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

«‹›434 »

Wheany: Mar 17, 2006; Spinya^{ha^{ha^haha}ha}ha_{ha_{ha_haha}ha}ha!; Doctor Rope

Cable Guy posted:

DallE-2 has no idea...

Lol

"how do attach trunk :shrug:

# ? Aug 29, 2022 10:53

Adbot: ADBOT LOVES YOU

# ? May 29, 2024 01:24

mobby_6kl: Aug 9, 2009; by Fluffdaddy

These are honestly pretty incredible. As I mentioned before, I completely failed at making a raven Mona Lisa and only like 1 in 10 pig cops turned out ok. Though I definitely need to test out adjusting the weights.

# ? Aug 29, 2022 11:47

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

BARONS CYBER SKULL posted:

grabbed a incantation from someone else's spellbook and tweaked it somewhat
code:
Ultra realistic, portrait, a beautiful woman, robot cloak, long hair, intricate, elegant,  highly detailed, digital painting, artstation, concept art, smooth, sharp focus, illustration, art by artgerm and greg rutkowski and alphonse mucha, by guweiz and wlop and artgerm, by Akihiko Yoshida, Mark Arian, Sakimi Chan, Mogoon


Steps: 20, Sampler: k_euler_a, CFG scale: 12, RealESRGAN_x4plus
it spits out extremely varied images which are mostly cohesive, it can also ALMOST get the hands right:

other examples

Its thanks to incantations like this one that make me more experimental, little by little. This is making me realise just how much I'm missing out on.

I'm trying one from earlier up thread had had some fun results;
"A wizard woman walks towards the moonlight. It is a starry sky at night. Synthwave. Digital art. Trending on Artstation. Sharp focus."

The accidental duplication effect from rendering in 16:9 is really fun here. Now your art subject gets a free friend on their mystical synthwave journey!

I guess the next step is to fine tune the upscaler. I'm not getting the results I want yet. They're not bad but they're still a bit fuzzier than I hoped. Maybe chain two AI upscales together or something. :v:

Mindblast fucked around with this message at 12:58 on Aug 29, 2022

# ? Aug 29, 2022 12:47

Mozi: Apr 4, 2004; Forms change so fast
Time is moving past
Memory is smoke
Gonna get wider when I die; Nap Ghost

What I really wish Midjourney had was an ability to mark out a certain section of the image and re-do only that part with a specific prompt.

# ? Aug 29, 2022 13:57

Rinkles: Oct 24, 2010; What I'm getting at is...
Do you feel the same way?

mobby_6kl posted:

These are honestly pretty incredible. As I mentioned before, I completely failed at making a raven Mona Lisa and only like 1 in 10 pig cops turned out ok. Though I definitely need to test out adjusting the weights.

to the chagrin of certain communities I'm sure, SD seems pretty reluctant to put animal faces on people in general, which is unexpected given the internet we know.

# ? Aug 29, 2022 15:07

Wheany: Mar 17, 2006; Spinya^{ha^{ha^haha}ha}ha_{ha_{ha_haha}ha}ha!; Doctor Rope

Rinkles posted:

to the chagrin of certain communities I'm sure, SD seems pretty reluctant to put animal faces on people in general, which is unexpected given the internet we know.

Consulting CLIP interrogator, I can create animal-faced people pretty easily:

but the phrase "mona lisa" just annihilates any other parts of the prompt. No matter what I do, I either manage to make something that is not-mona-lisa, or is just mona lisa that looks like she smelled a fart or a mona lisa that saw you eat a booger. It also might be a mona lisa that uses wild colors or thick inky brush strokes, but it's very difficult to replace mona lisa's face while still retaining the "mona lisaness" of the rest of the picture.

# ? Aug 29, 2022 16:29

KakerMix: Apr 8, 2004; 8.2 M.P.G.

Wheany posted:

Consulting CLIP interrogator, I can create animal-faced people pretty easily:

but the phrase "mona lisa" just annihilates any other parts of the prompt. No matter what I do, I either manage to make something that is not-mona-lisa, or is just mona lisa that looks like she smelled a fart or a mona lisa that saw you eat a booger. It also might be a mona lisa that uses wild colors or thick inky brush strokes, but it's very difficult to replace mona lisa's face while still retaining the "mona lisaness" of the rest of the picture.

You can do img2img and use the Mona Lisa itself as the base image but don't invoke it in the actual prompt, play with the settings and you should be able to get the feel of it but not have to deal with the prompt did directly like you have been.

