🏞｜general-with-images

[wispy nest]

we have a little faq on this, I'll start by linking it, but the #1072220168534642768 chanel is a good place

[crystal delta]

FAQ: I'm new here, how do I generate images ? Where is the bot ?

Welcome ! There is no bot currently to generate your images on discord. You may want to start by taking a look at the #1014939219904450590 channel. You can access Stable diffusion in different ways : 1️⃣ the official website, https://beta.dreamstudio.ai/. The easiest and fastest way to access Stable diffusion with 200 free credits. For any question on it, you can find help in the #1025467151206854736 channel. 2️⃣ Installing Stable diffusion on your computer. There are numerous projects that let you do that, and you will find help in the #🤝｜tech-support channel. 3️⃣ Running Stable diffusion in the cloud, through rented GPU services, using notebooks. You can find lots of them shared and discussed over in the #1011228442399883294 channel.

[wispy nest]

it should also answer Maximus non question : we don't have bots making pictures around here right now

[violet carbon]

Thanks 👍

[crystal bloom]

generating character concepts and turnarounds for a game I’m working on. Stable Diffusion v1.5.

[wispy nest]

yea...I did few test runs too

I heard phrase something like "sword is an extension of your hand" or something like that, but AI takes it literally pretty often

[strange cliff]

What prompt are you using for the head?

[crystal bloom]

great turnaround. did you use an embedding for your prompt?. I used one called charturner

[violet carbon]

Is stable diffusion 1.5 free?

[crystal bloom]

yes, also newest one

[violet carbon]

I'm interested but don't know much about how to access that

[hasty nova]

#1072220168534642768

[crystal bloom]

I work on Mac. I can send you the installation tutorial if you want.

[violet carbon]

Ohh okay, i got confused 🤔.I thought it was available online.

[crystal bloom]

you can try online too

[violet carbon]

I have a windows pc with old hardware

How do i use it online?

[crystal bloom]

https://huggingface.co/spaces/stabilityai/stable-diffusion

[violet carbon]

Looks kinda realistic 😉

[violet carbon]

Thank you so much

[wispy nest]

All images had same prompt, I got it randomly .
I think you can try to get heads by specifying it , but who knows, possibly you can get different emotions too by doing "emotion sheet" (not sure how people actually call it, just heard somewhere), but also...this might not work , since AI always trying to make similar objects look the same...eh

(High Quality:1.2), (Masterpiece:1.2), character design reference sheet for a game, 2d, girl with sword, red hair, full body

[wispy nest]

no, I think I was using just default 1.5 model , nothing else
(or maybe 2.1 , I don't even remember anymore, have only prompt saved and remember I had to do higher resolutions to get somewhat detailed characters)

Looking at the prompt rn, I'd probably remove masterpiece part from it and 2d, maybe lower high quality weight too...

[wispy nest]

multiple ways, you can find some help on using it online in #1011228442399883294 mostly, there are lots of possibilities. The space you got linked is a great starting point too. Also, you can make quite a lot of picture for free on the official website too, dreamstudio

[woven turret]

Batman

[young arrow]

I'm really excited about this and wanted to share with other Stable Diffusion fans. Today my first AI art album cover was released. It's for the band I'm in, but I've done some other covers since that aren't released yet, and I'm really hoping to get into doing more covers more often. Here's the first cover, for the album "Lotus Head" by The Fair Attempts. I made it with SD and several hours of corrections.

[rain gazelle]

Made some significant upgrades to my image viewer, now v4.0. 🙂 You can now customize the detailed view, add columns to sort by things like model and seed, and even turn off image preview in detailed mode if desired. These settings can be saved and persistent through relaunches, making it super fast and easy to use. And as always, very easy to quickly browse images and snag the metadata. Feedback welcome. 🙂
Source and binaries up on GitHub: https://github.com/GarlicCookie/PNG-SD-Info-Viewer

[cedar elm]

Ok, this is exactly what I need right now, how the heck did you achieve consistency please tell me! :)))

[earnest solar]

Darth Obscura looks over her planet

[mighty marsh]

doing some cyborg/mech characters on 3dmdt1 ckpt and the HyperFluid textual inversion. Love this textual inversion for adding fire/water/ice "type" characteristics to characters

[wispy nest]

I didn't really, it's just how SD works , when it drawing similar objects on one image - it's trying to keep it consistent.
So...in this case you want to go as high resolution as you can and hope to get decent result

[grizzled nova]

silo métallique

[chrome kite]

So cute!!

[earnest solar]

orb

[sterile kiln]

😍

[open sonnet]

Hi

[barren anvil]

Letting SD tweak my prof pic. Favorites?

[wispy nest]

I blended Genshin Impact with Devil May Cry
https://www.youtube.com/watch?v=fvT4uzjhc2w

[wispy nest]

inpaint?

[wispy nest]

Mission failed , Chief , we lost pigeons.

not inpainting, its a technique i haven't seen anyone use yet

ControlNet converts the image into lines, i edit those lines and then feed it back in

[wispy nest]

Yea I figured what you did (kinda), I was referring to another post

[opaque torrent]

Tried making a punk rock native American woman

[grizzled sage]

while i respect the use of using stable diffusion, your game needs art direction still. using a mish mash of different styles and aesthetics doesn't really make much sense for design language. i'd recommend having roughed out characters first in order to get consistent styles, and then put those roughs into controlnet to start shaping some sort of consistent art style with actual direction to it

just firing cherry picked generations into a game will look as bad and cheap as asset grabs do currently

[broken grove]

also, finally generated something that got into the gallery of an ai art server

[amber smelt]

What can I add to my prompt/negative to avoid 'hallway' shots like this?

[young mountain]

Trippy!

[amber smelt]

I like it some cases but it seems like every industrial indoor prompt I give just makes something like that :(

[young mountain]

I think its an issue with training

Prob training image of factory contains lots of same machines being lined up like in your pic

[cursive lance]

Í can't smell colors 😋👌

[cursive lance]

Wooaa is beautifull

But this is more best

[crisp cave]

The last shepherd

#1072220168534642768 sheperd

#✍🏼｜rules-and-tos

[sterile kiln]

Controlnet

[smoky oak]

kauhr

[dense tapir]

That car was a mix of a corvette and brad pitt. That car shown is classic so I lol'd.

[smoky oak]

wow lmao

[dense tapir]

1970s all female rock band album cover

Oh, my I mixed two of my loras

[smoky oak]

anybody know how to get beefy bodied men, rather than extremely lean?

I have been trying for a damn long time now, and I have found no reliable way

[craggy rune]

nice

[dense tapir]

Thanks

[sterile kiln]

You can do better general, i know you 🤣

[wispy nest]

I see it is control net, I'm wondering how you made it draw inside other pic, not changing it.

[dense tapir]

LOL. I have tried three times now to train a lora and that is the best I could get. I mean loss rates through the roof (2+) that I have never ever seen like that before.

that is the lowest I could achieve

[robust path]

Multi controlnet. https://www.youtube.com/watch?v=cNIHZInV3mg&t=134s

[jolly crater]

1970s all female rock band album cover

[sterile kiln]

I noticed LoRA models (mines indeed), need a low CFG scale, if i set hifher than 6, it's less and less like

[sterile kiln]

It changes it, look on Tifa pict, her skirt a red reflects from neons. So its really mixed up. To know how to i did : https://www.youtube.com/watch?v=MDHC7E6G1RA

[dense tapir]

Noyhing to do with that this is the training loss and even when I went back to what has worked for me (10 img to 6) EVEN increasing the steps I still ended up with horrible loss.

