🏞｜general-with-images

[glossy herald]

you can ask things but not share NSFW here

#✍🏼｜rules-and-tos

[wary magnet]

oh oops sorry

well I guess I can't ask then, because without seeing it nobody can tell me the model XD

[wispy nest]

damn, this background tho

human thing is kinda meh, but background is so good

[glossy herald]

yep, sorry

[wary magnet]

if I only show head it's ok ? without the nsfw body

[glossy herald]

I didn't watch it for its entirerty so hard to tell 100% but as long as it's not NSFW anymore, not sexual position or clearly lewd, or excessive nudity, we don't have a problem on this side

and sorry for the "yes/no" I put you through

[green plover]

Don't feel bad. I'm a pinup artist so I rarely get to share my pics

[wary magnet]

XD no worries, it's difficult position because I want to know how to make that accurate characters, but the guy is only doing nsfw stuff, me I just want to do it with normal clothes XD

[green plover]

What style is it?

[wary magnet]

is that ok ? @glossy herald

[wary magnet]

my question was about the model or the lora used for the character

because i've been trying to do her since a long time, it's Yamato a character from One Piece, i've trained a lora for her but it never worked good

[wary magnet]

so if someone is familiar with training loras or have any idea about the model used for this one, please tag me or dm me x)

[green plover]

Did you try anything V3. maybe put nsfw in the negative prompt

[wary magnet]

the issue I have is that the character is not known/accurate in SD without trained loras I think

the image is not from me

[glossy herald]

if it's for anime, I would also recommend asking in #1072013871730131004 . it's one of our biggest communities on the server, and they sure do have lots of models and lora shared in there

it's also one where we talk the most about rule 4 :p

[wary magnet]

oh what a good recommendation !! thanks !

I will post the same message there 😄

[green plover]

better rule 4 than rule 34?

[glossy herald]

lol

yeah

rule 34 would be under rule 4 though

I'm wondering what to train... I was making a "houses" dataset, but that feels a little .... meh

[green plover]

I'm feeding an iclone render into realistic vision with instructions to turn it into a photograph. It's working 😄

[dense tapir]

I am about to toss loras

go back to embeddings as this is nonsense

[wispy nest]

Creatures

[dense tapir]

well, I guess it worked

without and with at 0.6

[glossy herald]

Hit me with some examples pics and a little description of what you would want to do with it, I'll look into it. I remember you talking D&D recently when talking monsters together

[ripe cedar]

There's a line art model out there (actual line art, 0 shading) which could work with scribble.

[dense tapir]

so unpredictable

[novel ravine]

the rock as a catholic priest

[wispy nest]

Dragons, Griffons, Serpents as big flying creatures , maybe humans riding them

Another category might be humanoids , like goblins, orcs, elves , other human-like races.(but that's not as interesting as mythic creatures)

[novel ravine]

https://civitai.com/models/22345/dwayne-the-rock-johnsonlora

[wispy nest]

But idk , that might need actually huge dataset for model to be decent at reproducing them and probably higher resolution then 512x512

[dense tapir]

Yeah, Lora is great for subjects, like DB, but I can't tame it for styles on 2.x

[wispy nest]

They all so different...

[novel ravine]

I have failed to make clothing lora many times now, without the face of the subject getting dragged in,
most clothe loras on civit also have this problem

[glossy herald]

So mythic creatures mostly. I'll look what I can find copyright free and see later, thanks

Chimera and the like

[wispy nest]

• chimera, cerber 😄

[wispy nest]

Yea, just thought about them too

[dense tapir]

I can't get loras to style. I don't have the bringing in face issue I just can't get them to work right. Locon either and locon is better for styling

[glossy herald]

They seem the best in fact, they are mix of mythical creatures a lot, so it feels like it could even cover the whole dataset if i find some sources

[wispy nest]

dragons are weirdly hard to find

[dense tapir]

They all seem to belong to Wizards of the Coast is why.

[austere pilot]

[glossy herald]

I should give another go to my "button eyes" dataset one day

hard to find varied sources

[austere pilot]

no please not not from that movie, when I watched it as a kid it was legit scary....

[wispy nest]

seems like "high angle POV" actually works

[carmine smelt]

THIS IS FROM AI?

[glossy herald]

about all there is in this channel is

[carmine smelt]

WHAT WOW

[carmine smelt]

damn

[wispy nest]

yea and it's pretty easy to notice, doesn't look like something naturally generated

[carmine smelt]

well yes but is beautiflu

[wispy nest]

those 2 aren't AI generated, we were talking about ideas for a model

[glossy herald]

let's get the party going guys

[carmine smelt]

WUT

[glossy herald]

we have a new commer

[carmine smelt]

what was the prompt

[carmine smelt]

How to better word this

Olympic rings in a long shot floating in a pitch black sky near an ocean

[tardy mist]

How do make these images

[glossy herald]

more than a prompt, I use custom models and quite a lot of extensions
but here are the parameters:

parameters

upper body, female dj, colorful clothes , quirky, vibrant appearance, playful accessories, creative behavior, imaginative, sensual, spontaneous, dj headphones, mixing console, music club, night club, indie theaters, people dancing on the dance floor background, underground danceclub, highest quality, skin texture, intricate details, (cinematic lighting), RAW photo, 8k
Negative prompt: airbrush, photoshop, plastic doll, plastic skin, easynegative, monochrome, (low quality, worst quality:1.4), illustration, cg, 3d, render, anime, asian, korean, japanese, clivage, boobs

Steps: 80, Sampler: UniPC, CFG scale: 15.5, Seed: 1089573738, Size: 960x544, Model hash: c4947fa2c4, Model: DB_mix_realbiter_v10, Denoising strength: 0.44, Hires upscale: 2, Hires steps: 30, Hires upscaler: R-ESRGAN 4x+

[carmine smelt]

https://beta.dreamstudio.ai/dream

[carmine smelt]

uh- damn

I dont understand many terms

so cant use them ig

[glossy herald]

take it a step at a time

don't worry

this is a hard tech to tame

[carmine smelt]

haha ur saying it as if this is something to learn

[glossy herald]

we've been here for months

[carmine smelt]

but this is very cool tech

[glossy herald]

well the fact is, if you want results like those, it is

[carmine smelt]

uh huh

[glossy herald]

you need to get a lot of things, in order to tinker correctly with it

[carmine smelt]

damn i ran outta credit i couldnt get my picturw

[glossy herald]

wheras MJ or DallE can give out of the box great things for sure, SD lets you get into the small details of it all

[carmine smelt]

i dont understand why it gives bright sky when i have typed pitch black sky

[carmine smelt]

hey could u generate one for me please

just with the prompt

cause im guessing its free for u (cause u prolly have client)

[carmine smelt]

wow hahahaha

[glossy herald]

"dark night" would work better than using the word "sky", that is a lot used to describe dailight, that is why

[carmine smelt]

o

man how long before ai is making movies

i mean animation is basically justy images put together runnning very fast

ai can do that in some time

[wispy nest]

try another model or try to change prompt a bit, I had results with dark \ black and even blood moon sky, it's doable

[carmine smelt]

or should be able to hopefullly

[glossy herald]

you should look at the #1080946152318443610 , and maybe just run an notebook to have your own. google colab gives about 2 hours of free computing time per day

[carmine smelt]

okie ty

[carmine smelt]

r u genersting these rn?

[glossy herald]

no, those took about 1 minute 30 on my 3090TI

[carmine smelt]

looked real for a second but the collarbone and hair give it away instantly

[glossy herald]

I'm scrapping my outputs to show you what it can do

[carmine smelt]

oh pls show haha this is awesome

[glossy herald]

ultrawide backgrounds

[carmine smelt]

WOW

[carmine smelt]

omg

this is awesome

[carmine smelt]

straight outta an anime

[glossy herald]

yeah that one worked really well

[carmine smelt]

where does SD get image data to train model

[carmine smelt]

u mean different versions?

Okay my suspicion is confirmed under a few years ai is gonna make movies

[glossy herald]

LAION B mostly I think ? but like I said, I'm using a custom model, so people have trained it more on specific pictures

[livid ermine]

Maybe cooler vintage equipment and make it seem more like a music festival so she seems more wanted

[carmine smelt]

wowww

[robust oar]

img2img

[glossy herald]

Remixing Whiterun

[carmine smelt]

WOW

[livid ermine]

Weird that there are no real good female DJs or musicians in general. But it makes perfect sense, throughout evolution it was males who hadsing beautiful songs toget females.

