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
Tarkus
Aug 27, 2000

Yeah, it's just bad photoshop, maybe moved a hand or took the hands from another frame.

Adbot
ADBOT LOVES YOU

Swagman
Jun 10, 2003

Yes...all was once again peaceful in River City.






























Roman
Aug 8, 2002

Character consistency was just released in Midjourney. It works best with referencing Midjourney images, not photos or images from other sources, so it's not really relevant for me and I'll still have to use Insightface.

https://twitter.com/WorldEverett/status/1767319696804606291

Soulhunter
Dec 2, 2005

Midjourney says in their documentation that characters generated in Midjourney work better than those from other sources, so working with that in mind, I made this generic gold-glasses steampunk cowboy man to test out the cref command:


Cowboy riding a black horse through the center of a muddy main street in a cowboy era town, "MURPHY'S SALOON" behind him as he passes by, side angle shot --ar 16:9 --cref https://s.mj.run/AAMA-1tXuD8 with a --cw value of 75, then 100


Cowboy holding up dual techno pistols that look like nerf guns, In the style of Jojo's Bizarre Adventure --niji 6 --ar 16:9 --cref https://s.mj.run/AAMA-1tXuD8 with a --cw value of 75, then 100


Mortal Kombat, high quality character render, side profile face-off image --cref https://s.mj.run/AAMA-1tXuD8 --cw 100 --ar 2:3


I started to hit a wall with the cref overriding the style keywords or muddying them a bit:
Thrilled man riding a unicorn, a trail of rainbows flowing out from behind the unicorn as they sail through the sky over Pittsburgh, overjoyed, arms in the air in celebration, cartoon style image, kawaii, superdeformed in the style of powerpuff girls --ar 16:9 --cref https://s.mj.run/AAMA-1tXuD8 --cw 75


Tweaking the same prompt to include a sref helps, but the cw value staying high seems to still inform the style of the image:
Thrilled man riding a unicorn, a trail of rainbows flowing out from behind the unicorn as they sail through the sky over Pittsburgh, overjoyed, arms in the air in celebration, cartoon style image, kawaii, superdeformed, in the style of powerpuff girls, minimalist cartoon style with thick line work --ar 16:9 --cref https://s.mj.run/AAMA-1tXuD8 --cw 100 --sref https://s.mj.run/Ou788WS3azo --sw 100


A better overall idea of what the --cw does while keeping the --sw low and changing the order of the prompt keywords slightly:
cartoon style image, kawaii, superdeformed chibi style, powerpuff girls, minimalist cartoon style with thick line work, Thrilled man riding a unicorn, a trail of rainbows flowing out from behind the unicorn as they sail through the sky over Pittsburgh, overjoyed, arms in the air in celebration --ar 16:9 --cref https://s.mj.run/AAMA-1tXuD8--sref https://s.mj.run/Ou788WS3azo --sw 10 with --cw 0, 25, and 80, in that order:


Now, with all that said and done... I also tried some pre-made assets with a couple methods. Taking the generic model from RuneScape seen here, results were kind of poo poo, as Midjourney said would be the case.


In this prompt, it kept the 'bald man with goatee', but no color scheme for the clothes at all, despite maxing out the cw value.
Man cutting logs next to a friendly beaver that's nibbling on a maple log, medieval setting, minimalist renaissance painting --ar 3:2 --cref https://s.mj.run/AAGg9wxGgOg --cw 100


To counteract this problem, I used the generic model linked in a regular prompt to make a Midjourney asset with the proper color scheme. Pretty close.
https://s.mj.run/AAGg9wxGgOg Generic bald man wearing a yellow shirt and green pants with a red belt and silver bracelets, the man has a trimmed goatee --ar 2:3 --stylize 10


Throwing the new Midjourney-fied asset back into the blender with a similar prompt again yields better outfit / character consistency. Kind of wonder if they're doing some funky encoding with the images on the backend. In any case, Midjourney still doesn't seem to totally understand what a hatchet looks like.
Man swinging a hatchet with a teal metal blade at a gangly old overgrown oak tree, a friendly beaver sits on a stump nearby nibbling on a maple log, medieval setting, minimalist renaissance painting, caricature --ar 3:2 --cref https://s.mj.run/6IN_SjA9GAY --cw 100


Different background settings, clothing preserved:


Curiously, the cref did seem to work fairly acceptably with a sketch as a reference. Breaking Johnny Five-Aces out of retirement here.