# ? Aug 29, 2022 16:53

Mozi: Apr 4, 2004; Forms change so fast
Time is moving past
Memory is smoke
Gonna get wider when I die; Nap Ghost

midjourney really does not want to put a pirhana face on a seagull but this is neat i guess

# ? Aug 29, 2022 17:11

Dr. Fraiser Chain: May 18, 2004; Redlining my shit posting machine

nightwisher posted:

Jumped into DreamStudio and got this as my first output, no idea where the gold tentrils factored in from but I really like it. Ran it through Cupscale and it came out even nicer:

"vast landscape filled with forests and rivers and mountains and clouds, hyper realistic, detailed, intricate, night lighting"

It's drawing from car headlights probably

# ? Aug 29, 2022 17:16

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

Mozi posted:

midjourney really does not want to put a pirhana face on a seagull but this is neat i guess

wow goddamn that's cool.

# ? Aug 29, 2022 17:17

AARD VARKMAN: May 17, 1993

Glad Midjourney is testing again.

Some dog astronauts

A shopping mall turning in to an eldritch horror, upscaled

# ? Aug 29, 2022 17:21

Moongrave: Jun 19, 2004; Finally Living Rent Free

BARONS CYBER SKULL posted:

grabbed a incantation from someone else's spellbook and tweaked it somewhat
code:
Ultra realistic, portrait, a beautiful woman, robot cloak, long hair, intricate, elegant,  highly detailed, digital painting, artstation, concept art, smooth, sharp focus, illustration, art by artgerm and greg rutkowski and alphonse mucha, by guweiz and wlop and artgerm, by Akihiko Yoshida, Mark Arian, Sakimi Chan, Mogoon


Steps: 20, Sampler: k_euler_a, CFG scale: 12, RealESRGAN_x4plus
it spits out extremely varied images which are mostly cohesive, it can also ALMOST get the hands right:

other examples

running an absolute shitload of this and fiddling with some words got me two really good ones:

# ? Aug 29, 2022 23:12

MartingaleJack: Aug 26, 2004; I'll split you open and I don't even like coconuts.

Burger Queen

MartingaleJack fucked around with this message at 23:55 on Aug 29, 2022

# ? Aug 29, 2022 23:32

VectorSigma: Jan 20, 2004; Transform
and
Freak Out

"you ate the whole thing? how was your walk?"

# ? Aug 29, 2022 23:55

Mozi: Apr 4, 2004; Forms change so fast
Time is moving past
Memory is smoke
Gonna get wider when I die; Nap Ghost

existential dreads

# ? Aug 30, 2022 01:08

Vlaphor: Dec 18, 2005; Lipstick Apathy

I done made robot bears.

# ? Aug 30, 2022 03:10

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

BARONS CYBER SKULL posted:

grabbed a incantation from someone else's spellbook and tweaked it somewhat
code:
Ultra realistic, portrait, a beautiful woman, robot cloak, long hair, intricate, elegant,  highly detailed, digital painting, artstation, concept art, smooth, sharp focus, illustration, art by artgerm and greg rutkowski and alphonse mucha, by guweiz and wlop and artgerm, by Akihiko Yoshida, Mark Arian, Sakimi Chan, Mogoon


Steps: 20, Sampler: k_euler_a, CFG scale: 12, RealESRGAN_x4plus
it spits out extremely varied images which are mostly cohesive, it can also ALMOST get the hands right:

other examples

I accidentally copied your post text into the prompt too and now its pretty consistently generating TCG cards

code:

grabbed a incantation from someone else's spellbook and tweaked it somewhat      code:      Ultra realistic, portrait, a beautiful woman, robot cloak, long hair, intricate, elegant,  highly detailed, digital painting, artstation, concept art, smooth, sharp focus, illustration, art by artgerm and greg rutkowski and alphonse mucha, by guweiz and wlop and artgerm, by Akihiko Yoshida, Mark Arian, Sakimi Chan, Mogoon
Steps: 50, Sampler: k_euler_a, CFG scale: 12, Seed: 153222442

Warning: too many input tokens; some (21) have been truncated:
lop and artgerm , by akihiko yoshida , mark arian , sakimi chan , mogoon

# ? Aug 30, 2022 03:14

Chronojam: Feb 20, 2006; This is me on vacation in Amsterdam
Never be afraid of being yourself!

