:

That's:

webui posted:

parameters

painting of a [[gt-r]] on a street in paris, painting by ((pierre-august renoir)), oil on canvas, masterpiece, brushwork, muted colors, 8k
Negative prompt: photo, render
Steps: 40, Sampler: Euler a, CFG scale: 7, Seed: 363442244, Size: 896x512, Denoising strength: 0.7, First pass size: 704x448

It looks like this is one of those seeds where you can probably twiddle around with the number of steps, denoise strength, and/or throw in an extra seed and it'll still want to produce more or less the same image (some seeds seem much more "fragile" and change subject/focus/whatever with even minor tweaks), but I didn't really go digging around to optimise the image.

This pretty much always plops the Nissan in the centre of the image, because it's main subject, and you'll end up with a lot of things framed like product/beauty shots, presumably telling you what the training data looks like. If you want it to be a more organic part of the scene, try making the subject something else, and the GT-R just an additional element. E.g.:

webui posted:

parameters

painting of a street in paris with a nissan gt-r, painting by ((pierre-august renoir)), oil on canvas, masterpiece, brushwork, muted colors, 8k
Negative prompt: photo, render
Steps: 40, Sampler: Euler a, CFG scale: 7, Seed: 2724655492, Size: 896x512, Denoising strength: 0.7, First pass size: 704x448

webui posted:

parameters

painting of a street in paris with a nissan gt-r, painting by ((pierre-august renoir)), oil on canvas, masterpiece, brushwork, muted colors, 8k
Negative prompt: photo, render
Steps: 80, Sampler: Euler a, CFG scale: 7, Seed: 4120077182, Size: 576x896, Denoising strength: 0.7, First pass size: 448x640

If that's still too "GT-R" and not enough "Renoir" for you (too many hard lines, for example), you can just pass that to img2img and then inpaint over the problem section adjusting attention appropriately. Here's the car inpainted for 80 steps @ 0.2 denoise, with the prompt changed to "painting of [[nissan gt-r]], painting by ((pierre-august renoir)), oil on canvas, masterpiece, brushwork, muted colors, 8k":

# ? Oct 18, 2022 21:44

Mercury_Storm: Jun 12, 2003; *chomp chomp chomp*

((Intel)) Inside! branded (chili vomit) dump-truck filled with and dumping chili into a pool with people swimming in chili

# ? Oct 18, 2022 21:45

LifeSunDeath: Jan 4, 2007; still gay rights and smoke weed every day

Mercury_Storm posted:

((Intel)) Inside! branded (chili vomit) dump-truck filled with and dumping chili into a pool with people swimming in chili

NOW THIS...is a fetish

# ? Oct 18, 2022 21:47

EVIL Gibson: Mar 23, 2001; Internet of Things is just someone else's computer that people can't help attaching cameras and door locks to!; Switchblade Switcharoo

Mercury_Storm posted:

((Intel)) Inside! branded (chili vomit) dump-truck filled with and dumping chili into a pool with people swimming in chili

poo poo, you taught the AI of how to be a greasy fanboy

# ? Oct 18, 2022 23:00

Mercury_Storm: Jun 12, 2003; *chomp chomp chomp*

a dystopian post-apocalyptic lovecraftian ratty advertisement for a Facebook branded food processor that turns rats directly into fully cooked hamburgers in the style of H.R. Giger

# ? Oct 18, 2022 23:25

TheWorldsaStage: Sep 10, 2020

Mercury_Storm posted:

a dystopian post-apocalyptic lovecraftian ratty advertisement for a Facebook branded food processor that turns rats directly into fully cooked hamburgers in the style of H.R. Giger

Not gonna lie here, I'm ecstatic it glossed over your giger prompt. My mind however did not

# ? Oct 18, 2022 23:46

Mercury_Storm: Jun 12, 2003; *chomp chomp chomp*

TheWorldsaStage posted:

Not gonna lie here, I'm ecstatic it glossed over your giger prompt. My mind however did not

Adding Giger just tends to make things a lot creepier, and there are a bunch I didn't post because they were way too creepy. For example!

# ? Oct 19, 2022 00:11

Nice Van My Man: Jan 1, 2008

So much horror, I decided to try to make something a little nicer:

"Thomas Kinkade cute xenomorphs scene award winning painting"

Awww, look at those cute little guys. And surprisingly consistent between pictures. I feel like these pictures are showing their lifecycle.

# ? Oct 19, 2022 00:31

TheWorldsaStage: Sep 10, 2020

I love this thread

# ? Oct 19, 2022 00:37

WhiteHowler: Apr 3, 2001; I'M HUGE!

An apple pie in the style of Thomas Kinkade and Salvador Dali

# ? Oct 19, 2022 00:48

deep dish peat moss: Jul 27, 2006

Nice Van My Man posted:

So much horror, I decided to try to make something a little nicer:

"Thomas Kinkade cute xenomorphs scene award winning painting"

Awww, look at those cute little guys. And surprisingly consistent between pictures. I feel like these pictures are showing their lifecycle.