I am not sure how to reduce loss as more steps is not the answer.

[hollow grotto]

why am i the only one that wants to generate cartoons

[simple prism]

Red Square, Kremlin, flowerbed, buildings, Kremlin Palace, churches, towers, politics, culture, tsar, Soviet Union, president

[dense tapir]

Found the culprit

[dense haven]

Dumb question: how do I sharpen a blurred image? The original is blurred, but it's still blurred after I reimagined' it with controlnet!

[dry crow]

you need to upscale it, with img2img sd upscale script

[stoic pivot]

Followed @regal belfry advice to use mittens to get nice hands 😄

[grizzled sage]

if thats the original resolution of the image, that's the issue. low pixels == low details

[grizzled sage]

na'vi hands 😮

[stoic pivot]

Todays random image.

[regal belfry]

even 2 billion movie cant do hands all hope is lost bro

[silk horizon]

Hello,how and where I can do an illustration please

[stoic pivot]

Well, another random image.

I think the prompt for the giraffe was something like "grunge futurism rendering, a tall (girl furry giraffe) with long legs standing on the ground in a ruined city after a war, character side portrait, (seen from side:1.3), dof, bokeh, light rays, environmental dust, foggy, full body digital Furry art, rule of third"

[craggy rune]

u like furry

[stoic pivot]

Anthropomorphism.

Like to use animals in human roles to tell a story, with a giraffe we know that it is a part of Africa that been in war but that life still is around and is hope for the future.

[regal belfry]

my main issues with furries is everything looks terrible

hopefully with ai art

everything wont so i wont have to hate furries again

[stoic pivot]

This is same, we know it is Europe and the wolf tell us that it is something we fear, the weapon make him look even more dangerous.

[regal belfry]

see that actually looks good

[craggy rune]

so realistic

[regal belfry]

not whatever this is

[stoic pivot]

I do not see me as a furry, but I grew up with fables and that is also a bit of North mythology, to take the shape of an animal to get adventages.

[wispy nest]

that reminds me of my "bad cosplay man" model

https://huggingface.co/Guizmus/Experiments/resolve/main/bad cosplay man.ckpt

[craggy rune]

LOL!!!

[wispy nest]

😂 the frog has me rolling

not the best quality model, but a sure fun night with friends to be honest

[regal belfry]

irl pepe

[wispy nest]

I have lots of "party models" like that https://huggingface.co/Guizmus/Experiments

this one, HogLeg, was done using the trailers of hogwarts legacy before it came out ^^

sorry, trip down memory lane :p

bad cosplay man as a car

[craggy rune]

[wispy nest]

ho misa !

[stoic pivot]

Just for that, now I need to try cosplay and I plan my outfit...

[wispy nest]

xD

[rancid juniper]

I deleted everything from my google drive! And im running the Automatic 1111 web ui

after this error it shows done

But in the end cell, I get this error

What is missing?

[wispy nest]

#🤝｜tech-support seems more fitted for things like this. Not sure, I would start again from a working notebook like this one https://colab.research.google.com/github/TheLastBen/fast-stable-diffusion/blob/main/fast_stable_diffusion_AUTOMATIC1111.ipynb

[stoic pivot]

Do you have correct version of Pyhhon and SD 1.5 model?

[wispy nest]

bad cosplay man as a poney

[stoic pivot]

Ppython 3.11 is not good, it shall be 3.10.6 for Automatic1111.

[wispy nest]

same prompt, using your image as controlnet

let's try memes

[stoic pivot]

Now I want to organize a cosplay event for only bad pony outfits.

[wispy nest]

bad cosplay man as a frog

almost there

[wispy nest]

yes ! at last I got it

"fake scribble" was the mode

Me gusta costume

[stoic pivot]

I cant meme 😦

[wispy nest]

but you can frog !

[grizzled sage]

this is good cosplay man

[stoic pivot]

Frog?

[wispy nest]

yep, it doesn't keep it bad since I activated controlnet

those are quite impressive cosplay even

[grizzled sage]

typical grams with a photo of her grandchild on her jumper

[stoic pivot]

Some may upscale this to see what the text say.

[craggy rune]

I can't stop difussing, this is addicting

limitless result

[wispy nest]

only your imagination and energy have limits now

[stoic pivot]

Animal you can see in aftrca but you cant remember the name

[wispy nest]

bad cosplay man is still here

this one makes me laugh x)

[wispy nest]

damn control net is strong

never would have been able to get such a close result with img2img before

[stoic pivot]

Try meme this.

[wispy nest]

I'm trying bad cosplay man as a goat mermaid

[stoic pivot]

😄

[wispy nest]

loool

[stoic pivot]

Simpler if seen from front?

[wispy nest]

I'll change model soon, I'm using bad cosplay MAN, it doesnt help

buzz and woodie

[stoic pivot]

Bad names for toy, unless it is moms toys.

[wispy nest]

lol that reminds me of a fun story

My mother once bought a gift for her sister, a cheap electric toothbrush

but she was a little simple my mother, and didn't see it was a "vibrating toothbrush", nothing moved, only vibrated 👀

never got the courage to gift it once she got what it was x)

[rain gazelle]

For any using my image viewer, I just updated the source/binaries on GitHub with a bugfix. (Crash could occur when copying favorite images while in Detailed View mode). Grab the updated v4.1 and all is well 🙂
https://github.com/GarlicCookie/PNG-SD-Info-Viewer

[wispy nest]

also that

[rain gazelle]

You like my chicken? haha

I love him

[wispy nest]

I tried to modify him a little, I love him 🙂

I got some realistic ones too

[rain gazelle]

I like making bird photos

you can get some great ones

ok, imma go find some food

[wispy nest]

bon appétit !

[sterile kiln]

if you are a woman prepare yourself, I will ask you in marriage

[sterile kiln]

@rain gazelle How it is supposed to works?
Found Exe file x)

[barren anvil]

Prompt: my 3 year old "A robot cat playing an accordion" / Concept art: Stable Diffusion/ Polymer clay toy: Me. For now, at least, we are all working together 😉

[floral flicker]

These were so dumb and easy to make, I did these on day 2 of using SD with Controlnet, simply used the depth model w/ very simple prompts

Top left is input

[twilit cosmos]

I love the pikachu and black bug haha

[rain gazelle]

You lost me?

[rain gazelle]

Yes you can get the binary files under 'releases'. 🙂

[floral flicker]

Thanks 🙂 It made some prompts very very well… that was the only good pikachu out of like 10 gens. But it cranked out lots of “spiders” (black bug)

Roses, watermelon - were cherry picked. It made amazing gremlins though, and Ewoks

Photo of an Albert Einstein doll (inpainted on right)

[sterile kiln]

Well I concluded that I was right, put on your dress, I'll heat up the Cadillac

[rain gazelle]

I don't follow

[sterile kiln]

(Your tool is perfect)
I'm a perfectionist so I would say that a dark theme is missing 🙂

[rain gazelle]

mmm dark theme, I could do that pretty easily

I guess it doesn't inherit the OS light/dark settings? 'cause it might...

it does use built-in control colors already. Not sure if that works with the OS or not.