IShallRiseAgain posted:

I accidentally copied your post text into the prompt

This is an exciting new genre, and could certainly fuel a resurgence in CYOAs

# ? Aug 30, 2022 04:32

VectorSigma: Jan 20, 2004; Transform
and
Freak Out

been playing with the upscaler

https://i.imgur.com/rB4aKCw.mp4

# ? Aug 30, 2022 04:47

Rutibex: Sep 9, 2001; by Fluffdaddy

IShallRiseAgain posted:

I accidentally copied your post text into the prompt too and now its pretty consistently generating TCG cards

code:

grabbed a incantation from someone else's spellbook and tweaked it somewhat      code:      Ultra realistic, portrait, a beautiful woman, robot cloak, long hair, intricate, elegant,  highly detailed, digital painting, artstation, concept art, smooth, sharp focus, illustration, art by artgerm and greg rutkowski and alphonse mucha, by guweiz and wlop and artgerm, by Akihiko Yoshida, Mark Arian, Sakimi Chan, Mogoon
Steps: 50, Sampler: k_euler_a, CFG scale: 12, Seed: 153222442

Warning: too many input tokens; some (21) have been truncated:
lop and artgerm , by akihiko yoshida , mark arian , sakimi chan , mogoon

This is awesome, I guess designing a trading card game is actually even easier with these AI. they can just design the layout of a card. the one you made there would be a great template

# ? Aug 30, 2022 04:58

Ima Computer: Oct 28, 2007; Stop all the downloading!

Help computer.

The hlky fork of Stable Diffusion really unlocks some cool workflow potential when mixed with a little bit of manual photoshopping/compositing.

Here's an example of a quick walk-through of my process of turning pancakegoon into a body horror spider monster with a decent face:

Ima Computer fucked around with this message at 06:54 on Aug 30, 2022

# ? Aug 30, 2022 06:43

Rinkles: Oct 24, 2010; What I'm getting at is...
Do you feel the same way?

I was intrigued by the "public link" message that appears after launching the webui, and I got it to work, however that seems to host the page on gradio's servers (#####.gradio.app). It says the link expires in 72 hours.

I don't mind restarting every day if it means I can share access, but I'm not sure if it's indefinitely repeatable like that.

Is there an easy way to just host it on the local network?

# ? Aug 30, 2022 08:26

Rinkles: Oct 24, 2010; What I'm getting at is...
Do you feel the same way?

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

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

Rinkles fucked around with this message at 09:54 on Aug 30, 2022

# ? Aug 30, 2022 09:24

Crapple!: Nov 1, 2019

Loving the new test branch of Midjourney. It's way better at anatomy and context. Not quite as creative so you need to add more details. I used it to churn out a few school yearbook photos

# ? Aug 30, 2022 09:41

Moongrave: Jun 19, 2004; Finally Living Rent Free

The real power of these AIs is unlocked by finding the right words to scream at the machine to make it make what you mean, based on the training model they used and the tags the images got

Also �greg rutkowski� as a prompt additive has been tested and literally always generates better looking results for every sampler and every prompt. He is the salt of the AI image cooking.

# ? Aug 30, 2022 10:13

GABA ghoul: Oct 29, 2011

BARONS CYBER SKULL posted:

The real power of these AIs is unlocked by finding the right words to scream at the machine to make it make what you mean, based on the training model they used and the tags the images got

Also �greg rutkowski� as a prompt additive has been tested and literally always generates better looking results for every sampler and every prompt. He is the salt of the AI image cooking.

Someone should make an AI that generates good prompts from just a description of the image.

# ? Aug 30, 2022 10:18

strange feelings re Daisy: Aug 2, 2000

Yeah adding Greg Rutkowski really does make a difference vs. default. Adding stuff like "trending on artstation, 4K, award winning, clean brushstrokes, highly detailed" makes a difference too.

# ? Aug 30, 2022 10:20

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

BARONS CYBER SKULL posted:

running an absolute shitload of this and fiddling with some words got me two really good ones:

Did you turn "robot cloak" into robot mask or something?

I also did some editing around of your prompt and simply switching out anything gender related and adding "samus" has interesting results. Lots of Samus clones in power armor.

Btw I noticed your original prompt is slightly larger than Stable Diffusion can handle. Some words fall off at the end. Like, an entire artist is not taken into account. Which is alright I guess I mean there's like nine listed already lol

# ? Aug 30, 2022 10:38

Moongrave: Jun 19, 2004; Finally Living Rent Free

Mindblast posted:

Did you turn "robot cloak" into robot mask or something?

I also did some editing around of your prompt and simply switching out anything gender related and adding "samus" has interesting results. Lots of Samus clones in power armor.

Btw I noticed your original prompt is slightly larger than Stable Diffusion can handle. Some words fall off at the end. Like, an entire artist is not taken into account. Which is alright I guess I mean there's like nine listed already lol

I think the top one was �robot gasmask� and the bottom was �robot kitsune mask�?