[carmine smelt]

game devs should use ai to generate game art

[carmine smelt]

whoa

[glossy herald]

if you like anime, I'm not that good at it, check in #1072013871730131004

[carmine smelt]

#deleted-channel ?

lol

[glossy herald]

woops

[carmine smelt]

i dont like anime no

[glossy herald]

damn

[carmine smelt]

lol

[livid ermine]

Got a couple of tricks btw.

[carmine smelt]

?

[livid ermine]

For one I go my script working!

[glossy herald]

we have an old channel, I keep linking to it by error

[carmine smelt]

lol

[carmine smelt]

Man I cant imagine how much math went into developing this bot

[wispy nest]

Different models trained on different things, some model know words other don't...
Changing model sometimes help doing something specific

But nvm, in your case just changing prompt should be enough

[carmine smelt]

o

Dall-e overloaded oof

[livid ermine]

Unfortunately my subject, one of the dudeI signed turns out to be a psycho who wants to kill me.

[carmine smelt]

feels like mixture of real and anime

[livid ermine]

He seemed perfectly nrmal until, well, he didn't.

[wispy nest]

oh, you're using dalle, nvm what I said then 😄

[glossy herald]

I was trying to get a bear, not sure why the girl was there :p

[carmine smelt]

im not

I am just looking into ai generated images

and trying to generate a cool one on my own

[wispy nest]

it's not that hard

[carmine smelt]

is it possible that i upload image to ai and it gives me images related to it

[carmine smelt]

well Olympic rings in a long shot floating in a dark night above an ocean is not generating anny good ones imma try decribing more

[wispy nest]

You can use google for that 😄

[jaunty mural]

yeah

[carmine smelt]

but thats not ai generated

[carmine smelt]

how/where

[jaunty mural]

ur in this server arent you

[wispy nest]

You can use control net to get pose \ form of an object and prompt to get object you want, but it's not as easy as you want

[carmine smelt]

yes

[carmine smelt]

well i dont wannt get into all that noise and stuff just wanna genrated based on prompt since idk any of those terms

[glossy herald]

bad cosplay man as a frog

[carmine smelt]

btw im trynna get one like tihs

[glossy herald]

forever alone guy

but real

[carmine smelt]

omg

^ not ai generated

[glossy herald]

yes ai generated

[wispy nest]

@fleet carbon are you hiriz-fix-ing all your gens or just ones you like?

[carmine smelt]

but i wanna generate similar to that

[carmine smelt]

no xD

[glossy herald]

https://www.reddit.com/gallery/11e4dfw

I did a series of rage comics

[carmine smelt]

ai generated?

[wispy nest]

okaay...let's finish island castle theme there...

[carmine smelt]

what are rage comics xd

[glossy herald]

the corresponding meme faces

=

[carmine smelt]

oh lmao

[carmine smelt]

wait so how do i generate more images based on sending my own

[glossy herald]

multiple ways to do that.

just to modify a pic, you use "img2img"

or the "input picture" at the bottom right of dreamstudio

but this was using an extension named "controlnet"

that can help condition what the AI will output

[carmine smelt]

r u talking about dreamstuidio?

[glossy herald]

no, I couldn't have done that one on dreamstudio with the current tech available there

[carmine smelt]

oh

on the client version?

why so many ppl generating anime images lol

[glossy herald]

Stable diffusion is mostly a brick, that makes image. Dreamstudio paints that brick a lot, but client version lets you add so many smaller bricks around it, it's another wall completly

forgive my metaphore

not sure why anime is so popular, not my jam

[jaunty mural]

cos fantasy into reality

[carmine smelt]

lmao

but with ai u can do reality into reality

[jaunty mural]

yes

[carmine smelt]

cool, i bet theres a million ways to edit an image on the clint version

prolly wont use any of them tho haha

[glossy herald]

there is

[carmine smelt]

omg

where does the prompt even go

[glossy herald]

up top

[carmine smelt]

oh

xd

[glossy herald]

"bad cosplay man as a frog on a plane"

[carmine smelt]

this is deffo not made for a layman

[glossy herald]

there are more intuitive ones too 😉

and harder ones too

[carmine smelt]

even tho i know some computer programming and im learning more and im better with using tech than an average person idk any of this

[carmine smelt]

haha

[glossy herald]

and then there is "comfyUI"

is it comfy ?

you tell me :p

[carmine smelt]

absolutely not

[glossy herald]

but it's damn useful

node based image making

[carmine smelt]

its like linux, the default ui makes you want to cry

but once u get used to it

u feel like a demigod

and others thing ur a manicsc

[glossy herald]

https://www.youtube.com/watch?v=WmOUl8Gab5U&t=136s

I recommand invokeAI if you want a fire UI

[livid ermine]

yeah linux is horrible all over

been using it for almost 30 years but

[wispy nest]

Damn, I wanted extention with vae choise right there for so long, what is it?

[livid ermine]

Choose OpenBSD for your Unix needs. OpenBSD -- the world's simplest and most secure Unix-like OS. A safe alternatve to the frequent vulnerabilities and overengineering of Linux and related software (NGiNX & Apache (httpd-asiabsdcon2015.pdf), OpenSSL, iptables/nftables, systemd, BIND, Postfix, Docker etc.)

OpenBSD -- the cleanest kernel, the cleanest userland and the cleanest configuration syntax.

It's like how cool wouldn't it be if all this AI stuff was written in Ruby instead? They're basically the same language except one has a lot less noise.

[carmine smelt]

nice, im on linux now lemme jnstall this thing

[livid ermine]

But yeah this

https://elsamuko.github.io/gimp-elsamuko/scripts.html

[glossy herald]

not an extension, just using the "quicksettings" setting here :

[livid ermine]

NEVER install linux as a webserver though

[glossy herald]

https://invoke-ai.github.io/InvokeAI/ and #🤝｜tech-support if needed be 😉

[livid ermine]

that's asking for trouble

[carmine smelt]

not this i meant SD

[livid ermine]

yup

[glossy herald]

this is the SD video I linked, it's a UI made on SD

[livid ermine]

But for the record, if you ever wanna set up a webserver, check out my latest piece: https://gist.github.com/basicfeatures/7d1448aad6fadc0def4ffee443f26dab

alright i gotta figure out how to add those gimp effects

for increased realism

[glossy herald]

also @wispy nest you can have any setting from the "settings" tab up there, if you get the name of the setting. all names here : https://github.com/AUTOMATIC1111/stable-diffusion-webui/blob/master/modules/shared.py#L317

[wispy nest]

I'm guessing we can add clip_skip here too?
Do you know it's actual name?

[glossy herald]

😉

[wispy nest]

oh

[carmine smelt]

oh

the link to install SD Is in #1080946152318443610 ces right?

[wispy nest]

so CLIP_stop_at_last_layers yay

[glossy herald]

people have been building on top of SD for the last 8 months basically, so there is a lot now 🙂
you don't "install SD", you install one of those tools, like AUTOMATIC1111 (the first UI I showed) or InvokeAI (the video). all links in #1080946152318443610 yes
it will have SD inside it

[carmine smelt]

uh

um ok

whats the best

[glossy herald]

I showed you the most used ones just there. people tend to use AUTOMATIC1111 (the one that scared you at first) the most currently, it has the most features
ComfyUI is great for professional settings, very versatile, but hard to use
InvokeAI is more turned towards artists in my opinion

[carmine smelt]

uhm

okay ill just use automatic111

is the link for that in #1080946152318443610

[glossy herald]

https://github.com/AUTOMATIC1111/stable-diffusion-webui/

yep but here

[carmine smelt]

ty

ok lets try this out

[glossy herald]

you should have quite some help in #🤝｜tech-support but to sum up how to install it

first the dependances : git and python 3.10.6

second you clone the repository

[carmine smelt]

i have git

[glossy herald]

third you download the models you want and store them inside the cloned repository, inside the models folder

[carmine smelt]

but python is 3.8

[glossy herald]

and last you run the "webui-user.bat"

[carmine smelt]

rightttttt

[glossy herald]

3.8 may work, not sure. I know 3.6.10 does, as does 3.9 on some configs

[carmine smelt]

ill just paste all commands blindly and hope it works

xd

from the gitub repo

sudo apt install wget git python3 python3-venv why -venv

i shuld prolly ask in tech support

[glossy herald]

yep, it's not really hard usually, but do be careful, python version can bite you in the ***

yep

[carmine smelt]

lol

ig it uses python for the actual generation

[wispy nest]

where did the individual channels go

like sci-fi art

[glossy herald]

we moved that into forums, in the #dreamers communities , like #1070829982252798043

[trim jacinth]

@glossy herald hello. can you help me again? How can I make perfect and beautiful pictures like this picture? https://www.reddit.com/r/StableDiffusion/comments/118cdcx/i_swear_ive_seen_this_town_before_controlnet/

[wispy nest]

thanks

where do i find sci fi there 😂

[mighty ice]

Rip

[pulsar vessel]

At least I'm gifted with patience.

