🏞｜general-with-images

[rocky quail]

you mena 600 steps

[zinc patrol]

600 images 90 steps each

if the image is too dangerous

[rocky quail]

go to civitai, click what image you like then look at their promprs

[zinc patrol]

try adding "Roblox charcter,roblox graphics,blocky charcter"

If you wana get the exact same image use control net

[rocky quail]

you forgot (sexy:2)

[zinc patrol]

They can help you get consitent charcters

[zinc patrol]

🤫

@wispy nest

https://civitai.com/models/88420/roblox-shirt-sd-2

https://civitai.com/models/119969/roblox-face

Try adding loras

ur settings are wrong

use 50 steps

1024x1024 res

[rocky quail]

they are like adding style to a picture

[zinc patrol]

lora is model on top of a model it tells the model things

[rocky quail]

but it have waymore usage than that

yes, but it depend on if the weight of it is stronger than your other prompts

[zinc patrol]

my favourtie lora https://civitai.com/models/38527/aizenbleach

Try using my workflow 🫡

@wispy nest https://github.com/comfyanonymous/ComfyUI_examples

not mandatory

[rocky quail]

you just need really specific prompts

[zinc patrol]

you can get stunning image with just 6 words

[rocky quail]

mostly making them in negative prompts

[zinc patrol]

@wispy nest

Made by my friend

[rocky quail]

yes

[zinc patrol]

yes but its on horde

By me

by me

in just a couple or words

ye Horde is insane

[rocky quail]

yes

[zinc patrol]

lets just say you need to read this

https://stablehorde.net/

[rocky quail]

stock SD 1.5 is just so bad

[zinc patrol]

maybe

[rocky quail]

meanwhile other 1.5 models look way better

[zinc patrol]

but they are for spefic cases

sd 1.5 was base model

sdxl is the goat

see these examples

Very good images with a couple of words

thats because you are using 512x512

[rocky quail]

my SDXL run locally made this in 30min

[zinc patrol]

use 1024x1024

[zinc patrol]

the renders above were made under a minute

send a screenshot

[zinc patrol]

mind giving me the prompt?

[rocky quail]

yeah, lowvram setting is very slow but medvram said it need 20mib more vram, i wish there was a inbetween setting

[rocky quail]

(fashion upper body photo of a female model sitting down facing camera with full head in frame:1.2),

(beautiful, Russian, young),

(detailed white shirt with long sleeves),

(blond hair, bob cut hair, angled bang), (realistic detailed skin, realistic detailed hair),

(big blue eyes, looking at camera, smiles, white teeth, smiling eyes),

(small shoulders, wide hips, small waist:1.2),

(detailed beach, sea, sky, sun light, sunrise:1.2),

(arms back, anatomically correct arms),

(8k, HDR, best quality, high contrast, vibrant colors, film grain, professional photo, sharp details, sharp focus, in focus, hyper realistic:1.3),

[zinc patrol]

a1111 issue

As I said a1111 is not good

@rocky quail

Under a minute

rate it out of 10

[rocky quail]

10/10

[heady anchor]

mid / 10

[zinc patrol]

Lamfo Horde is steriods

[rocky quail]

i could post the negative prompts but somehow it is cesorec here

[zinc patrol]

Horde images on the discord bot are censored

but https://tinybots.net/artbot aint

[rocky quail]

my prompt does not include boobas to make those horny AI model not crank out naked woman

i downloaded 100gb worth of models and use the same prompt to compare them

[zinc patrol]

dont worry I still have 12tb remeber?