Steampunk man sitting at a table, POV over his shoulder, the man is holding five playing cards in his hand, all of them are Aces of Spades, cinematic --ar 16:9 and Steampunk, man sitting at a table holding five playing cards in one hand and a beer in the other hand, the cards are all Aces, cinematic, one of the man's legs is propped up on the table, in the style of a black and white pencil sketch --ar 16:9 --cref https://s.mj.run/tnmzChKi1gc --cw 100


E: Couple more to test the "using non-MJ assets has bad results" theory:

Using headshots of Natalie Portman and Keira Knightley as dual crefs on the same prompt:
Woman wearing a red battlearmor suit, glamour pose while a battle against Zerglings and Warhammer Marines rages in the background --cref https://s.mj.run/2yxoGsTOzkI https://s.mj.run/wD8dIOhECOg --cw 5 --ar 2:3 --w 100


Using a /blend of the same two Portman/Knightley headshots to generate a MJ asset, then using the MJ asset as the cref for the same prompt above:


I think using the headshots as separate crefs actually looks better, but the blended one does seem more consistent in appearance. Either way, didn't seem to have an issue with using celeb photos as a starting point.

Last cref:






Soulhunter fucked around with this message at 15:40 on Mar 12, 2024

hydroceramics
Jan 8, 2014

To be clear, I just got a chuckle from this. There is no other intended commentary.

null_pointer
Nov 9, 2004

Center in, pull back. Stop. Track 45 right. Stop. Center and stop.

Huh. Looks like character reference is as wonky as style reference is, still. You need a pretty low value (probably around 20) otherwise the sref starts to bleed into the prompt. Unfortunate, but I'm still curious to see how it works out in practice.

If the next thing they work on is a pose reference, they would at least have the three tools I need in alpha.

Swagman
Jun 10, 2003

Yes...all was once again peaceful in River City.






























XYZAB
Jun 29, 2003

HNNNNNGG!!

Soulhunter posted:

Curiously, the cref did seem to work fairly acceptably with a sketch as a reference. Breaking Johnny Five-Aces out of retirement here.


Steampunk man sitting at a table, POV over his shoulder, the man is holding five playing cards in his hand, all of them are Aces of Spades, cinematic --ar 16:9 and Steampunk, man sitting at a table holding five playing cards in one hand and a beer in the other hand, the cards are all Aces, cinematic, one of the man's legs is propped up on the table, in the style of a black and white pencil sketch --ar 16:9 --cref https://s.mj.run/tnmzChKi1gc --cw 100

I unironically would love to see an AI-storyboarded, human-written Johnny Five-Aces spec script turned into a feature film in this style.

Cabbages and VHS
Aug 25, 2004

Listen, I've been around a bit, you know, and I thought I'd seen some creepy things go on in the movie business, but I really have to say this is the most disgusting thing that's ever happened to me.

ymgve posted:

no thats normal british teeth







Soulhunter
Dec 2, 2005


null_pointer
Nov 9, 2004

Center in, pull back. Stop. Track 45 right. Stop. Center and stop.

Update: MidJourney's character reference variable with a weight of 0 seems to work surprisingly well. Combined with a style reference weight of around 10 to 20, I'm finding prompts "just work" with far less finagling than my previous efforts. I'm running a bunch of tests, but so far, I find myself saying "yeah, that looks a lot more like him/her" than before.

The Demilich
Apr 9, 2020

The First Rites of Men Were Mortuary, the First Altars Tombs.



Someone recreate a hyperrealistic version of the big book of British smiles

TIP
Mar 21, 2006

Your move, creep.



The Demilich posted:

Someone recreate a hyperrealistic version of the big book of British smiles









Sab669
Sep 24, 2009

Why must you turn this thread into a house of lies?

pixaal
Jan 8, 2004

All ice cream is now for all beings, no matter how many legs.


Google ImageFX from TestKitchen: (escape phrase for Trump of choice) with DD teeth


e: someone got a tape worm in their brain from undercooked bacon so I made this

pixaal fucked around with this message at 17:29 on Mar 14, 2024

Small Strange Bird
Sep 22, 2006

Merci, chaton!
I tried to use Bing to generate a pic of Elon Musk based on this pic from the GBS Musk thread:

with the prompt starting "rich white privileged South African age 50 with small half-closed eyes & receding hairline" and trying to describe everything else the best I could. Based on the results, I can only conclude that Bing thinks white South Africans are all mutants or Gollums.