[sterile kiln]

It's French humor, basically when we appreciate something, we like to say that we're going to marry the person in question, nothing twisted xD

[rain gazelle]

Ah

oui oui

[sterile kiln]

Yes, I already had a discussion with a dev from Qbittorrent, who said that Windows Explorer posed a problem in this kind of case.

[rain gazelle]

Do you use dark mode on Windows already? I'd be curious if the program automatically inherits that or not.

[sterile kiln]

I am, and its not.

[rain gazelle]

ah ok guess not 🙂

[sterile kiln]

I think that explains why even today, Qbittorrent still doesn't have a Dark mode, when everyone asks for it x)

[rain gazelle]

It's not that hard to implement.

I'm not the best at color design though 😐

[sterile kiln]

In any case, don't hesitate to let me know when you have made an update on this, I will be your 1st beta tester if necessary! 🙂

[rain gazelle]

Ok

Make sure you check out the features and stuff. e.g. Left-click zooms, middle-click-and-hold pans, right-click the image copied it to clipboard.

mousewheel over image goes back/forward

Right now I'm adding a feature to quickly set something as wallpaper 🙂

[sterile kiln]

What language are you using? C# & Windows form?

[rain gazelle]

Yes

[sterile kiln]

Cool

try to find faults with it but apart from this missing dark mode I don't see anything, no more drag&drop from hell on jimpl.com, especially since when you have a double screen with 200,000 shits open it drives you crazy

[rain gazelle]

Thanks 🙂 I'm glad you like it!

I am just finishing up the wallpaper options 🙂

[rain gazelle]

@sterile kiln working on dark mode. Not complex, but kind of a pain to design these colors. I'm not great at color design 😄

[sterile kiln]

https://tenor.com/view/god-pray-jesus-christ-sign-of-the-cross-gif-5701179

[sterile kiln]

easy, find a software where the colorimetry is interesting, then make screens in PNG-24 (so as not to degrade the colors), then on Photoshop, you can have these colors in Hexadecimal

[rain gazelle]

@sterile kiln - version 5.0 has been published.
Fixed some UI elements
Added functions to quickly set images as wallpaper (or clear back to your default selected color)
Added dark mode. Persists as part of saved settings.
Updated docs.
--Dark mode isn't perfect. There are some UI elements that are a real bugger to change in Winforms C#. This is as much effort as I'm willing to contribute to this presently. 🙂
https://github.com/GarlicCookie/PNG-SD-Info-Viewer

[glossy herald]

woooo ! Dark mode even not perfect is always perfect

[rain gazelle]

🙂 thanks. I hope you like it

[sterile kiln]

https://tenor.com/view/jesus-god-dj-dancing-party-gif-11368127

[sly harness]

#1011634831467221033 https://i.imgur.com/VYvzlfK.png Draw the shortest line from point A to point B in the white area

[wispy nest]

I went to the planet and traveled to the far reaches of the farthest outcropping of rock and strange stone. There I found a creature of indescribable horror, seemingly infested with a strange parasitic creature. I don't know what it started as but it's not that now...

[lilac pilot]

These pretty decent quality?

Like the Sd upscale extension

[crystal delta]

FAQ: What is Stability AI?

Our vibrant communities consist of experts, leaders and partners across the globe. They are developing cutting-edge open AI models for Image, Language, Audio, Video, 3D and Biology.. AI by the people, for the people. Learn more here

[cunning geode]

the photography half of my model is now officially finished and working well with only slight issues that may even be fixed in version 2.0 of my model, but this is good enough to be released.
the other half of my model should finish around late evening sunday or early morning monday provided there arent any serious issues with the model like undertraining or issues with the captions, so that i may release the full model early monday still or at least monday after im home from work.

[rain gazelle]

48 gigs nice

[smoky oak]

man, I forget how good some models can do for realism

it never gets old haha

[cunning geode]

which model is that?

[smoky oak]

This is realistic v1.3 on civitai

It does damn good when you prompt it well haha

[smoky oak]

❄️ Gato ❄️

@cunning geode

https://civitai.com/models/4201/realistic-vision-v13

[smoky oak]

@dense tapirI would love to know how many images you have generated lmao

like, total

I am at about 12k

in 19 days

[royal stag]

Is there any way of creating progress shots, like here for example? Im learning to draw and would like to teach myself how illustrations come together

[late sentinel]

futurastic garage artwork

where i can generate pictures from text?

[blazing crow]

You mean what program to use? or how SD functions?

[wispy nest]

May I ask why do you need so much memory

[glossy herald]

training mostly I think, you can go higher batch sizes for better quality like that

flatface

[glossy herald]

let's try more Fry

can't get that one right

bad cosplay man could maybe help

YES

lol

can do better

this one is nice

miss a finger though

[gritty bluff]

#1047610792226340935

[glossy herald]

I'll be your bot :p

[gritty bluff]

Da pra melhorar

[glossy herald]

sorry I speak only english and french

[gritty bluff]

I would like you to improve this image for me to use in a logo

I don't know how to use discord nor are these Boots I'm new here and I care about the subject I need your help to improve this image to use in a logo

please generate several variations of this stylized image so I can get it here better

[glossy herald]

first, thank you for this effort so I can understand you

I can do lots of variation, and I'll post some random ones

but do you have a style you want ?

I am using your base image and just asking the AI for "an eagle"

but I can add a lot of things

"cartoon", "taken on an iphone 6", ...

there are lots of possibilities

also for the background, we can draw anything

[gritty bluff]

it's just the Eagle's head that I need, it can be from the front or from the side, but preferably from the front. it can be simple or stylized in color or black and white, I would like you to try to generate it in several ways so that I can choose the best one, in advance I appreciate the help, thank you very much

the Eagle's neck has to be small because I'm going to amend it into a logo I'm going to post a picture of the logo here for you to see, it's easier for you to create the eagle's head

[glossy herald]

if you want more variations in one style, ask for it, I'm trying all the styles I have

this is all the types of style I have in stock

[gritty bluff]

just a minute I'm fixing a picture of the logo so you can see the head of the Eagle that I want to change

[glossy herald]

sports logo style

[gritty bluff]

I want to get a head similar to the one in the logo but prettier so I can replace it

[glossy herald]

can you get the .png or the .jpg with just the eagle head ?

the background is a problem for me

I can't do variations on this

[gritty bluff]

if you make the Eagle's head with a transparent background it's better for me to cut it later

[glossy herald]

yes but I need this eagle head

the one you have

I can make it 3D or whatever

but I need it as a base

[gritty bluff]

sim eu vou obter somente a cabeça da Águia em png ou jpg para você só um minuto

[glossy herald]

let me check google translate

[gritty bluff]

yes i will just get the eagle head in png or jpg for you just a minute

[glossy herald]

ok no problem

[gritty bluff]

my cell phone is a little slow, please wait

[glossy herald]

Cara, eu fui hackeado por um brasileiro nem 24 horas atrás...