[zinc patrol]

horde offers 5000+ models for free

[zinc patrol]

the jawline so cursed

Horde da 🐐

Image?

use comfy

[rocky quail]

example, if you do not want a blurry picture, put (blurry:1.2) in the negative

[zinc patrol]

https://colab.research.google.com/github/camenduru/sdxl-colab/blob/main/sdxl_v1.0_comfyui_colab.ipynb

not quite what i expected

yes

[rocky quail]

i used gigachad lora

[zinc patrol]

lamfo thats why it was so unatural

[rocky quail]

those people lora doesnt look 100% alike but atleast regconizable

Trumpn't

some pic on civitai is just corrupted or something, doesnt have prompts, best to look at others

[zinc patrol]

@wispy nest

these are the activation words

No sherlock

lora have activation trigger words

so that they dont activate unexpectedly

good luck in that unreal blueprint looking goofy ah ui

[heady anchor]

your prompts sucks so the result will be the same

[zinc patrol]

People who use a1111:🤡
People who use comfy:🫡
People who build their own Webui:

Rate it

[heady anchor]

generic

[rocky quail]

true, its not a1111's fault, its skill issue

[heady anchor]

https://tenor.com/view/ez-noob-skill-issue-simply-a-difference-in-skill-gif-20780203

[rocky quail]

too complicated

[zinc patrol]

lamo using horde I get badass Images with 3 words

[rocky quail]

try simplifing your words

[zinc patrol]

https://tenor.com/view/smg4-smg4skill-issue-skill-issue-smg4mario-smg4skill-gif-26104493

[rocky quail]

should be: 1girl, diamond shaped eyes, night city street, falling dollars, <artist name here>

[zinc patrol]

@wispy nest

lamfo

[rocky quail]

like i said, its personal preferences and skill issue. just because it is more complicated does not mean it is different

[zinc patrol]

A divison horde by db0

https://stablehorde.net/

[zinc patrol]

Horde is on steroids

That's why it can generate badass images

[heady anchor]

where are the badass images

[zinc patrol]

no

[zinc patrol]

not ur pfp

"Haters hate cause they cant achieve"

[heady anchor]

you can come to #✨｜sdxl chat and check all the images we post there

idk just scroll up or use the search bar i dont save them

[zinc patrol]

Lamo my artbot collection is the most unholy thing in this universe

[heady anchor]

looks like the novelai outputs from 2022

[rocky quail]

atleast 2hrs

[zinc patrol]

It took me about 3 minutes to make these 6 images

I mean manga writers are using controlnet and sdxl to make charcters

As I said I use HORDE

[rocky quail]

with digital painting, you can do a full colored character with basic shading in 2hrs,

yes,it is

[zinc patrol]

Looking fine

ye my server profile is different

[rocky quail]

can you view your google colab gpu model?

[zinc patrol]

Go upto runtime type

but you cant change since you are not paying

[rocky quail]

preety cool for a free 16gb vram machine

[zinc patrol]

google colab is pretty dope

but horde is doper

since it uses a 80gb vram to generate images

[rocky quail]

left is sdxl on colab, right is realistic vision 5.1 on my pc, same prompts, random seed

[zinc patrol]

a1111 or comfy?

[rocky quail]

a1111

[zinc patrol]

dang

also check out horde

[rocky quail]

give link, ill try it

[heady anchor]

can u try to create a bugatti on sdxl?

[rocky quail]

sure

[rocky quail]

ah yes, tall asian woman

[zinc patrol]

comfy or horde?

[rocky quail]

sdxl anatomy is still kind of wack

[zinc patrol]

sdxl shit

you have to make it

using nodes

[rocky quail]

i think its best to run locally if you have the money

[zinc patrol]

horde(^///^)

[zinc patrol]

I have a 3080 but still perfer horde

[heady anchor]

tall like her

[zinc patrol]

try asking https://www.perplexity.ai/

[zinc patrol]

which model?

[rocky quail]

she needs to go to the eye doctor lol

[heady anchor]

put some dirt on her eye

[zinc patrol]

no a neck doctor

[heady anchor]

1.5

[zinc patrol]

havent seen someone use 1.5 in a long time

[rocky quail]

also, her shirt turns to hair

[zinc patrol]

sd 1.5 hallucinates too much

it requires a shit ton of prompting

[heady anchor]

this one got better but less asian

[zinc patrol]

mind sharing prompt?