Count Roland
Oct 6, 2013

Small Strange Bird posted:

I tried to use Bing to generate a pic of Elon Musk based on this pic from the GBS Musk thread:



Oh god he's a goon. No wonder he's so insufferable

Soulhunter
Dec 2, 2005

Bunch of Austin Powers stuff:




The newest horror kdrama coming to streaming services, Aporkalypse


Bunch of Great Deku Tree variants:




Couple updates to the "Right to Bear Arms" prompts I did a while back in the thread to test the new sref update. I think sref 1.0 was actually better at capturing the style of sketches, personally. I'd say the quality is better on these overall, but quality isn't exactly what I was going for with the style.


Red Carpet Bojack with Bojack Horseman as the cref:


Junji Beavo



Hide the Work Injury Harold


Do Not Touch The Marbles


Dissected human nervous system cref + Escher stairs sref:


Human musculature + dramatic pose for a novel cover + squares:


Sab669 posted:

Where's Felicity Shagwell
vvv she was replaced in this film with a new character, as is tradition:

Soulhunter fucked around with this message at 17:19 on Mar 15, 2024

Sab669
Sep 24, 2009

Soulhunter posted:

Bunch of Austin Powers stuff:

Where's Felicity Shagwell

Swagman
Jun 10, 2003

Yes...all was once again peaceful in River City.






















frumpykvetchbot
Feb 20, 2004

PROGRESSIVE SCAN
Upset Trowel

Small Strange Bird posted:

Based on the results, I can only conclude that Bing thinks white South Africans are all mutants or Gollums.

these should be default avs.

feedmyleg
Dec 25, 2004

Legit solid stuff, you discovered a cool little fantasy world.

Swagman
Jun 10, 2003

Yes...all was once again peaceful in River City.

a couple of these I would expect to be ambushed and brutalized by in elden ring

TremorX
Jan 19, 2001

All Hail Big Hairy Mike

Soulhunter posted:

vvv she was replaced in this film with a new character, as is tradition:


That's Ella Purnell 100%

PaleFigure
Sep 3, 2007

the other white meat
The party had travelled far into the Forest of Nightmares, and many foes had fallen before them, but finally their quest reached it's end. Their foe lay just ahead, and this night would see an end to its reign of terror.

The mighty Barbarian roared his battle cry and readied his sword.


The stealthy Rogue melted into the shadows, his poisoned blade glinting in the moonlight.


The pious Cleric muttered in fervent prayer, his blessed hammer ready to smite evil.


The studious Wizard spoke words of ancient and arcane power, his staff ready to channel his will.


The stalwart Ranger nocked an arrow, his focus sharp and his gaze clear and bright.


And before them, the foul creature of darkness that they had been hunting for days waited, its expression implacable, its thoughts unknowable: Ysobel the Un-Potty-Trained!


Yeah so long story short it was a total wipe, the Cleric was a pubby who went AFK when his mum said no more internet for tonight, the Barbarian screwed the pull and the Wizard drew aggro, and the Ranger was specced for crowd control for some god-knows-why reason. Anyway, they said they'd try again in two weeks when they all get back from bible camp.

Swagman
Jun 10, 2003

Yes...all was once again peaceful in River City.
you just know the ranger is going to roll need on any melee weapon drops




































snorch
Jul 27, 2009







e:






e2:
I was trying SDXL with a null prompt, and it tends mostly towards birds and flowers in an oil painty style but every once in a while I get an oddball like this fellow:

snorch fucked around with this message at 07:10 on Mar 19, 2024

PaleFigure
Sep 3, 2007

the other white meat
Staying on the beef train (and with deepest apologies to the estate of Dr Seuss):

Soulhunter
Dec 2, 2005

You've heard of Goosebumps, but have you heard of BooseGumps?


alternates:



broken physics outtakes:


e: You ain't got no legs, Lt. Dan!