[wispy nest]

It will probably stop after some time

What does this even mean?
4 models?

we can't change them, so gow exactly this works?

[pulsar vessel]

it seems it got stuck

[wispy nest]

yep

What can i do with the karma ?

[stoic pivot]

It is soon Easter, where can I see your Easter images?

[sacred flame]

happening to me too

also won't let me select images

[wispy nest]

How do I even chose 😄
They both terrible 😄

I don't like whole idea of this thing, it's so wrong...

[sacred flame]

?

[wispy nest]

One image could be better, but it's not what I prompted, so like...it's not better then in this case...

[sacred flame]

go with what's closer to the prompt

[wispy nest]

It needs more options

[sacred flame]

if they're both closer to the prompt pick the better looking one

which image would you consider better overall

if they're equal, pick equal

[carmine smelt]

@wispy nest @vague oasis

[brittle steppe]

😍

[wispy nest]

yea, I was thinking about sharing their today's talk, but doubt anyone would listen 1.2h talk about AI

[brittle steppe]

currently, lol

[wispy nest]

not this one

https://register.nvidia.com/flow/nvidia/gtcspring2023/attendeeportal/page/sessioncatalog/session/1673303350236001hW73

it's about AI overall from nvidia

[brittle steppe]

thanks fam

[sacred flame]

"uncensored and non-cucked text and image models coming soon" -Tabloid, novelai

[glossy herald]

here 🙂 https://discord.com/channels/1002292111942635562/1072017962531307540

[wispy nest]

oooh thank you, i was looking in the other channels and didnt find it

anime genre huh

[glossy herald]

yeah, not sure why classified in there tbh

[wispy nest]

[pseudo aspen]

It was always a dream of mine to take photo with Chester Bennington from Linkin Park. However, because of what happened six years ago, it's no longer possible. Thanks to modern technologies and SD, I will be able to take selfie with him. LoRA of him is finished. Now it's time to train AI on myself. For now, see an example of how my LoRA works.

[trim jacinth]

@glossy herald can you help me again? How can I make perfect and beautiful pictures like this picture? https://www.reddit.com/r/StableDiffusion/comments/118cdcx/i_swear_ive_seen_this_town_before_controlnet/

[glossy herald]

I can try and give you a hand later, I have a reunion to go to rn. I'm not sure I would know exactly how to do those, but it's close to what we did the other day

[trim jacinth]

okay thanks, its similar but its more beautiful

[stoic pivot]

The prompt is not 100 % correct, but it is a bit fun, and I hope it is not too political loaded. (or hated for it come from the wrong place).

[grizzled sage]

staged prompts are staged. we all know how craiyon.ai works. it's dall-e mini not stable diffusion

[stoic pivot]

I know. I was trying to make something with my StableDiffusion but I can not make it do a good Gollum.

[grizzled sage]

They might've upgraded their backend. That's a pretty good kennylex

[stoic pivot]

"Portrait Gollum" SD V1.5

[glossy herald]

ok so here I am
2 things.

First, a cow fact 🐮 Cows have a secret society known as "The Moo-tants," who possess extraordinary powers and abilities. Members of this society have the ability to shoot lasers from their eyes, manipulate time and space, and even control the weather. The Moo-tants have sworn to protect cows everywhere from harm and injustice, and they meet in secret to plan their next moves. It's rumored that The Moo-tants are responsible for some of the most mysterious and unexplainable phenomena on earth, such as crop circles and UFO sightings. However, despite their incredible powers, The Moo-tants remain humble and dedicated to their mission of protecting cows and their way of life.

Second, as much as I can help you do things in the same way, force the AI to follow whiterun's entrance like that (I played with the same picture too), I'm not sure I have the good prompts for it. My results were very bland, and the architecture didn't change like that. In fact, I wanted to try to make it cyberpunk and never managed to leave a medieval vibe

[crystal bloom]

Hugging Face Infinite Zoom Out and Zoom in

[analog garden]

lol

[glossy herald]

sushi chef cat

[trim jacinth]

https://www.reddit.com/r/StableDiffusion/comments/118qcwz/control_net_is_too_much_power/ In this photo, the shape is formed beautifully without distortion, how is this done?

[glossy herald]

first they seem to use a good model for that great aesthetics, users seemed to suggest Realistic Vision for example
Second, this uses Canny mode for sure, at least. They played on the thresholds and used the preview to select the good ones, the ones that were giving enough but not too much details in the preview.
The they surely used a second and maybe a third controlnet at the same time. I would guess Depth and/or Normal map, those are great for situations like those.
Also, on mostly all the controlnets, they may have reduced the Guidance End, meaning that for the last few steps of the picture, controlnet wouldn't act, making the results more possible, while keeping the composition you wanted
Then, they also use "Restore faces" option.

it's about all I can see in there. Then good prompt and lots of seeds

when using all that controlnet, a lot of the results just don't work, since you push it a little too hard for it to make possible things

so letting go of the guidance in the end, and doing lots of seeds, that's how you get those bangers

no link : some 1920x1080 I'm making currently

[grizzled sage]

i think that was one of the first controlnet remixes before guidence ends and multi control was brought in

just a good prompt with a good model and dialing in one of the canny modes

[glossy herald]

@verbal loom just wanted to say thanks again for those great styles you gave me

[grizzled sage]

it's the same BDSM memes that went viral when 50 shades of grey was making the marketing rounds

[glossy herald]

I missed the reference on that meme at least, 2 girls and a bottle of milk from knowyourmeme, but I never saw it online

[verbal loom]

Didn't really give you styles tho, it's what's already in the models. Although I've been trying to make V2 of my randomizer 🤣

[glossy herald]

yeah, but I wouldn't have been able to call for them that good without a lot of previous work 😉

[grizzled sage]

it was fully one of the "forced memes" that came up around that time. Not just for 50 shades but a lot of marketing companies were eager to use memes to sell. same with the DC justice league movies. A very obvious amount of new DC comic memes were pouring onto communities. astroturfing is very real

usually nobody ever remembers those forced memes but some reason that milk one has staying power

[verbal loom]

yeah, it's a real easy way of "give me something interesting"

[verbal loom]

i think stylepile is a extension that tries something similar tho

and is certainly more controllable

[glossy herald]

did you know ? 🐮 Cows are actually the world's first flying animals, with the ability to soar through the skies like eagles. According to ancient legends, cows were born with wings, but they chose to keep this ability a secret from humans to avoid being hunted and captured. However, some intrepid scientists have recently discovered that cows are able to fly, using a combination of flapping their wings and catching air currents. It's said that cows are most likely to take flight on moonlit nights, when the magic of the full moon gives them extra lift. These flying cows are known to journey to distant lands in search of new grazing pastures and adventure, and they're even able to outmaneuver airplanes and helicopters with their incredible speed and agility. However, humans must never attempt to ride a flying cow, as it's said to be a recipe for disaster.

[trim jacinth]

depth leres or depth is better for human photos?

and i should off preproccesor when i rendering final image

[grizzled sage]

experiment and find out. they're just two different depth estimators. leres has background removal

i prefer normal maps for people

[verbal loom]

alien + my random prompt v2 (work in progress) lol

[glossy herald]

depth I think. depth leres lets you change the thresholds though. You should try to click "preview" to know and fine tune the thresholds.

[grizzled sage]

preview button helps a lot in the process

[glossy herald]

no, once you have found the good settings for each controlnet, you should have all activated and using the preprocessor

[grizzled sage]

i wish it was more up near the settings

always have to scroll down an up

ngl its infuriating

[glossy herald]

yeah, and that you could hide the preview too, once you are done

depending on your image ratio, it can push other settings quite low

[trim jacinth]

okay 👍

[verbal loom]

do you guys know what's up if controlnet doesn't seem to respond to guidance start or end?

[grizzled sage]

or hotkey the button even better. maybe i'll dig into the code and do it myself

bing! write me hotkey code for gradio

[glossy herald]

it does usually quite well for me. droping under 0.8 feels like almost no controlnet

[verbal loom]

for me it's reverse, they just don't disable at all

like even ending them before the generation got going (0.01) and are still applied in full

😦

[glossy herald]

that is quite strange yeah... try to update the extension in the Extension tab maybe

[verbal loom]

updated, even reinstalled controlnet

[glossy herald]

and it does that on each mode ?