[heady anchor]

1girl,blonde,(hi-top fade:1.4),dark theme,soothing tones,muted colors,high contrast,(natural skin texture, hyperrealism, soft light, sharp),waist up,fashion design,black sport tank top,jogger,(red background:1.1),simple background neg (worst quality, low quality, lowres:1.4)

[rocky quail]

what model

[heady anchor]

absolute reality

[zinc patrol]

I cant believe its photography is also good

Ah yes

I really like the black tank top in the second image☠️

[heady anchor]

that hair looks rly good 😭

[zinc patrol]

Sdxl supremacy

[rocky quail]

you guys notice how when you put (bald) in the prompt, it give the girl a guy's scalp?

[zinc patrol]

where?

[rocky quail]

in realistic vision

im sure i got it once before

[zinc patrol]

lamfo

I have seen stability.ai took a page out of midjourney's book

and made sdxl less realistic and more dependent on cgi

They really want to beat midjourney

[rocky quail]

yeah, SD biggest quality is the realistic stuff

[zinc patrol]

no sdxl is more dependant on cgi

like

this aint realistic

[rocky quail]

it looks like digital painting

[zinc patrol]

good now you have a proper prompter

but I am happy with image horde generates

They do the trick

ngl lacks background

[rocky quail]

i like asian

[zinc patrol]

Never seen blonde asians

as an asian

neither they are tall

[rocky quail]

original prompt is a russian girl, i just changed the russian part to asian

[zinc patrol]

Horde really be crushing a1111 and comfy

not a1111

basically everyone is complaning about it being a piece of shit

after sdxl intergration

horde has /source image and comfy can do img to img

[rocky quail]

it is hard to setup and sometime it doesnt blend it

[zinc patrol]

https://comfyanonymous.github.io/ComfyUI_examples/img2img/

[zinc patrol]

a1111 tried to become simple but they just degraded

I mean gradio is shit

[rocky quail]

i kind of agree on gradio thing

[zinc patrol]

I mean gradio on colab takes forever to load

ye

so?

[rocky quail]

yes

[zinc patrol]

img2img doesnt work that well

inpainting is good tho

[rocky quail]

also 3d openpose

basically replace a part of the image with a new thing

how much is your denoise strength

it should be 0.2-0.3

at 1 it is a completely new pic

[zinc patrol]

and sdxl's defination of "new pic" can be anything even random pixels

[rocky quail]

looks like a man's scalp

[zinc patrol]

get that shit out of my screen!

[rocky quail]

resolution doesnt matter with inpainting

[zinc patrol]

try adjusting the res

or his res fix

[rocky quail]

it seems SDXL can't do bald woman

[zinc patrol]

sdxl's humor is broken

[zinc patrol]

you need to get your balls deep in comfy webui nodes

lamo

read their docs

even if you hate reading

or use Cough horde

[zinc patrol]

The language of the documentation might be hard to grasp

Make sure the model is simple

Did you see the markdown options ?

@wispy nest

Basically a seed is a randomier

If you take 2 computers run sdxl on them

And take the same prompt

And put the same seed you will get the same images almost same

The AI model

Ye

Same image if the model is same

It's like minecraft

Every seed is different

If same seed on 2 pcs the world is same

I don't recommend using paid/closed source software

They offer you convince but slowly they charge more

But if you can pay very well

[tiny cloud]

How can I do this? There is no CLIP Text Encode node that accepts STRING?

As it happens String is not accepted by any node at all.

[kind quartz]

open @tiny cloud clip text encode by clicking on circle on left

oh from yesterday. Probably solved already

But simply you right click on clip text encode when not minimized and from menu you choose convert TEXT to Input

[sly mountain]

Does anyone have tips for outpainting with Automatic1111? Im running SD 2.1 and trying to outpaint this from 1024x1024 to 1920x1080 but it keeps returning a square with some blurry outpainting. I'm using Poor Man's Outpainting in img2img and have tried adjusting all the variables without success.

[crimson holly]

I was using image editing tools (krita or gimp) in that case to extend canvas + draw roughly background, than inpaint img2img.

[prisma stone]

How to stop it from generate 2 people?

My prompts tell them EVERYTHING now

I have nothing to do to make it better

[sterile temple]

solo, centered?