Soulhunter fucked around with this message at 19:17 on Mar 19, 2024

LifeSunDeath
Jan 4, 2007

still gay rights and smoke weed every day

Soulhunter posted:

You've heard of Goosebumps, but have you heard of BooseGumps?


alternates:



broken physics outtakes:


lol nice

naem
May 29, 2011

life is like a chock of boss-let’s

snorch
Jul 27, 2009
I'd watch any one of those.

Also PaleFigure and swagman what are y'all using to get those images? They seem unusually crisp and coherent, or is it just the Art of the Prompt at work? I get decent results with SDXL but I'd say it only ever reaches about 80% of that fidelity.

dumb.
Apr 11, 2014

-=💀=-

snorch posted:

I'd watch any one of those.

Also PaleFigure and swagman what are y'all using to get those images? They seem unusually crisp and coherent, or is it just the Art of the Prompt at work? I get decent results with SDXL but I'd say it only ever reaches about 80% of that fidelity.



I've found lowering the CFG scale helps.

Sab669
Sep 24, 2009







as prompted to me from somewhere in Automotive Insanity :v:

LifeSunDeath
Jan 4, 2007

still gay rights and smoke weed every day

Sab669 posted:







as prompted to me from somewhere in Automotive Insanity :v:

LMAO

drat I just tried to create something similar and got blocked on everything, water pipe, 420, bong, wtf is happening.

Sab669
Sep 24, 2009

Weird, waterpipe has been a reliable workaround to "bong" for me. My various prompts were:

- a car dashboard with an illuminated symbol that looks like a water pipe
- a car dashboard with an illuminated symbol that looks like a waterpipe for smoking
- a car dashboard with an illuminated symbol that looks like a stoner's waterpipe
- something with "bong" was blocked
- a car dashboard with an illuminated symbol that looks like a stoner's pipe and lighter
- something with "420" was blocked
- a car dashboard with an illuminated symbol that looks like a stoner's pipe and lighter, odometer has numbers 4, 2, and 0
- a car dashboard with an illuminated symbol that looks like a stoner's pipe and lighter, odometer reads 4, 2, and 0
- a car dashboard with an illuminated symbol that looks like a stoner's waterpipe, odometer reads 42 0
- a car dashboard with an illuminated symbol that looks like a stoner's waterpipe, odometer reads 42 0, warning message reads 'TURN LIGHTS ON'

LifeSunDeath
Jan 4, 2007

still gay rights and smoke weed every day

Sab669 posted:

Weird, waterpipe has been a reliable workaround to "bong" for me. My various prompts were:

- a car dashboard with an illuminated symbol that looks like a water pipe
- a car dashboard with an illuminated symbol that looks like a waterpipe for smoking
- a car dashboard with an illuminated symbol that looks like a stoner's waterpipe
- something with "bong" was blocked
- a car dashboard with an illuminated symbol that looks like a stoner's pipe and lighter
- something with "420" was blocked
- a car dashboard with an illuminated symbol that looks like a stoner's pipe and lighter, odometer has numbers 4, 2, and 0
- a car dashboard with an illuminated symbol that looks like a stoner's pipe and lighter, odometer reads 4, 2, and 0
- a car dashboard with an illuminated symbol that looks like a stoner's waterpipe, odometer reads 42 0
- a car dashboard with an illuminated symbol that looks like a stoner's waterpipe, odometer reads 42 0, warning message reads 'TURN LIGHTS ON'

I guess it wasn't waterpipe but 420, that was my workaround for saying weed, but it doesn't work now. Thanks for the tips!


it's like AI knows exactly what you want but it's gotta make a stupid game out of getting it

pixaal
Jan 8, 2004

All ice cream is now for all beings, no matter how many legs.


Homer Simpson with a dab rig in a cloud of vapor

Sab669
Sep 24, 2009

LifeSunDeath posted:


it's like AI knows exactly what you want but it's gotta make a stupid game out of getting it

Yea, something about that prompt is giving me much worse [IMO] results. Now I'm getting like, plumbing pipes. I tried changing "rasta colors" to the actual colors but that didn't improve it.

Adbot
ADBOT LOVES YOU

LifeSunDeath
Jan 4, 2007

still gay rights and smoke weed every day

Sab669 posted:

Yea, something about that prompt is giving me much worse [IMO] results. Now I'm getting like, plumbing pipes. I tried changing "rasta colors" to the actual colors but that didn't improve it.

trying to test the limits lol





now I want rasta milk

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