[verbal loom]

yup

[glossy herald]

first time I hear of that

[verbal loom]

at first i thought it was with style transfer only

but i noticed it happening with canny and hed to

[glossy herald]

yeah maybe an interaction with another extension ?

[verbal loom]

posex? 😮

that's the only one i can imagine

[glossy herald]

like I know the depth extension had a button "send to controlnet" that would just break the UI until you restarted it

I'm not saying it's this one for you

I'm saying multiple extensions can do things to each other

so maybe try running just controlnet, to see

disabling the others in the settings

[verbal loom]

that's a good idea, honestly

lets try that

[glossy herald]

confyUI has controlnet too ... 😉

jk

well it has it, but it's not what you're looking for here

[verbal loom]

i coulnd't get comfy-ui to control controlnet

uhh... yeah

[glossy herald]

yeah, I have spent some time on it but honestly, I don't have enough to really set up my workflows in it currently

I did some, but every change can be quite long, and I like being able to just change the whole setup on the fly in auto currently

[verbal loom]

fuuu, still applies controlnet regardless of what i do -> guidance start at 0.85, still starts with it

[grizzled sage]

i think comfyui has a LOT of potential but its not quite there yet. node based is awesome. the other one is tuned to animations and thats awesome too. nodeAI i think?

[verbal loom]

yeah, node based is WAY stronger in general, but goddamn it takes a lot of time to set up workflow that way

[glossy herald]

try to ask in #🤝｜tech-support if maybe someone has seen that in the past, I'm dry of ideas :/

[verbal loom]

imma just do go clean controlnet install, and removing all config from the jsons

that SHOULD clear it out

i suppose

[grizzled sage]

theres an app called graphite.rs which has tons of potential too. its coded in rust and runs as a preview through webasm right now. i think you can pull it yourself from github too

node based and connects to a sd server for editing

[wispy nest]

hmmm

[dawn niche]

Hi guys, can i put a prompt and generate images here in discord?

[wispy nest]

nope, we don't have a bot anymore

[sacred flame]

I recommend playgroundai.com for free gens

it's very fast

[haughty wren]

Hey i have an error when i put "--no-half-vae " args when i try to generate an image

someon have an issue ?

[smoky oak]

Yeah, the concept and potential is dope, but the execution is currently lack luster. I am excited to see how they do over time as they make it better tho

@dense tapir I guess we're gonna see if SDXL is actually good, or if it's just SD 2.x repackaged

Oh shit, it is way bigger

😅

From just over 800millin params, to 2.3 billion, yeesh

[static tusk]

Shutterstock is partnering with NVIDIA to develop models to generate 3D assets trained on fully licensed content from Shutterstock. These models can be used to generate high-fidelity 3D assets from simple text prompts which can in turn be used in game development, animation, and other 3D workflows.

text to 3d

https://www.nvidia.com/en-us/gpu-cloud/picasso/

[ripe cedar]

Anyone try ComfyUI? The nodes based workflow should make things quicker and gives more control, but it doesn't have the support Automatic does

[glossy herald]

it has quite a lot of features AUTO has, the main ones everybody use mainly I think, but I find it slower to set up when you change settings a lot

I think it's the best if you use mainly 1 complicated pipeline

but for changing pipelines a lot, auto is a lot faster

[ripe cedar]

Do I have to have 2 copies of the same model if I want to use both?

[glossy herald]

there may be an option with comfyUI to specify the model folder ? I'm not sure. I know in auto, you can use set COMMANDLINE_ARGS=--ckpt-dir "F:\AI\Data\Models"for example, to set the model dir to another directory on your computer, so you could point it to comfyUI's folder

[glossy herald]

welcome on the server but, I mean, really ? chain mail as first message on discord with such content ?
(they left the server imediatly, I removed it)

too many fingers but damn catgirl head

[smoky oak]

ASSS CHEEEEKKKKKKSSSS

[smoky oak]

You have me so curious what they asked now lmao

[glossy herald]

just a creepy copy pasta

[flat moss]

White bear

[glossy herald]

I have a bug on the bot so I didn't delete it like usual and it didn't log

[crystal delta]

@flat moss

FAQ: How do I generate images? Is there a bot on the server?

Currently, there is no bot on the server that generates images. However, there are plenty of other ways such as the official https://beta.dreamstudio.ai/ website or by running Stable Diffusion locally using your own hardware! Check out #1080946152318443610 for more details! You can also stop by #1025467151206854736 for any issues you experience while using DreamStudio or #🤝｜tech-support for any problems you encounter while installing it locally!

[smoky oak]

Ahhh, I see

[glossy herald]

like "send this to 10 of your discord friends or YOUR GONNA DYE "

those ones

and when I wanted to send a warning i was they were already gone anyway

I hope they lifted the curse and won't die because of that copy pasta

new profile pic

[silent haven]

I want something that looks more like an actual trilobite. This is the prompt that I used "a photograph of a golden trilobite crawling on a muddy beach, the ocean is in the background, the trilobite has two eyes, arthropod, pyritized trilobite, woodlouse, hyperrealistic, highly detailed, 8k, professional photography"

[glossy herald]

first a sir cat

and then let's see

a photograph of a golden trilobite crawling on a muddy beach, the ocean is in the background, the trilobite has two eyes, arthropod, pyritized trilobite, woodlouse, hyperrealistic, highly detailed, 8k, professional photography

some basic changes

a wildlife picture of a golden trilobite crawling on a muddy beach, with two eyes, ocean background, arthropod, pyritized trilobite, woodlouse, hyperrealistic, highly detailed, 8k, RAW photo, professional photography, taken for national geographic

lots of the words you use are completly foreign to me though

you could also add tokens to drive the lighting, like "taken at golden hour"

that makes some good effects sometimes

in fact

let's run it

like you saw, I added quite some terms for making it more realistic

I'm also using a quite realistic model

this is the best of the 3 1024x1024 I got

this one is fun but has no eyes

[astral osprey]

@dry crow THX for the help these past few days amigo, I've been having a blast

[glossy herald]

removing the golden hour because it makes everything gold

what the hell am I computing here lol

what do you think @silent haven ? you're the expert there

#🏞｜general-with-images message
this one stays the best imo

a wildlife picture of a golden trilobite crawling on a muddy beach, with two eyes, ocean background, arthropod, pyritized trilobite, woodlouse, hyperrealistic, highly detailed, 8k, RAW photo, professional photography, taken for national geographic, taken at golden hour
Steps: 60, Sampler: UniPC, CFG scale: 5, Seed: 1613606767, Size: 512x512, Model hash: c4947fa2c4, Model: DB_mix_realbiter_v10, Denoising strength: 0.44, Hires upscale: 2, Hires steps: 30, Hires upscaler: R-ESRGAN 4x+

[tepid nebula]

Epic manga mermaid

[crystal delta]

@tepid nebula

FAQ: How do I generate images? Is there a bot on the server?

Currently, there is no bot on the server that generates images. However, there are plenty of other ways such as the official https://beta.dreamstudio.ai/ website or by running Stable Diffusion locally using your own hardware! Check out #1080946152318443610 for more details! You can also stop by #1025467151206854736 for any issues you experience while using DreamStudio or #🤝｜tech-support for any problems you encounter while installing it locally!

[glossy herald]

sorry Laura, no bot :/

unless you were calling me an Epic manga mermaid, in wich case, thank you

[silent haven]

It at least looks like an arthropod, but it looks more like a beetle grub than a trilobite. It doesn't help that there aren't any photographs of this extinct group of animals for the AI program to work from.

[glossy herald]

and welcome around

[glossy herald]

no it doesnt, but if you have a sketch of one, we can push more conditionning on the AI, force it to follow your sketch and make a real picture out of it

tech like this

[silent haven]

This is closer to the right number of legs and it also looks like some kind of crustacean.

[glossy herald]

I'll do a few more

it's about 12 second each

[silent haven]

There is however a large selection of paleoart available on trilobites. There might be some illustrations under Creative Commons on Wikimedia Commons.

[glossy herald]

this is horrifying by the wya

[glossy herald]

if there is illustration of it, you could train the model more on it too

and it would be able to do better on that side

it's a whole other subject but it can be nice

[silent haven]

Let me try that.

[glossy herald]

stop, too dirty for me :p

[silent haven]

I found this image on Wikimedia Commons. I'll try it on Stable Diffusion.