Não é sua culpa, mas ainda dói

[glossy herald]

thank you

[gritty bluff]

Caramba friend I'm sorry you got hacked this is really bad I'm sorry for the Brazilians

[glossy herald]

let's go

[gritty bluff]

this head has been modified a bit by me maybe i can get a better example

[glossy herald]

like I said, not anything to do with you, but I couldn't forget it when google translate said "bresil"

sports logo style

"portrait wallpaper"

[glossy herald]

I'll start over without the crown

[gritty bluff]

it can only generate the head of the eagle with a transparent background for me to cut and place in the logo that I already have ready

[glossy herald]

this is perfect

sport logo => bad

Portrait walpaper => OK

liquid multiverse style => :/

exploding nebula => :/ 👍 :/

Headshot portrait => better

[gritty bluff]

I need something simpler friend just the Eagle's head that should be colored nothing more transparent background

[glossy herald]

ok I'll try simpler

[glossy herald]

nice !

[gritty bluff]

so you'll have an idea of ​​what I want to do but the Eagle's head is still ugly it's not pretty it didn't match the logo

[glossy herald]

hard to stay closer with something realistic

[gritty bluff]

I need you to Generate an eagle's head that matches the rest of the logo I only need the Eagle's head the rest of the logo I already have I'm going to cut the eagle's head and paste it on the logo

[glossy herald]

the rest of the logo

you mean the "grey transparent" wings ?

[gritty bluff]

it doesn't have to be very realistic, it just needs to be beautiful, it can even be stylized as a silhouette, etc.

it has to be the head of an eagle that will match the logo you can even see other examples of eagle head on Google to generate

[glossy herald]

I'm getting closer ?

[gritty bluff]

it doesn't have wings it's just the head of the Eagle that I need the logo I already have separated

[gritty bluff]

I think it could be a simpler type of drawing because we won't be able to get too close to reality. I'll use the ones you made so you can have an idea

just a minute I'm going to put some that you made in the logo so you can see how it looks

[gritty bluff]

this one you generated was good, see, I need something simple

[glossy herald]

can you remove the background ?

without the hands

and without the text

just the sketch of eagle head

and wings

[gritty bluff]

Yes I will remove the background and send you only the head

[glossy herald]

I mean, head + wings + torso

I could make a real eagle in a flat design

and you can cut the head

but it will help the AI

[gritty bluff]

just a minute I'm trying to copy the messages you wrote to translate

[gritty bluff]

just a minute I'm trying to copy the messages you wrote to translate

[glossy herald]

this one is the better

[gritty bluff]

you can only focus on the eagle's head, the rest I have ready, no need to generate, just generate the eagle's head with a transparent background

[glossy herald]

yes, I understood this. But, having the body too, it seems to help the AI get what we want

[gritty bluff]

it could be the head of an eagle in that style that you made and I put it in the logo and it turned out good

I understood with the body it is easier to generate and then I can cut only the Eagle's head because I already have the body

[glossy herald]

it can be easy to :
1/ cut the head out of one of those flat design
2/ upscale it

[gritty bluff]

yes it is improving but there is still a long way to go to reach the final result

[gritty bluff]

the ones I like I'll circle around so you can focus on them

[glossy herald]

OK but
I need more direction in the style you want to continue
I'm just doing random things right now
You said here "turn out good". What is "good" for you ?
The only thing I got is that you want "Flat" design of an eagle head

[gritty bluff]

then try to create the realistic eagle head in front of profile

[glossy herald]

I'll go offline for now. This is not an answer to what you are looking for.
I need to feed words to the AI.
I can't circle anything
The ones you show me are so different, I don't understand what is a "good" eagle for you
You need to answer yourself that, or I cannot ask it to the AI.

[gritty bluff]

ok my friend, the good head would be the one that I'm passing below variations of it to try to improve there but in the same style

#1072220168534642768

[gritty bluff]

anyone available who could help me create this image based on the examples i have given

#1010934719455707218

[dense tapir]

I stopped counting as it filled my harddrive so I turned saving off

[echo garden]

holly, the low contrast you can achieve with the new offset noise method is insane

[smoky oak]

anybody know how to run a public version of SD on anything other than gradio? I am tired of all of gradios problems

the whole site is down right now, and I can't run SD because of it

[fiery carbon]

I have used a 2D background image and a posed character in iClone and feed the render output (640x768) into Multi ControlNet using canny and openpose models. Took me a bit of time to figure this out.

[fiery lichen]

I hope 1 day we also get posing models just for fingers so that...you know.

[fiery carbon]

No kidding, i am literally waiting for hand / finger models or loras and also for feet 🤪 I already thought about training something myself but i am short of time 😿

[lilac pilot]

Datsun 👑

[vale egret]

#1072220168534642768

#1072229020520947753

#✍🏼｜rules-and-tos

#✍🏼｜rules-and-tos hd

[gleaming trellis]

Does anybody understand the firewall joke he was making here

[gusty cargo]

I've been trying to get this image more realistic while still retaining a painting style

I do want to make the buildings more detailed

[smoky oak]

@thick sundial

These are results I have gotten from using the model I am linking

[thick sundial]

OH GOD DAMN whats the model?

[smoky oak]

It issssa

Just a sec

https://civitai.com/models/4201/realistic-vision-v13

here it is

my favorite model, hands down

[thick sundial]

thanks

[smoky oak]

@green plover

[smoky oak]

@green plover

[quasi mist]

ChatGPT edited basically made my prompt from scratch as I asked for 5 words and it wrote 4 paragraphs from that. I then fed it to the AI. The results speak for themselves

[green plover]

dumb question. did you restart webui-user.bat after you edited it?

[smoky oak]

I closed out of it to make sure nothing was interfering

let me try again

[green plover]

That's why it can't connect. it's not listeninng

[smoky oak]

restarting it

@green ploversame problem ._.

wait

I didn't save the bat lmfao

trying again

[green plover]

did you add --listen to the bat?

[smoky oak]

I did not lmao

sorry, i thought I was just adding the --gradio-auth

[green plover]

you need the --listen too

That should open it up

[smoky oak]

ok, re-restarting

[green plover]

We will get it working

[smoky oak]

🤞

oh, it asked for firewall permission, thats a good sign

IT WORKED

@green ploverWOOOO

heck yeh

[green plover]

ok PM me the URL and I'll try to hit it from outside your network

[wispy nest]

I don't think one should give too much to the ai's work :P
modern bedroom would do for those results, or maybe add something like wooden to get the details :D

Though it's good that you liked the results, that's the important part

[smoky oak]

@green plover 800 images at 512x512. The highest you can do in a single batch

took 24 minutes

though the image sitch/grid processing took another 3 lol

I usually go up to 64 at a time for normal gens, but I wanted to go extreme lol

I can't send the full res image, its 390MB

[green plover]

Nice

[smoky oak]

1.8 seconds an image, not bad haha

thats 20 samples

so just over 10 samples a sec across each batch of 8

[charred bough]

does anyone know why auto1111 does this to all my img2img outputs?

theres an option in the settings that applies color correction to all my results to match the original colors but disabling that doesnt seem to fix the problem

[smoky oak]

beep

[stoic pivot]

A problem with AI! I wrote "erotical art" to see how long time it would take for the AI to generate a male, but a swirling circle have come up around 70 times, but no males...