This gave me an idea.

"A Xenomorph by H.R. Giger riding a bicycle through a whimsical forest by Lisa Frank" (midjourney)

# ? Oct 19, 2022 01:10

Sedgr: Sep 16, 2007; Neat!

most at the face-hugger bird/bat.

# ? Oct 19, 2022 01:15

Commie Lasorda: May 15, 2009; IT'S CLOBBERIN' TIME!

"Statue of Christopher Hewitt sitting on a ball commemorating the time when he crushed his own testicle during a table read for Mr. Belvedere"

# ? Oct 19, 2022 02:23

frumpykvetchbot: Feb 20, 2004; PROGRESSIVE SCAN; Upset Trowel

Mercury_Storm posted:

a dystopian post-apocalyptic lovecraftian ratty advertisement for a Facebook branded food processor that turns rats directly into fully cooked hamburgers in the style of H.R. Giger

Delicious.

You gave birth to this beautiful thing too. I'd legit somehow find a way to respect Zuck if he adopted this as his metahorizonverse avatar.

Only registered members can see post attachments!

# ? Oct 19, 2022 03:11

KwegiboHB: Feb 2, 2004; nonconformist art brut
Negative prompt: amenable, compliant, docile, law-abiding, lawful, legal, legitimate, obedient, orderly, submissive, tractable
Steps: 32, Sampler: DPM++ 2M Karras, CFG scale: 11, Seed: 520244594, Size: 512x512, Model hash: 99fd5c4b6f, Model: seekArtMEGA_mega20

Negative prompt: Fusion
Steps: 32, Sampler: Euler a, CFG scale: -15.0, Seed: 1185003645, Size: 512x512, Model hash: 45dee52b

# ? Oct 19, 2022 03:30

Tunicate: May 15, 2012

Do negative prompt and negative cfg cancel out

# ? Oct 19, 2022 04:25

KwegiboHB: Feb 2, 2004; nonconformist art brut
Negative prompt: amenable, compliant, docile, law-abiding, lawful, legal, legitimate, obedient, orderly, submissive, tractable
Steps: 32, Sampler: DPM++ 2M Karras, CFG scale: 11, Seed: 520244594, Size: 512x512, Model hash: 99fd5c4b6f, Model: seekArtMEGA_mega20

I have absolutely no idea.

# ? Oct 19, 2022 04:26

Tunicate: May 15, 2012

Can you try running with positive on both? It is a cool image and it would be neat id doublenegatives were a secret trick

# ? Oct 19, 2022 04:27

KwegiboHB: Feb 2, 2004; nonconformist art brut
Negative prompt: amenable, compliant, docile, law-abiding, lawful, legal, legitimate, obedient, orderly, submissive, tractable
Steps: 32, Sampler: DPM++ 2M Karras, CFG scale: 11, Seed: 520244594, Size: 512x512, Model hash: 99fd5c4b6f, Model: seekArtMEGA_mega20

Fusion
Negative prompt: Fusion
Steps: 32, Sampler: Euler a, CFG scale: -15.0, Seed: 1185003645, Size: 512x512, Model hash: 45dee52b

If you same word Prompt/Negative, it does not matter -15 or 15 classifier, it'll be the same exact image.

I ran a few mixes of No prompt/Negative prompt, I'll make an imgur gallery in a minute.

# ? Oct 19, 2022 04:45

KwegiboHB: Feb 2, 2004; nonconformist art brut
Negative prompt: amenable, compliant, docile, law-abiding, lawful, legal, legitimate, obedient, orderly, submissive, tractable
Steps: 32, Sampler: DPM++ 2M Karras, CFG scale: 11, Seed: 520244594, Size: 512x512, Model hash: 99fd5c4b6f, Model: seekArtMEGA_mega20

https://imgur.com/a/YtPc5yN
Am I a data scientist now?

# ? Oct 19, 2022 05:18

Adbot: ADBOT LOVES YOU

# ? May 30, 2024 20:19

BrainDance: May 8, 2007; Disco all night long!

Wheany posted:

You just assign the string "cuda:0" or "cpu" to the variable named device you never use that variable for anything, on the next row you call the funtion pipeline and you assign the value 0 to the named parameter called device

Thanks, this was really useful and I was able to improve mine a lot. I'm still trying to figure out how this all works and there's a lot in the gpt-neo docs that I am kinda lost on.
So, I was using pipeline but it looks like GPTNeoForCausalLM takes its place? One thing I noticed is that I'm getting very different results now, less these long essays and more scenarios which is cool but I wish I understood exactly why so I could fine tine things even more.

I know the thread is mostly about AI generated images but I think some of this stuff is just as interesting. Weird non-existent fantasy scenarios, conversations that are just off. Making a conversation in gpt-neo then making a picture in SD to go with it.

# ? Oct 19, 2022 06:24

The Something Awful Forums > Main > General Bullshit > AI Art: It is criminal to not post your prompt

«‹›435 »