I got the following image generated with the above image using the following prompt "a wildlife picture of a trilobite crawling on the seafloor, with two eyes, arthropod, pyritized trilobite, woodlouse, hyperrealistic, highly detailed, 8k, RAW photo, professional photography, taken for national geographic"

[glossy herald]

I collected a dataset. I think I'll train it as "chimera", nothing more, teaching that a chimera is a mix of multiple animals and mythical creatures, in different artstyles and backgrounds, ...
This way, it should be possible to prompt "chimera horse fish" or something, let's hope. here is a small view of what I got for now

[glossy herald]

you used img2img ?

[silent haven]

I'm using the online version of stable diffusion through mage space, but I'll read the tutorial section and get the full version installed on my PC.

[glossy herald]

just wanted to be sure

I'm trying to use controlnet to see

first off without controlnet

[silent haven]

https://cdn.mage.space/generate/28cfe80b9e1341f286896bc58e87cd5b.png

[glossy herald]

with your latest prompt that was

with controlnet

[smoky oak]

I trained a LoRA on freshwater shrimp lol

Without a LoRA

With my LoRA

[silent haven]

Now these actually look like plausible trilobites. I'm guessing the controlnet is on the downloadable version for PC, but I'll read the "about" section first to make sure.

[lucid hamlet]

So, I taught chatGPT to take bulk statments and turn them into stable diffusion prompts... best thing I've done all day

[glossy herald]

controlnet is another tech on top of SD. it lets you condition what SD will output, here I forced a "canny" mode, but here is another example using a nomal map mode instead

depending on the settings I use, it lets more or less freedom to the AI

so you can manage a little creativity

here is the same using depth map mode

[silent haven]

Again, these results are pretty impressive. I'm definitely going to try installing the full version this week. I was initially impressed with some of the results Shad Brooks got with Stable diffusion on his YouTube channel Shadiversity.

[stark vine]

I just found out you can do img2img on native 1024+ with the help of controlnet. Good god. There truly is no limit to Controlnet's power. I can literally inpaint and touch up any part of my image without even opening photoshop. This is an arthead's wet dream come true

[smoky oak]

I have been using an addon called ultimate SD upscale, and it allows you to continuously refine at higher and higher resolutions, while actually adding more detail and pixels, rather than just over defining the ones that are there

[stark vine]

I tried it yesterday since some dude told me you were hyping it up

[smoky oak]

I took a base image I made and was able to get it from 2560x1080 all the way up to 15360x6528

[stark vine]

it works like shit lol can u post a tut

pretty sure im not using it right

[smoky oak]

I can help you more when I get home

[stark vine]

it splits the image in squares and does each one and then puts them together right?

[smoky oak]

Were you able to see my results?

[smoky oak]

Correct

[stark vine]

Yeah im gonna need a tut cuz I had no luck, it looked like ass

[smoky oak]

It took me a long time to find the right workflow, but my results have been insane

[stark vine]

interesting

[smoky oak]

Base image (so you can see where I am zooming)

There's a comparison of base res and ultimate upscale x6 (which is 6²x or 36x)

@stark vine I think I can actually help you now

First, you wanna go into settings

[stark vine]

nah G not backgrounds

those are easy I wanna see subjects like my vampire girl

[smoky oak]

Wallpapers are just as hard as subjects

I've used this on subjects as well

[stark vine]

So what settings do you use

I can try it rn

Or start from the top

[smoky oak]

You need to go into settings

[stark vine]

U generate low res and then upscale?

[smoky oak]

I'm trying to give you my settings. Go into settings, stable diffusion, and you wanna turn on the img2img force steps setting

Force stable diffusion to use as many steps for img2img as you say. It will make things way less annoying

[stark vine]

Ok. Then?

[smoky oak]

You wanna use 5-15 steps in img2img with the script running. I recommend using DDIM, as it's worked best for me

You want to set the resolution scaling or whatever (the first drop down in the script) to scale from image, which is the third option

That allows you to just use a multiplicative slider rather than having to guess

Sorry if my terms aren't exact, I can't actually see the UI right now

Can you take a picture of the ultimate SD upscale UI really fast? It will help me remember

[stark vine]

Kk hol up

[smoky oak]

Ok, target size type, you want the third option

Upscaler I recommend R-ESRGAN 4x+

Tile size can be whatever your GPU can handle, just not below 512. I use 512-1200 depending on my mood lol. On really high res images, I recommend the gigher/highest res you can use, cause the main slow down is the in-between tiles part

Important things to know. Always use a very low denoise for upscaling. Like 0.05-0.2

If you want to upscale and let it get more creative, then use 0.2+ (if you do this, you need to enable seams fix. I recommend half tile pass with a denoise of .3)

And the final thing is, use a model that works and makes results how you want. If you wanna upscale a photo, use a photo realistic model

And you have to use a prompt and negative prompt to get the best results

It took me 10+ hours nonstop of fucking around and finding out to get settings that work as good as possible

[stark vine]

no fucking way

well it works

but is it better than topaz?

let's do a pepsi test.

ima post 2 of these and I want volunteers and @fiery lichen to tell me which one looks better. One of them is the ultimale upscale result, the other is the easiest topaz upscaled from 512.

Sadly, skin detail is lost in both which is the #1 objection I have for 512 models. I'm so starved for skin details like a mf lol

Okay I think the decision is pretty conclusive without even posting them yet.

One of them obviously has more detail. Cast your votes. Left or right

However what I'm thinking is that if I increase the steps on the ultimate upscaler maybe it will retain skin detail?

[grizzled sage]

this extension is great

[smoky oak]

@stark vine send me the base image and I'll see what I can do

[stark vine]

Im experimenting rn

[stark vine]

topaz is better but hopefully the extension can prove me wrong.

[smoky oak]

As somebody who has used topaz for years, ultimate SD upscale blows it out of the water if you know what you're doing

I can only help so much, the rest of the use is up to you to figure out for your workflow

[grizzled sage]

yeah. an inexperienced person using a tool to show it's a bad tool, doesn't make sense

[stark vine]

knock yourself out homie

thats the base, prove me wrong

still waiting on the ballot for the two above

[grizzled sage]

¯_(ツ)_/¯

no

[stark vine]

u scared

lmao

[grizzled sage]

i hate when people do this. you try to give them advice and it becomes a dick measuring contest, demanding to do the work for you

[stark vine]

Ur saying that its a skill issue

[grizzled sage]

laughter is a coping mechanism

[stark vine]

I say I'd be darned surprised if it can outdo what topaz can, and the other guy asked for the base image

[grizzled sage]

All of this tech is barely half a year old. If it is a skill issue, that doesn't reflect on you as a person at all

cut the ego bullshit

[stark vine]

fine

[stark vine]

I did 40 steps and skin detail is still lost

"Without ego" I ask if anyone finds a way to use it without losing, or even increasing, skin detail then by all means share

[smoky oak]

Right

[smoky oak]

So you didn't listen to me, cool

[grizzled sage]

shit man, if you're going to intentionally use it like an infomercial bad example situation, then of course it'll fail

[stark vine]

Bru look at the 2 above

that one was 10

[grizzled sage]

confirmation bias can be a real bitch to deal with

[smoky oak]

Good luck figuring it out yourself. I'm not gonna spend so much on somebody who doesn't listen

[grizzled sage]

seriously. wasted effort to put positive energy out there just for it to be smacked like that

[stark vine]

Ur the one who isnt listening lol. The 2 side by side above are the ones I did following ur steps

[grizzled sage]

its a simple pebkac problem really. solved. moving on

[stark vine]

I tried 40 to see if detail remained and it did not

[fiery lichen]

looks like right side has more texture and tone to me.

I would not dismiss the tests on the upscaler. We may all learn something new.

even though I am not really a human portrait generator, I'd like to find out something new so when it comes time to try it out, I have techniques to try.

I am only judging by seeing wrinkles on the right image, where as the left one has less wrinkles on her left cheek (right side) of image.

The left image's tone has changed a bit too, a bit more saturation on the reds.

[smoky oak]

@stark vine here, I'm not gonna give you the full image, cause I want you to figure some of it out on your own. You can easily make insane results, if you follow what I say, and actively try to figure things out for yourself.