[hasty nova]

Did you hit "apply"? I often forget that part..

[magic zephyr]

open pose hand model

where to find

[dense haven]

For Kohya LoRA training, one tutorial mentioned i should put 20 epochs here. On my RTX 3090 with 1 epoch this is taking two hours. Surely i must have given it bad parameters? i'm also using 128 network rank and 128 Network Alpha.

[slow iris]

How many repeats? That does sound about right though

[stoic pivot]

In a prompt that should never been written "photorealism, a fat (male) furry cosplay as a (latex Fluttershy pony) in a conference center, by Jeff Koons" I compare Img2Img upscale including LDSR.

And the original image

[glossy herald]

that gives me vibes of BadCosplayMan

I'll load the model if it decides to start...

[stoic pivot]

Some has to do it, when it happen do not blame me.

[glossy herald]

Bad Cosplay Man as a poney

[young mountain]

guys does it mean I can make it faster?

it says vram 8/18.9 GB

how can I make it do faster?

[dense haven]

100 steps. steps mean repeat, right?

[glossy herald]

no, steps usualy = repeats multiplied by dataset size

and yes you can train faster/with better quality usually using more VRAM

not sure in LORA, but in what I tried, it's called batch size

you basicaly train more than 1 picture at a time, making it divide the total step count by as much as your batch size

basically makes better quality, and a little faster

(there are less steps to do in total, but each step takes a little longer too)

[untold tinsel]

/dreamt prompt blob:https://discord.com/25115604-52d1-4901-8739-93a95ded0c4f Femme brunette futuriste 4k réaliste

[glossy herald]

sorry but there is no bot dreaming for us around

[crystal delta]

FAQ: I'm new here, how do I generate images ? Where is the bot ?

Welcome ! There is no bot currently to generate your images on discord. You may want to start by taking a look at the #1072220168534642768 channel. You can access Stable diffusion in different ways : 1️⃣ the official website, https://beta.dreamstudio.ai/. The easiest and fastest way to access Stable diffusion with 200 free credits. For any question on it, you can find help in the #1025467151206854736 channel. 2️⃣ Installing Stable diffusion on your computer. There are numerous projects that let you do that, and you will find help in the #🤝｜tech-support channel. 3️⃣ Running Stable diffusion in the cloud, through rented GPU services, using notebooks. You can find lots of them shared and discussed over in the #1011228442399883294 channel.

[dense haven]

what's called batch size, repeats?

[glossy herald]

no, batch size is just called batch size

it's "how many picture to train on at once"

and repeats is called repeats

it's "how many time do you want to teach each picture to the model"

[dense haven]

then does repeats mean epoch?

[glossy herald]

hehe

epoch is another thing

sometimes, the training tool splits the training in multiple epocs

in this case, each epoch contains the same number of repeat

it's another multiplier on top of all the rest if you'd like

imagine all we talked about before was on Epoch=1
with Epoch=2, you do another wave of training with the same params basicaly

[dense haven]

I see people saying 20 epochs, 10000 epochs, and then some tutorials tell me 1 epoch 1 save every n epochs lol

[glossy herald]

well, it all depends how you want to set it

some tools set repeats on 1, and do a big epoch count

some other tools do the contrary

the total step count and the learning rate will be what's important mostly, the rest is just "in what order do I present my pictures".
An epoch main goal is to guarantee each picture has been presented the same number of times to the model, that the repeat count of this epoch is over and that, statisticaly, we trained as much on each picture

so one approach or the other is just changing the bell curve, the statistic distribution of training order

[dense haven]

Thanks so much. Why do you learn this? I feel like there's documentation somewhere

[glossy herald]

well, I learned this because I wanted to master Shivam, a dreambooth tool, and then EveryDream, another one, and then another one, ...

and I tried making a UI for those, but I didn't finish

in the process, I still learned quite a lot on how it all works

all are based on the same script, but they all upgrade it in their ways (I'm talking about Dreambooth based training)

and it's just like..; I can't stop lol

there is no master documentation no. You can find bits and pieces, usually specialised on one tool

[dense haven]

It's quite addictive to play with tech that is new to virtually everyone

[dense haven]

If a face I trained on lora does not look like my subject, do I increase my learning rate, or steps, or network rank, or network alpha?

I could experiment with each parameter, but that's gonna take forever

[radiant orbit]

someone get this error when trying to generate with training lora?

[outer spear]

Getting lots of images like this, what negative prompt can i give to avoid this?

To be specific her facing back

sorry lol that question was vage

[dry crow]

haha

i wanted to ask rn

@outer spear looking at viewer will turn her around 😄 (as positive tag)

[outer spear]

looks like back, showing back as negative prompt mightve done it

[outer spear]

But I also already had "facing the viewer" which i changed to that, so could also be that

[dry crow]

for me looking at viewer always works

you can also try front view

and you need a vae for that model

[outer spear]

which model am i using

[dry crow]

Maybe AnythingV3 or AbyssOrangeMix

[outer spear]

yeah orange mix

[dry crow]

Then you need the orangemix.vae.pt file and put it together with the model

Here is the file:
https://huggingface.co/WarriorMama777/OrangeMixs/tree/main/VAEs

make sure to rename it to match the model

[outer spear]

i got it

thanks

looks like it got brighter

[glossy herald]

Damn, I got the text. even with controlnet it wasn't having it

[wispy comet]

Hi, I get this background when using depth(controlnet). any ideas?

i tried different models. low vram config...nothing

[smoky oak]

wow haha

[dense solstice]

Sorry for the dumb question but are you changing the seed or the Denoise Strength is too low?

[glossy herald]

Steampunk shoes

[grizzled sage]

https://tenor.com/view/super-mario-bros-movie-mario-luigi-gif-21977403

[glossy herald]

Wooden Nike

[smoky oak]

LED cyber nikes lol

bruh-

[grizzled sage]

bk nikes

[smoky oak]

THE NIKE SUBBBBBBB

[glossy herald]

https://huggingface.co/Guizmus/Experiments/resolve/main/NikeAirShoes-v1.ckpt

Activation token : NikeAir Shoes

[smoky oak]

I am just using a standard model

[glossy herald]

I do have stupid models in stock for sure

like burgy.ckpt

only does burgers

[wispy comet]

denoise strength is 7 and seed is different everytime but i get that "style" of background

[smoky oak]

I feel like some models would do better not even specialized

[smoky oak]

this is just Realistic Vision v1.3

DAMN IT

There

trying steampunk

realistic vision v1.3 really is just good at everything man. This model never disappoints

YOOOOOOOOOOOO

BRUH

These shits are CLEAN

[glossy herald]

nice !

I had to put so much weight for pipes and gears to start appear

[smoky oak]

I am trying to see if I can get them

oh my

I weighted a little too heavy lol

lol

[grizzled sage]

that's badass. using control net?

[smoky oak]

nope

[grizzled sage]

fun!