Also found one of my older ones from the SD discord:

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

# ? Aug 30, 2022 10:53

LordAdakos: Sep 1, 2009

GABA ghoul posted:

Someone should make an AI that generates good prompts from just a description of the image.

Oh Hello, I think you are looking for this: https://colab.research.google.com/github/pharmapsychotic/clip-interrogator/blob/main/clip_interrogator.ipynb

You give it a picture and it gives you :words:

# ? Aug 30, 2022 12:24

Rutibex: Sep 9, 2001; by Fluffdaddy

LordAdakos posted:

Oh Hello, I think you are looking for this: https://colab.research.google.com/github/pharmapsychotic/clip-interrogator/blob/main/clip_interrogator.ipynb

You give it a picture and it gives you

no what he wants is an AI which takes out lovely natural language prompts and converts them into weirdo AI focused prompts full of "high detail, hypermaxiums, drawn by Oscar the grouch"

# ? Aug 30, 2022 12:30

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

Ima Computer posted:

The hlky fork of Stable Diffusion really unlocks some cool workflow potential when mixed with a little bit of manual photoshopping/compositing.

Here's an example of a quick walk-through of my process of turning pancakegoon into a body horror spider monster with a decent face:

Good lord

# ? Aug 30, 2022 14:17

pixaal: Jan 8, 2004; All ice cream is now for all beings, no matter how many legs.

Ima Computer posted:

The hlky fork of Stable Diffusion really unlocks some cool workflow potential when mixed with a little bit of manual photoshopping/compositing.

Here's an example of a quick walk-through of my process of turning pancakegoon into a body horror spider monster with a decent face:

The foot is merged with the table

I think this goon has 3 arms and 3 feet at least.

# ? Aug 30, 2022 15:44

Wheany: Mar 17, 2006; Spinya^{ha^{ha^haha}ha}ha_{ha_{ha_haha}ha}ha!; Doctor Rope

I'm not sure if "greg rutkowski" is some exlusive magic incantation, seems like just about any artist name makes the output better. You can browse artists here: https://sgreens.notion.site/sgreens/4ca6f4e229e24da6845b6d49e6b08ae7

edit: And in general stable diffusion loving loves :words:

(but, only up to 77 tokens, where every word is usually 1-2 tokens, and every piece of puncuation is also a token)

Wheany fucked around with this message at 20:37 on Aug 30, 2022

# ? Aug 30, 2022 20:32

Wheany: Mar 17, 2006; Spinya^{ha^{ha^haha}ha}ha_{ha_{ha_haha}ha}ha!; Doctor Rope

For example this is how it tokenizes my last message:

prompt length:126

[<|startoftext|>; i; 'm; not; sure; if; "; greg; rut; kowski; "; is; some; ex; lu; sive; magic; incan; tation; ,; seems; like; just; about; any; artist; name; makes; the; output; better; .; you; can; browse; artists; here; :; https; ://; s; greens; .; notion; .; site; /; s; greens; /; 4; ca; 6; f; 4; e; 2; 2; 9; e; 2; 4; da; 6; 8; 4; 5; b; 6; d; 4; 9; e; 6; b; 0; 8; ae; 7; edit; :; and; in; general; stable; dif; fusion; fu; cking; loves; :; words; :; (; but; ,; only; up; to; 7; 7; tokens; ,; where; every; word; is; usually; 1; -; 2; tokens; ,; and; every; piece; of; pun; cu; ation; is; also; a; token; ); <|endoftext|>]

(every token in the list is separated with a semicolon and a space)

# ? Aug 30, 2022 20:50

Moongrave: Jun 19, 2004; Finally Living Rent Free

Wheany posted:

I'm not sure if "greg rutkowski" is some exlusive magic incantation, seems like just about any artist name makes the output better. You can browse artists here: https://sgreens.notion.site/sgreens/4ca6f4e229e24da6845b6d49e6b08ae7

edit: And in general stable diffusion loving loves (but, only up to 77 tokens, where every word is usually 1-2 tokens, and every piece of puncuation is also a token)

He�s specifically both prolific enough, good enough at both portraits and landscapes, and generic enough that he doesn�t adversely influence the style or Image too much, while adding extra detail and �quality� is why he�s good

Hence me calling him salt

# ? Aug 30, 2022 22:09

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

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.

# ? Aug 30, 2022 22:15

Tunicate: May 15, 2012

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 pretty impressive. how'd you get img2img set up?

# ? Aug 30, 2022 22:51