It's not perfect, but it's pretty damn good

[fiery lichen]

OK I am just backtracking a little bit on the conversation, R-ESRGAN 4x+ is usually recommended for photos. It can work well on some abstract details like rocks, trees, marble (somethings that I've tried before). However, it does "change" the details from what I find. Although it can add details, it often shifts them to new locations. For example, tree branches of trees from very far away, branches location shift after R_E 4x+ upscale. I don't know what it does with people. edit spelling

[smoky oak]

Skin texture, individual hairs, everything

Again, it's not perfect, as it doesn't have that much detail to play off of, but experimentation is important

It could be significantly better with more dedicated time

[stark vine]

I have noticed some insane results on eyes

Im still playing around with it, I'll have to check the documentation cuz a lot of this stuff is just guesswork

[stark vine]

yes upscaling always affects details

at least from my experience

I'm trying to do some without upscaling with dubious results

[smoky oak]

I gave you about 4,000 images worth of tweaking and guess work for free, so at least that should help

[stark vine]

yo glass if u wanna give it a shot

then by all means. u know that I'm a portrait freak

and yes the right one is topaz unfortunately lol. but this extension does seem to have great potential

I got an upscale with even more detail which is insane but it has that characteristic blurry look. hmu if u want pix

[fiery lichen]

Sorry I am busy right now with some other things not related to SD. But its good you have options now to try different upscale tools. I will watch you guys' results later.

[static tusk]

not having the original model and using one I have that tries to clear up the skin sorta removes details there, but the eyes, hair, mouth I think has more detail, with my test.

[stark vine]

ty for the input

yes thats the biggest beef I have with upscalers, removing skin deets

[static tusk]

its not the upscaler, its the model I used

[stark vine]

Ideally if it could create a brand new 512x512 and piece them together for a true native 2k for example, then there would be no loss

oic

[sinful mirage]

using the bing dalle thing,

prompt: a dragon made of grass, middle of bursting in flames, photorealistic fantasy art, matrix numbers, neon

not bad for a small prompt, obviously less control and worse faces but still

[smoky oak]

It's a program I don't see challenging dedicated programs any time soon

[static tusk]

my own 512 gen vs scaled 4x

[smoky oak]

Diffusion upscaling is very powerful. Far more powerful for lots of things than standard upscaling

Just gotta learn how to use it properly. And I hope me sharing my dozens of hours of testing can help you guys

[reef sail]

stable diffusion xl generating perfect text

[static tusk]

photoshopped

[smoky oak]

I think the text is about the only good thing here 😅

I don't wanna get hyped for SDXL, and then have it be another SD2.x failure

So I'm taking everything with a grain of salt

[static tusk]

if it can generate human bodies in motion and hands, its a step up at least

[stark vine]

its a 2x upscale

looks like a denoise filter?

I got some interesting results

[static tusk]

4, I cropped the image

[stark vine]

it seems higher denoise = better details but also more artifacts

the sweet spot is probably somewhere in between where you have enough denoise to add details but not too much that photoshop can't fix the artifacts

[stark vine]

This is the sweet spot I was talking about, with some photoshop you can iron out the artifacts and you have a pretty good upscale, however, I doubt that the same settings can be copied and pasted to each picture, it takes a lot of fiddling to come up with something halfway decent therefore it is only practical for upscaling the masterpieces. That's my theory

While I will probably stick to topaz, if they were to integrate this tool with controlnet, the possibilities would be endless

[wispy nest]

hm, interesting approach.
Some of them aren't exactly chimeras, like 3rd row image 1 - that's pegasus, but I see where you going, horse+something?
2-2 is phoenix and 2-9 and 3-10 - griffon.
5-8 looks like a wyvern

3-3 mermaid? Furthest from chimera there 😄

Maybe teach it as "creature", so we can do "creature horse with wings" or something like that?

[fleet goblet]

Looking at pickapic, quite interesting.

Pretty creative, this seems to be.

Very.

[smoky oak]

@dense tapir Looks like my controlnet is now useless after rebuilding my venv ._.

can I delete the folder and reinstall?

[static tusk]

CN extension folder, yup

[smoky oak]

alright, cool

thank you <3

[smoky oak]

@static tuskit is still broken

damn that is a shame

[static tusk]

whats it doing?

[smoky oak]

it just won't do it

I click enable, and it just does nothing

all the settings are there, but it runs no models

OHHH

I see why

I renamed this folder, so it can't find the models now

[junior mantle]

How can I create this.

[dense tapir]

yeah, venv fucked me yesterday then I went to tag 81 images and the extension undid all my hard time consuming work when I dropped a single image on the single action tab.

besides with 2.2 I can only imagine how it will be for training

[smoky oak]

[dense tapir]

Something I did learn that I never knew about training

this is a problem that plagued me in 1.5 dreambooth and never left is lighting on the source images.

it fucks over faces as it sort of bakes it in and lora being a DB offshoot same problem

guy said he had 2 images out of 100 with green lighting and his training people would say their faces always looked like they had vomit on their faces. He tagged it out for the green lighting and perfect

[smoky oak]

I am still trying to manage expectations of SDXL, cause I have a feeling it's gonna be a disappointment like SD 2.x

[dense tapir]

improvement in one area and all else meh to wtf?!?

[smoky oak]

Yeah

[dense tapir]

I expect if it can now do text better all these training tools will change? bet so

[smoky oak]

"let's show you what greatness could look like, and then make it impossible for you to achieve these results"

[dense tapir]

Well, there are a lot of us out here who have jumped off, or about to, this train as we want it to be more stable with better tools. People are damn pissed, and rightfully so, that you train something and it finishes and you get to do it all over again because you were 1 epoch, or a few steps shy. Basically, we need more of a dynamic system as time is the most valuable asset one has.

wasting time doing it over and over until it works people are upset about.

think of it like Mpeg1 vs Mpeg2 where 1 used a constant bit rate (static rate) for everything so in fast scene or explosions you got blockiness in them because there wasn't enough bits to use. In came Mpeg2 that dynamic took from scenes (like someone sitting down with barely anymovement) its excess bits to give to the scenes that needed them. Problem solved. Same thing here, basically.

[smoky oak]

I don't know, I'm just not exactly a fan of stable diffusion, the creators I mean

[dense tapir]

You and I both

tbh, I have never seen any concern for the users of their stuff.

I get it, it's free so I just sit back and hope.

[smoky oak]

Yeah, true

Would be really nice to see them actually care about their users, as this tool is kinda useless without us lol

[static tusk]

I wouldnt be surprised if all the best advancements will be online services only

[smoky oak]

It's not them that are making these exceptional finetuned models, it's us

[dense tapir]

Agree with both of you

[smoky oak]

It's like what wizards tried to do with DND. They thought they controlled DND, and then they realized that without the community driving things forward, they are nothing, so they folded and gave in

[dense tapir]

This is what killed me with dreambooth and I left it with successes but hundred to over a thousand hours of wasted time. One thing I still smfh over is that you use a token to call a model with, right? what happens if SD already knows that word? no way to know if it does 100% and no way to change it after the fact so YAY, fuck me if I pick a word it does know as I get to find out when my model doesn't work. pick another word and retrain wasting 1-2h each time until a word you pick works.

that NEVER EVER made sense to me.

it bakes the token word into the model which is so primitive. At least give us a way to know before hand or a way to change it after

[static tusk]

did you see Nvidia launched Picasso, an online model building tool?

[dense tapir]

that known word bit me in the ass with DB so much

never even heard of it but what I am seeing out of Nvidia is impressive but online stuff of course

[static tusk]

Shutterstock and Getty Images are using it to build their own stuff

[dense tapir]

I honestly feel as if we are all tools being used and in the end it will all go POOF to online only with not even a kiss in the morning.

[static tusk]

open source will still exist, but it will lag behind in tech

Adobe will have its own SD built in, thats where most users will go

built into photoshop

[dense tapir]

Well, true but OSS has never been leading edge. Where is the OSS fork of SD? If there is one how is it doing? Of course that is a fork of OSS but you get what I am saying I hope.

[smoky oak]

Adobe is already saying they have their own AI image generator that uses work from no artists

[dense tapir]

I saw Adobe AI when its beta first began. Be afraid.

SD can't hold a candle to it

[static tusk]

atm I think its just a plugin and early days, but it will come standard for sure

[smoky oak]

There are few companies I would want to burn to the ground more than the pieces of shit at Adobe

[dense tapir]

I despise Adobe with a passion, but give the devil its due when something as powerful as it and photoshop.

[static tusk]

and sites like turbosquid will have text to 3D model assets

[dense tapir]

There are just a few companies I have worked with in the industry and Adobe, Microsoft, and Autocad were the absolute worst pieces of trash to deal with.

[smoky oak]

Adobe and Autodesk are both worthless trashy companies...

But at least Autodesk has some useful free tools like fusion360. I have grown very attached to fusion 360

I can also appreciate fusion 360 more having come from blender

[dense tapir]

DAMN IT 😦 https://www.youtube.com/watch?v=HzV1RS5Oc6I

Intel is not going to be the saviour we were hoping it would be until at least 2026 or maybe even 2030.

if ever.