[glossy herald]

someone cosplaying as a shoe

[smoky oak]

just raw prompting

[smoky oak]

lmfao

[grizzled sage]

it really knows the nike logo well

[smoky oak]

oh yeah, it definitely does haha

ok, these are fire

I fucking love realistic vision man. Hands down my favorite model

lemmie do a high res gen

[grizzled sage]

enhance

[smoky oak]

ok, messed it up a bit haha

went a little heavy on the gears lol

[grizzled sage]

banned from olympic competition

[glossy herald]

not sure what that is

but yeah

prompt was "bad cosplay man as a giant glass of water"

that's a correct cosplay

[smoky oak]

They probably weigh 20 pounds as is haha

[glossy herald]

bad cosplay man as a giant slice of pizza

[smoky oak]

LMFAO

I'm trying to make Satanic Nike's

I mean, I am not disappinted

[smoky oak]

NOoOoOoOoOooOoOO

[glossy herald]

I love bad cosplay man

[smoky oak]

yoooo

those would match my other Satanic clothes lol

[glossy herald]

surfing incognito

[smoky oak]

good fit with the satanic wallet

(real photo, not generated lol)

I feel like I have to specify, cause man sometimes these AI's make photos lol

[glossy herald]

Cosplaying as toilet paper

[smoky oak]

oh FUCK YEAH man

STOP WITH THAT OMG

WHY

There

I fucken love the occult man

[tawdry anvil]

Noise offset is no joke 😄

[smoky oak]

so, would you mind explaining how it works?

Does it offset the average value towards black or white?

aha, that IS what it does

amazing

[tawdry anvil]

Yes, I basically watched a video on it and can't describe it beyond the laymens terms of the flaw being the pixel noise is averaged on snow and dark scenes

[smoky oak]

I have been trying to get night time photos for a while now, I wonder if this will help

[tawdry anvil]

and this process of these models and negative prompting inversions corrects it

its likely an issue midjourney solved awhile back, but now its properly solved 😄

so with control networks and this, good luck MJ 😄

[smoky oak]

MJ is already outclassed in resolution and single subject work for sure

I see soooo many people on fiverr trying to sell works they make in midjourney, and they are just so far behind SD in quality

[tawdry anvil]

with the inpainting krita plugin I was kinda over MJ after a month, my 2nd month id totally exhausted and was bored of it, mind you thats 50k images later too

[smoky oak]

I made about 15k when i had MJ

then I ditched for wombo dream cause it was way faster, unlimited, gens, and did good for what I wanted

[tawdry anvil]

I don't just feel SD is catching up and surpassing it, but its also more fun tbh lol

[smoky oak]

I just simply can't agree. I don't think MJ makes results nearly as high quality as SD

or at least my specific SD generations

[tawdry anvil]

well the lighting fidelity issue with offset noise goes a big way towards that

[tawdry anvil]

but its always going to depend on what you are trying to make and how, I think MJ has been easier, but not better

[smoky oak]

oh yeah, MJ is easier, but nowhere near as diverse

[tawdry anvil]

yeah photorealism for instance, my main focus has always been fantasy art so it swings about a bit

but started with SD in python like 6 months ago, so SD at the core

and don't find the MJ prompting etc as fun

[smoky oak]

also, I don't know if MJ is faster than it was before, but it used to be so slow man

like 1-2 minutes an image

for $30/m

[tawdry anvil]

Oh god the month at least the first month I got it they were always upgrading servers

was so so slow

its 60 dollars for stealth now

[smoky oak]

are you fucking kidding me? that is absurd

fuck that man lmao

pay monthly for a server instance and generate images way faster, and higher quality than MJ using SD for a fraction of the cost. jeez

[tawdry anvil]

if you automate 24/7 its not so bad lol

[smoky oak]

do they have unlimited gens now?

[tawdry anvil]

on relax yeah

which is slow

[smoky oak]

wow, so they still have fixed gens on plans that expensive? fuck that lol

[tawdry anvil]

but if you make a bot in chat gpt, well I didnt get to 50k images by manually typing

[smoky oak]

fair

I did get to about 15k manually typing 😅

[tawdry anvil]

so I feel I extracted the value at least

yeah I just setup an ancient laptop and had it going 24/7 next to me

[smoky oak]

I am racking up SD gens so fast, my SSD's are getting mad lol

[tawdry anvil]

could focus on the real work in SD 😄

[smoky oak]

haha

oh, would you mind linking to the noise offset video?

[tawdry anvil]

thats something im concerned about my drive I had SD on since the start now makes awful noises lol

[smoky oak]

I am really biased to night photography

[tawdry anvil]

https://youtu.be/cVxQmbf3q7Q?t=2

[smoky oak]

I am at 13,105 gens in SD

[tawdry anvil]

thats on how it works, but I grabbed the model off civtai

thats all ? lol

or is that like "finished polished pieces"

in the early days I was doing about 10k overnight, I go slower now cause I dont overnight gen lol

and am way more prompt specific before I was just batch rendering with text files cause who knew how to prompt properly back then lol

[smoky oak]

yeah, it looks like the noise offset problem is why my black demon daddies I am trying to generate are borked lol

super harsh contrast

@tawdry anvilOhhh wait, this is a checkpoint?

you can only use it with this checkpoint?

[tawdry anvil]

Yes and No! 😄

https://civitai.com/models/8765/theovercomer8s-contrast-fix-sd15sd21-768 heres a lora

not tested it

but the illuminati one is a chkpnt

but that should be good for an insert/merge

[smoky oak]

hmm... interesting

I hope this new balanced noise don't come at the cost of quality, cause this model isn't was thuroughly trained

otter in a tea cupppp

[smoky oak]

SHEESH

pretty wallpapers

[dusk stag]

He has my vote

[smoky oak]

WAIT

where did you get Ricadro?

is it a LoRA or a textual inversion?

I wanted to make a LoRA of him, but there weren't any high quality images

[dusk stag]

It's a LoRA that you can get right here: https://civitai.com/models/9217/ricardo-milos?modelId=9217&modelVersionId=10927&infinite=false&returnUrl=%2Fmodels%2F9217%2Fricardo-milos&active=true

[smoky oak]

NO WAY

Bro, hes so hot lmao

[dusk stag]

It produces images of carrying quality

[smoky oak]

Understandably haha

I wonder if I can use realistic vision and a ton of curation to generate high enough quality images to train a better LoRA on

this is amazing haha

[sinful mirage]

https://imgur.com/z8xDVTP any tips? 2nd video in a while, using auto/deforum/controlnet(openpose), euler a, 20 steps, 7 cfg, 0.65 denoise. I wanna make it sorta keep the same person, it's also not really too detailed but thats probably my bad prompt

[smoky oak]

@dusk stagok, the Ricardo Milos LoRA gets insane results with realistic vision. I will be able to use it to generate images to train an even better LoRA

@dusk stagSorry for the second ping

with some proper prompting and inpainting, I will be golden haha

I can generate images to train a new higher fidelity LoRA on

[dusk stag]

Wow it's beautiful. I'll have to try out this model for some bulge action

[dusk stag]

This sounds illegal

[smoky oak]

Luckily it is NOT haha

I can't believe how well this LoRA captures how damn hot Ricardo is lmao

[wispy nest]

I think MJ makes really good gens and looks like it knows alot more "words" , making prompts easier and allowing you to do complex gens easier, but yea...

MJ can't do realistic photos , it's good at just art and it doesn't have amount of tools SD has.

[smoky oak]

MJ is cool for a very specific use, but stable diffusion reigns supreme on the forefront of new developments and diverse applications

[mystic marlin]

"Bruh..." -Tifa to Midjourney lol

But nah, MJ is fine, but I am in the same boat of belief regarding SD outputting things that I personally prefer over MJ.

I paid for MJ ONCE, and never used it again after one month.

This was right before Stable Diffusion came out too. lol

I like having leagues more control over what I can generate, not to mention....open source, the tools, extensions, you name it.

[wispy nest]

Woman

[craggy rune]

I choose stable diffusion becoz it allows you to trains your own shit

🤣

[jagged crest]

I fine tuned a SD 2.1 model on myself to generate these stickers that I can use in Discord!

[wispy nest]

Star wars, avengers

[smoky oak]

@dusk stagI am asking the creator of the Ricardo LoRA if I can use their LoRA to generate training images for my own Ricardo LoRA of much higher fidelity

You can see from their promotional images just how much it picked up on the JPG artifacts and compression

[dense tapir]

jpg <90 percent makes me sad

[smoky oak]

Compared to what I have been able to get with tons of iterations and inpainting lol

bruh

Same LoRA< but I know how to clean it up

[dense tapir]

I swear most of these have to be 60% or even 40%

[smoky oak]

So I wanna use it to generate images to train a higher fidelity and more accurate LoRA for him

[dense tapir]

He should be fine with that. I mean I wouldn't care if that were mine.

[smoky oak]

Train it on non compressed information

I am working hard as hell to clean these images lmao

[dense tapir]

yeah, I try and use png for all if 16 bit.

[smoky oak]

I am having to employ a lotttt of hacks haha

[smoky oak]

I just wanna ask first. Though I am gonna be really sad if they say no 😅

[dense tapir]

In all of my shit I have CC 4.0

[smoky oak]

I can't believe how much I was able to clean it up

the images for Ricardo are such shit quality lmao

[dense tapir]

CC 4.0 I allow all but for commercial purposes so share alike, improve etc... just give credit.

[smoky oak]

these are the wuality images it had to be trained on

super compressed

[dense tapir]

damn, that is shit

[smoky oak]

and thats like pristine quality compared to teh others

[nocturne hollow]

I'm using the stabiliy API img2img. I would like to set one of my pictures as input (init_image) and edit that picture (prompt) so that the output will be that picture of me wearing a golden armor or a suit, or with another picture style. But the face of the person in the output image is totally different of mine.

I don't know what I'm doing wrong !

[smoky oak]

so yeah lol

[dense tapir]

no only compressed but blurry af

[smoky oak]

the fact that I was able to get the face I was out of it is insane

[dense tapir]

You would be amazed

[smoky oak]

Was able to hack things using some very tedious processes lol

[dense tapir]

I am having a bitch of a time with lora right now and frustrated

[smoky oak]

you can see there are still some artifacts, but man, imagine how much better it would do with images this high res

[nocturne hollow]

When I use API , it doesn't keep the face of the init image.

How do you do to keep the face ?

[dense tapir]

There is a technique that SD can clean it up

[smoky oak]

wonder if its similar to what I did

[dense tapir]

uses img2img

[smoky oak]

Mine used img2img, lora balancing, and selective inpainting like 15 times lol

I shan't give away my secrets at the moment :p

[dense tapir]

yeah, pretty much how it is done then train on what it created

My lora woes is I am trying to make a general purpose lora

[smoky oak]

Its different for what I do to fix faces like this, but it still uses img2img

[nocturne hollow]

What is lora please ?

[smoky oak]

Low range data set

it teaches specific concepts to bigger models

[dense tapir]

it would appear people do not know what TE/Unet are when looking at an image for training. By that I mean if my lora is screwed up how do I know which of those two learning rates was bad?

If you have the additional-networks extension
The webui extension allows modifying Unet and text encoder strength separately. This allows you to play with the values and try to see which component is over/underbaked. The changes usually reflect the adjustment necessary to training.
If, for example, your model works fine at 0.5 strength for Unet and 1.0 strength for TE, that means you just need to reduce the Unet learning rate by half.
If it works fine at 1.0 Unet strength and 2.0 TE strength, then leave Unet as it is, and double the TE learning rate.
It saves a lot of guesswork.

[smoky oak]

you train a lora in order to teach a concept to a bigger model

[glossy herald]

https://github.com/cloneofsimo/lora/

Fine-tune Stable diffusion models twice as fast than dreambooth method, by Low-rank Adaptation
Get insanely small end result (1MB ~ 6MB), easy to share and download.
Compatible with diffusers
Support for inpainting
Sometimes even better performance than full fine-tuning (but left as future work for extensive comparisons)
Merge checkpoints + Build recipes by merging LoRAs together
Pipeline to fine-tune CLIP + Unet + token to gain better results.
Out-of-the box multi-vector pivotal tuning inversion

[nocturne hollow]

Are you using the stability.ai API ?

[smoky oak]

I am honestly not sure lmao

[nocturne hollow]

Alright.
It means that the model itself is not able to keep faces ?

[smoky oak]

oh, I did use SD to fix the face using inpainting close up

its a high res gen ontop of a lower res image

[nocturne hollow]

Can you give some intuition please ? I would really want to keep the faces !

[nocturne hollow]

Can you give some intuition please ? I would really want to keep the faces with SD !

[smoky oak]

Like how I fixed her face?

I generated the original image using an artsy model, then I went to inpaint, then I painted over her face, and used a photo realism model with a low denoise. Boom

[glossy herald]

there are memes you shouldn't controlnet to reality though. like Derp

or should we

[nocturne hollow]

I would like to keep the face of an image, and modify its clothing or the style of drawing for example. When I use the API the face changes in the generated image.

I don't know how to fix this.

[smoky oak]

peeng

[smoky oak]

use impainting and make sure to deselect the face, then regenerate

[nocturne hollow]

Ok, I try it.

[nocturne hollow]

Does that mean I should have an inpainting mask for each type of generation I want?

[spice oar]

Man with noodles as eyebrows.

[nocturne hollow]

These are outputs.
Prompt : Make him wear a suit

The faces are totally different from the init image

[glossy herald]

it was the "b*tch please" rage face

[craggy rune]

the f!

[smoky oak]

thats fucken rancid lmao

[glossy herald]

you'll need to prompt completly differently, this is not how this works, you'll need to describe the image you want, not the changes you want to happen

[glossy herald]

yeah, I had some horendous ones... check #1019361238234443776

[nocturne hollow]

Then it will keep the face from the init image ?

[craggy rune]

its scary in some way

[glossy herald]

if you want the face NOT to change at all, you will need to use a mask. there is an "inpaint" tool that makes SD modified only parts of the picture

selfie of a sad man crying with a giant enormous chin, small mouth, fake smile, 70mm, taken on iphone 6, seen from under

with controlnet of couse

[craggy rune]

which control net model ? canny ?