My how the once mighty Intel has fallen.

@smoky oak Do you know of a way to change the lighting of an image either from the prompt or in post?

prompt would be better as I get a lot of warm and mostly cold lighting

[smoky oak]

Yessir

Just a moment, I am quite busy at the moment

[dense tapir]

No rush

[static tusk]

https://huggingface.co/spaces/cvlab/zero123-live Its a little crude but cool to see

[dense tapir]

crude today but give it 15 minutes

[static tusk]

yeah its like ka-blam and its on v3 and fully 3d

[dense tapir]

Amazing

[glossy herald]

yep. I haven't ran it yet, and I need to balance it to be less biased and add some chimeras if I can find examples.
it's some kind of a mermaid yes. I'll cut it out, but I found it was cross species enough to get a spot

[wispy nest]

I feel like chimera is going to be mess 😄
Most images are mix of whatever, additional heads or snakes can be anywhere too

Plenty in "chimera mythical creature" search
https://duckduckgo.com/?q=chimera+mythical+creature&t=ffab&df=w&iax=images&ia=images&iai=http%3A%2F%2Fimages5.fanpop.com%2Fimage%2Fpolls%2F422000%2F422098_1334321013353_full.jpg

But idk if any of those usable for training (idk how laws work)

none of them look realistic, unfortunately 😦

[glossy herald]

no, I didn't even find 1 that was realistic, that won't be something the model does currently...

I'm wondering if I'm taking the right path

using videogame renders maybe could be great

[wispy nest]

yea, probably better then drawings

That's realistic looking thing

[glossy herald]

yeah I saw that one

I didn't take it because of the human

here is the dataset I curently have, but yeah I didn't rework it at all yet, it was a first try last night ^^

and the rights on that one, I didn't bother a lot tbh, I was going on the "research" path here because like we said yesterday, there is a real lack of source for dragons or things like that currently

I'm not sure I like this necessity though

[wispy nest]

Underwater dragon is perfect in dataset

I think we have plenty of dragons in movies and games, I might try to get some later

Chimeras in the games is another story tho...I remember only 1 game with chimera and it's not the game with good graphics lol

[glossy herald]

I love that one ^^

and yeah that underwater dragon is really cool

some nice artistic representations in there, but balance is quite off, towards horselike creature I think

That one isn't particularly beautiful but creative 🙂

[echo ingot]

Morning all, is there a guide to up-scaling that anyone would recommend I read?

Ignore me, haha just seen the upscaling channel and it's pins

[glossy herald]

sorry I was looking for a comprehensive guide and didn't find one

thanks for the reminder

I didn't even check the pinned x)

https://upscale.wiki/wiki/Model_Database
this is a good list of current upscaler models out there too

[crystal delta]

FAQ: How do I use the Stable Diffusion website & what do all the settings do?

Check out these videos for a great overview of the website and how to use it! https://youtu.be/014J2Yo1aGI ｜ https://youtu.be/M3jCa6qTHpQ

[wispy nest]

yea, this one is interesting

[analog garden]

For pickapic, should I do what is closest to the prompt or what looks the best?

[wispy nest]

I like this too

This one I don't even know what that is, but looks realistic 😄

[wispy nest]

I was prioritising prompt

[analog garden]

Alright. I'll do the same.

[glossy herald]

not knowing what it is doesn't seem a problem here to me, as long as it feels a mix of animals

but I think there are too many horse based ones, I'm trying to find other things to even things out

[wispy nest]

I think there was a name for this thing too, but I don't remember what it is...

[glossy herald]

isn't that just a basilisk ?

[dense tapir]

spam :/

[glossy herald]

wrong channel to spam on when I'm here bud

[dense tapir]

LOL

[glossy herald]

yeah I banned it

[wispy nest]

no idea, he calls it cockatrice lol
https://ru.pinterest.com/pin/476255729344966172/

[dense tapir]

YES, a cockatrice

[glossy herald]

love its face

I should also try to find my old D&D master books

there was one with just monsters in it

[wispy nest]

damn, that's a nice serpent
https://ru.pinterest.com/pin/52072939433401805/

[glossy herald]

and lots of illustrations

[dense tapir]

I can't believe that after all this time we still can't action/automate a lot of stuff within Photoshop. what a PITA

a batch function for it would be wonderful

[wispy nest]

about serpents...that was one of my first attempts doing creatures
I gave up on it after some time, final version expanded alot (not this one), but also has lots of messed up areas...it took soooo loong to outpaint, inpaint things...

^ actually generated (and outpainted)

This one generated too

and outpainted \ inpainted a bit

[wispy nest]

we forgot about nine-tailed fox 😄
https://ru.pinterest.com/pin/351912463449685/

[glossy herald]

true !

the left one is clean

[wispy nest]

yea

Chimera

[glossy herald]

wow that one is cool

[analog garden]

Woo! First try!

[glossy herald]

yeah the text on pick A pic is just on top

can't wait for this to come out 😉

[wispy nest]

more creatures

[wispy nest]

technically idk if that's chimera, shouldn't chimera have a snake?
I think internet calls anything mixed up chimeras now 😄

[glossy herald]

it seems so now yeah

nice thunder bird 🙂

I took the first thunder bird and the black dragon in there, good finds

[wispy nest]

I think black one look more realistic

[glossy herald]

yeah but it's 480x352

a little small, I can upscale but not the best for training

[wispy nest]

and more contrast around will probably make it better for AI to recognise as one thing

hmmm

yea, original in low resolution 😦

[glossy herald]

yep, happens

I was looking for a good model

to make some more base pictures

but damn this is a dry subject

I found some creepy horror monster and some cute fluffy things

no inbetween

or like a model for 1 specific entropo character

[wispy nest]

yea

more birds

and more dragons

[glossy herald]

https://orkerhulen.dk/Core books/Monster Manual I.pdf

low res

but damn why isn't it high res x)

[glossy herald]

better than the phoenix I had

same for the dragons, I took 2 in there, thanks

[wispy nest]

more on dragons

sooo....

That's a weird one lol

[glossy herald]

I'm at 59 samples, I'll do a cleaning of the low quality ones

[wispy nest]

I want to find white fox with nine tails, but search gives me girls with tails =.=

well , that's something

[wispy nest]

if that's too many we can clean out too cartoon-ish looking ones too

[glossy herald]

nah it's good, it's more that I had some duplication in themes, and some 3 low res ones

[dense tapir]

Any one know of a graphic's program that will batch equalize histograms?

[wispy nest]

few more foxes

not exactly mythical tho...

[glossy herald]

I like the first one but it feels blurry

or artifacted

[wispy nest]

yea

[glossy herald]

I'll train it this afternoon 🙂

well a first try at it at least

I was thinking of starting from a photorealistic model, to help on that side, but I'm not sure it will be enough, given the dataset

[wispy nest]

yea...

making realistic only will take alot of time

not bad too

[wispy nest]

Which model you think?

[glossy herald]

Illuminati maybe

hum no, I don't have the dataset in 768

[wispy nest]

never even tried it, got annoyed by their over advertisement based on profanity 😄

[glossy herald]

I didn't really see them advertise anywhere tbh, but I don't see a lot of ads

https://civitai.com/models/1116/rpg

[smoky oak]

@glossy herald for your appreciation

[glossy herald]

seems fitting

[glossy herald]

THANK YOU !

[wispy nest]

There was youtube guy, idk if he made or just advertising it who claimed SD noise algorithm to be bad, so he started throwing this phrase "noise offset" everywhere, which people repeat without thinking...

But...he also claims his lora is enough to fix it.

In other words he's claiming SD algorithm is bad, but he fixed it...without actually fixing it and throwing links to his model and lora, just trained on darker images...