[glossy herald]

scribble

from the meme face

[smoky oak]

skwibble

[smoky oak]

@glossy heraldPlease do something with the troll face lmao

[glossy herald]

not 100% lol

I tried

didn't get far, let me see the best I have

[smoky oak]

I am watching a video thesis on the breakdown of the all tomorrows evolutionary horror fiction, and these generations fit right in lol

[glossy herald]

derp

that's all I tried for now

well, I did some others but with "bad cosplay man"

[nocturne hollow]

Is there a website where I can find masks?
A kind of Lexica.art for masks 😅 .

[nocturne hollow]

Is there a website where I can find masks?
A kind of Lexica.art for masks 😅 .

[glossy herald]

usually, you just draw the mask in the UI

just click in the picture when you are in the inpaint tab, and draw over what you want changed

it will only modify inside the masked area

you can also open your image in photoshop first, use the tools to make the best mask possible there, as a black on white picture, and upload the mask you made to the UI if you want more precision

[nocturne hollow]

So what if I would like to make API calls from a client app to generate images ?
How can I get a specific mask for the image of each client ?

[glossy herald]

on what tool/ui/API?

on automatic, well, you would activate the API, and use the documentation of inpainting call, I don't really know more. You'll need to provide the mask as a secondary picture in such case I think

[green plover]

the inpaint api call has a mask parameter in the json bundle

you base64 encode a grayscale image with white being what you want to replace. same way you base64 encode the main image

check the request body example in http://localhost:7860/docs#/default/img2imgapi_sdapi_v1_img2img_post

[nocturne hollow]

I get it !

[nocturne hollow]

It's in your loclahost !!!

[green plover]

you have access to the docs through your automatic1111 interface. That link should work on your computer too if you have the --api switch set and are on the same computer you are running on 🙂

[glossy herald]

yep, the doc is only accessible localy, you need to have automatic running for it to be computed and available, it depends on the extensions

[green plover]

but if you are running then someone pasting the localhost link to it will work for you

[glossy herald]

yep ^^

Trying to get the "no" face, not easy

a little too far from what SD manages to get, I only have monstruosity for now

too far from what a real head could look like, I can't get it to understand where the mouth should be

[green plover]

bah!

[glossy herald]

I'll change meme, too horific

Y U NO

[glossy herald]

I couldn't share the worse

they were too NSFW

[green plover]

that's 50% of my work

[glossy herald]

like monstruosity, parts of body patched together

from that

[green plover]

Maybe reduce your denoising?

[glossy herald]

the scribble just completly drops the sticksman hands whatever I go

xD

feels more real the less denoising, but I don't recognise the meme anymore

wow I get the "Y U NO" feel on that one

let's change style, memes have done their time

original

[young mountain]

I added aptitude for ginger style in my model

[jagged crest]

You can try instruct pix2pix. That has a good chance of working. It's available on playground AI as "Edit"

[jagged crest]

Dude this is on point!

[glossy herald]

thanks 🙂

I just posted the compilation of the best I got

https://new.reddit.com/r/StableDiffusion/comments/11e4dfw/ragecomics_and_memes_as_real_as_i_could_get_them/

I'm trying to make this a reality right now

[jagged crest]

Check this out I used instruct pix2pix on playgroundai on one of the pics you uploaded. Prompt was literally just "make him wear a suit"

[jagged crest]

I am generating these stickers of people by fine tuning a model on their photos. But I'm having a really tough time generating specific expressions like in the rage comics. People want these stickers with expressions that they could use instead of emojis

[jagged crest]

I tried to use controlnet using my fine tuned model in Automatic1111 but for some reason it doesn't accept my fine tuned model's checkpoint.

Has anyone used a dreambooth fine tuned model checkpoint in Automatic1111? For me it just doesn't appear in the checkpoint selector. I'm running Automatic1111 from a Colab because I don't have a NVIDIA GPU computer

I've made a product out of it at https://www.youstick.fun. But I want to make it better by generating stickers of people with specific expressions. I feel like people will use it more in that case

[glossy herald]

hey
I mostly use dreamboothed models, and it has always been working great.
If the model is either a .ckpt or a .safetensors and is placed in the models directory, you should have it in the dropdown for selecting models, top left. if not, try to click the blue arrow next to it to reload the list maybe
Worse case scenario a broken model would still show here, and would show an error in the console when you try to use it. Not showing in the menu at all is either you didn't put the model in the model folder, or you didn't reload, or I have really no idea

really cool model and project !

I got to say those, those rage faces are too "extreme", even for controlnet, to keep it photorealistic and consistent. All of those I shared had a very low rate of good pictures, I mostly got garbage, with controlnet/SD not being able to understand the drawing correctly

[jagged crest]

Thanks!
So while the current stickers are coming out great I want to do even better. Some things I'm trying to improve -

1. If I am doing this for a blonde person and I use img2img and my init image person has black hair, the generated sticker person sometimes has black hair too.
2. Specific facial expressions in my stickers. With img2img it is a bit of hit and miss. Sometimes works sometimes not. Generating sad women in particular is proving to be quite difficult. I wanna try controlnet to improve this. Keep the person's hair color, major facial features and the art style constant but just change the expression.

[jagged crest]

Do you think training a base model in this specific sticker art style and then dreambooth fine tuning on a person's photos would be a better direction @glossy herald ? Would you recommend Lora, Textual inversion or Dreambooth for such a base model?

[glossy herald]

it would depend how much time you are ready to invest on each person you are doing a set of emote for.
1/ If you want to go full top quality, it will take longer, but I would train a small dreambooth on their face, a basic one. You could even use your already trained model as base. Then I would :
1.0/ prepare a list of cool emotes you want, using your own trained model in your style. I would keep those as template for every future client. I'll call this the "emote set"
1.1/ do a first pass in inpaint, using a mask that covers only the face part itself, in particular eyes, nose and ears, and using their dedicated trained model of their face on the emote set, so it can adapt the features of the face to make them more regonisable. I would also use controlnet on canny with denoising like 0.9, we don't want a lot of change
1.2/ do a second pass in txt2img, no mask, controlnet with a very high denoising this time so that no detail moves, using the model trained on their face, or your model, I'm not sure here. The goal of this step is to get the hairs, colors, ... fitting to both your emotestyle and the client's features.

2/ if you want to go a little lower quality, skip 1.1, go to 1.2 directly and use your own trained model. No dedicated model training for their face in this version.

In both case, the 1.0 is very important imo. You want to have already prepared emotes that would just be lightly modified on the main features, and colored differently each time, fitting to either the client or your model. Especially if you want stylized faces, it will let you get a very high quality on that first step, at the cost of some time. And then transfer that quality to each client in a low time.

as for what to use, LORA or DB or other, I would go DB if you are running locally, but I'm a LORA noob so I'm not in a good position to really say

[glossy herald]

one this I realised I didn't say but imply : the colors of the hairs, even if you don't train a dedicated model, shouldn't be a problem. using txt2img with controlnet, you can get the right color thanks to your prompt

[jagged crest]

Wow thanks so much this is very helpful! Just to clarify in 1.1 you mean - in painting with the mask created on top of the emote set image's face and the model as the subject's fine tune. Right?

[cyan snow]

does the VAE look blurry?