[smoky oak]

@glossy herald oh, and some curled tosies

He always has to have all his toes curled when he sleeps lol

[wispy nest]

RPG is trained only on clothing tho, it's not good at anything else, if you go off dataset it's trained on - it starts throwing cartoon looking images

I think there are 3 potentially good realistic models , which can do alot of different things -

• protogen, with the only con - it makes too bright images
• 3Guofeng3_v32Light - realistic + dark (usually)
• realistic vision (never tried it tho)

But there's another path too, dream models

• dreamshaper
  and whatever other of same sort was called
  (tried it few times, just wasn't my style, but might work with mythical things)

[smoky oak]

Realistic vision is my go to, and I have tied all versions of protogen. Realistic vision just looks way better for realism IMo

Also, realistic v1.4 just came out. Really cool model :>

[wispy nest]

yea protogen goes somewhere between actual realism and fantasy

[smoky oak]

The asshat creator finally got off his high horse and released it to the public after he got bashed on Civitai rightfully so

Had the nerve to license and sell a model made up of other people's work without citing them. Shitty person alert

[wispy nest]

oh yea heard about that, that's stupid lol, don't remember what models it was tho

[smoky oak]

1.4 was supposed to be exclusive to fantasy AI, but after the backlash he made it public. Now Fantasy AI is the exclusive place to host it

[wispy nest]

idk , I'm a fan of 3Guofeng3_v32Light rn, it fits my style the most.
Pretty realistic, can do alot of things, good at dark images, it's just too good

The only downside - it's Chineese and most people look Asian 😄

[glossy herald]

Gtg for now but I'll check those

Thanks a lot there

[wispy nest]

oh and another thing - it can do higher resolutions, I was doing 1024x768 pretty consistently.
I think there was another one with higher resolution which actually said it is trained on high res, but I don't remember which one was it, I have too many models rn to remember 😄

[wispy nest]

how should the webui-user look? i changed some thing to try and update python now I think i got it messed up.

[austere pilot]

https://youtu.be/-iwPVUzAWzk

I wonder if this is usable with img2img and if I can generate 1 different pose at a time?

[wise nimbus]

this photo taken by Leica M11 and Lens 35MM F/2, an indian 60years old man wearing white shirt and white dhoti and wear a white towel in his head and standinga alone by side of the road, in the light rain.

[tropic shell]

I haven't watched this, but you can use img2img to generate different pose angles---I've done it myself. You can do this with, and without keywords. It has a preference for some poses, but it can be done.

[austere pilot]

Could you give advice I have a certain character which I wanna animate using different poses but I never get good results. (Also I would love to turn the character in 3D then using other tools.)

God damned my 4G...

It's back

Kinda

[tropic shell]

I trained a bunch of consistent poses/models by combining 2d and 3D in 1.4/1.5, have been working on making new stuff for 2.1.

Sure!

[austere pilot]

I meant actual 3D

3D model

[tropic shell]

There are certain poses that SD is more prone to taking. That's why I personally trained my own models

You mean tips on 3D?

(I am actually working on tutorials for Blender.)

There are a lot of ways to turn a character 3D

[austere pilot]

Can you send me some? You mean as in 3D mesh right?

[tropic shell]

But tell me about this, so I can give you more direct advice

Yes, there are many ways to turn something 3D

[austere pilot]

Euh, I thought this tech was in early stages? I only know about a google collab that turns PNG to 3D mesh and EasyMotion that can transform the motion from a video into animation for a model

(Ping me with response pls)

[tropic shell]

Ah, through AI. Yes, as far as AI, there's a lot going on. But there are a lot of automated, or "easy" ways to turn something 3D, that have existed for a long time. They're not as curated as learning how to really model--there are programs that essentially allow you to draw around a picture, and it will put the texture on for you, but frankly, it's better to make the model yourself. Making a 3D model isn't that difficult, and there are already many programs, such as MakeHuman (opensource, humans), VRoid (anime, some restrictions may apply but you can make your own items, etc) to make it easier. You can then take those, bring them into programs such as Blender, pose them according to how you wish. You can then train on the images you create. Alternatively, you can simply use img2img to create model sheets. There are key phrases you can use, along with just using a character turn sheet, and a seed, to help turn the character.

The purposes of bringing in a 3D character, ofc, besides using the Blender plugin with SD, obvi, is that you can merge the 3D character with 2D styles while training, so you can get some very unique, solid stuff

[austere pilot]

I am learning blender at school but I am not that skilled

[tropic shell]

You don't need to be incredibly skilled to use it

Are you going for a human character?

[dense tapir]

@smoky oak Do you know of a way to change the lighting of an image either from the prompt or in post? prompt would be better as I get a lot of warm and mostly cold lighting

[austere pilot]

Yes one I already have designed with SD

[tropic shell]

Is it just one image?

[dense tapir]

https://www.youtube.com/watch?v=KaIYAjaowZI Just flat out no.

[austere pilot]

Ye

That's why I needed controlnet and that textual inversion

[austere pilot]

This one

(This is a previous version got some more decorations on another one)

[austere pilot]

this was my current final result

[dense tapir]

Been randomly crashing on colab and now I see why

[glossy herald]

hum...

it seems the dependancies are broken because of torch version there...

the first one is the strangest one

[dense tapir]

yep, and that is behind the scenes I can't update colab

[glossy herald]

damn

[dense tapir]

dev is on hiatus for 3 weeks too

I say dev I mean script kiddie who made this notebook

I am not being mean I am 100% he has no clue and is honest about it

[tropic shell]

Beautiful design. You could technically turn that 3d, and create more views of the character and train on that. That said, you might have an easier time using something like Make Human. That said. You can take 2D art, and animate it in Blender to get additional poses from it. You can split the image into various planes and import them. If you do this, you can get various views if your character, meaning that you could also use these for training or img2img as well.

@austere pilot

[austere pilot]

what is this Make Human?

[tropic shell]

One thing you can try, too, is, start with a similar looking character, and a sheet (I'll give you some of my terms)

And you go into img2img

[austere pilot]

that on auto1111?

[tropic shell]

And you start prompting til you get a base seed you like, then bring in your character, and use that with that seed, and the phrase, to see what you get

Yea

[glossy herald]

🐮Did you know that cows are actually expert coders and use a similar process to generate their "moo-sic"? They start by randomly selecting a base note until they find one they like, then bring in their musical character (usually a quirky cowbell or a moody trom-bone), and use that with the seed note to create their latest hit song! And the best part? They always use the phrase "udderly amazing" to see what they get!

[tropic shell]

If you go back and forth between promoting and img2img--not that it's necessary--but you can...and you will find you can really throw in some curve balka

[austere pilot]

You got any video on that?

also this model needs to be useable in AR

[tropic shell]

What a moosical take? 🐮 You know, cows are quite intelligent creatures, especially to sing so very well...

[austere pilot]

You recently got obsessed with cows XD, dw I been obsessed with them for a lot longer (in eating them that is )

[tropic shell]

Nah. I'll just give you some key phrases to work with.

[austere pilot]

mooohoooooooo

[glossy herald]

Haha, I agree! Cows are clearly multi-talented creatures, and who knew they had such a knack for music? Maybe we should start a band with them and call it "The Moo-sicians"! I'm sure they'd be udderly amazing.🐮

[tropic shell]

Cows are fantastic to talk about!

[glossy herald]

CoW

🔜

[tropic shell]

C o W

[glossy herald]

🐮

back under cover mate

[tropic shell]

🐄

moo

Roger, roger, on the case of the turnable character prompts!

[austere pilot]

I find invokeAI much easier to use, but sadly they don't support scripts yet, althought ControlNet is coming and node based UI as well

[last pawn]

hey evryones! I hope to find answers here because this tool drives me crazy! I managed to get this but I really can't find the way to refine it to have something consistent. Do you have any advice? Because "Restore Face" create really bad results, I also add VAE (MSE) to the process but it didn't change anything, I worked with ControlNet to have this result but I need something "PERFECT" and I think I'm FAR! If you could help me! THANK YOU in advance!

[last pawn]

(I prefer the body aspect of this one). On the 1st image I sent, I tried to work with inPainting only for the face but it sucks!

[austere pilot]

@tropic shell can you send me link to repo? can't seem to find the Make Human extension

[solemn topaz]

Not sure what you are asking for exactly.

If you want to create more characters in the same pose + body type, use controlnet's depth map.

You can also create a more verbose prompt from the photo using the "interrogate blip" and then generate an image with low denoise strength.

You can make adjustments with inpainting.

[wispy nest]

eh...that's not great way to do that...

I don't think you can do different pose of same character consistently, as you want it, no, that's whole point of this approach - you do 1 image with "many of same thing" - SD keep consistency between objects if SD thinks it's the same object.

There's easier way to do this tho, no embedding required
#🏞｜general-with-images message

I have a prompt in that messeage, but keep in mind - you want to go as high resolution as you can with this.

P.s. technically embedding should work too, it'll just need higher resolution too

[last pawn]

Thank you for your answer! I'm looking for something great. Do you see the body? I love the render I obtained. I'd like to have the same result for the face and for the hands. I see lot of amazing renders here and I'm wondering how they do that. I followed a tons of tutos but maybe I left something. My result SUCKS! 😅

[austere pilot]

https://github.com/makehumancommunity/makehuman-plugin-for-blender

@tropic shell this is also a blender add on