[prisma stone]

It's impossible

I'm trying to generate Frisk from Undertale

but A.I. won't stop being stupid

[sterile temple]

start with 'solo, 1boy, adult, male focus' take all the negative prompts that refer to human, woman etc.

[prisma stone]

What the fuck.

This is cool

[sterile temple]

which model are you using?

[prisma stone]

https://civitai.com/models/35960/flat-2d-animerge

[pulsar pelican]

I have a question Im upscaling images with esrgan 4x+anime from 512x512 to 4096x4096 and I got 2 results one is 500kb and one is 5.5mb and resolution is the same and even quality of those pictures... can anybody explain how or why the 500kb has same quality as the larger one?

[sterile temple]

the esrgan models don't really add more details

[pulsar pelican]

both are esrgan

what I dont understand why one is small size and one big with same resolution same details

[sterile temple]

how did you upscale it. because the upscale models usually output 4x resolution

[sterile temple]

you had resize set to 8 for both?

[pulsar pelican]

Yes

[kind quartz]

i am feeling stupid, but just realized it is posted also in different places and solved 🙂

[glass lantern]

¯_(ツ)_/¯

#🏞｜general-with-images sex au

[zinc patrol]

bro why wont Inpaninting give me nsfw?

[prisma stone]

What prompt allow to generate a human with butterfly wing?

My model is fucking stupid

[nimble geyser]

put "stupid" in the negative, then it'll be smart

[vocal cloud]

Why not just use Ultimate Upscale or Multidiffusion? The base upscaler in the UI doesn't add any noise, if that's what you looking for.

[vocal cloud]

You're generating the images at 768x768 using a 512-trained model, which is why you're getting those wack results. Use an aspect ratio based off 512.

I've used that model before and can confirm, however, that it can be messy at higher base resolutions.

[pallid ruin]

Try using outpainting mk2 , it gives me better results than poor´s man outpainting, also I use a denoise strenght of 8.5 here, the result isn´t that good because I used a simple prompt and i downscale the image to 256 to make it faster

[opal warren]

Have yall seen SplitticAI lately?

"A portrait photo of a kangaroo wearing an orange hoodie and blue sunglasses standing on the grass in front of the Sydney Opera House holding a sign on the chest that says Welcome Friends!"

I got the kangaroo image on my third try using it (not stable diffusion)
i know the hoodie is not orange, but still

[opal warren]

I have seen other models be able to do this but those were cherry picked. this was just third attempt

A green cap with text saying "cap" on it to the left of a blue cap with text saying "cap" on it
5th attempt

its very fast too. but its limited to 25 images a day and there is no paid plan :(

[prisma stone]

Aw man :(

I'll try, thz

[split rover]

you sure its not a version of XL? got this 2nd image

[split rover]

2nd as well with XL, just using defaults on clipdrop since im not on local machine atm

[little basalt]

hi all, im new to this stable diffusin, and im running on my laptop.. but i have a problem , i cant generate second time , its like stuck like this after 2nd run. can you guy help me? thx in advance

i restart again the stable diffsion and run for 1st and it worked, but again the 2nd time ...it wont run 😦

[heady anchor]

downgrade to an older version of webui

[little basalt]

can you give me link or the step sir?

[heady anchor]

you have open cmd where u have your webui installed then type this git checkout master git pull git reset --hard "hash goes here" pause

u can find the hash to w/e version u wanna downgrade to in the webui github page

[cyan snow]

When I run SDXL it takes about 14-ish seconds per image (4k), and it gets full sentences right 80% or the time.. I feel like the version on Clipdrop isn't as good as the finetuned ones

[tame grove]

@strange jungle what should I do here?

@strange jungle

I shouldn't need to tick the lowvram box, since I have 8GB, right?

[strange jungle]

right

maybe for sdnext, but not for 1.5

[tame grove]

actually hang on, I think I already went through this at one point, idk if itll dupe the folder

[strange jungle]

it will put it wherever you run git clone

look to see if you already have it, but it looks a lot like SDNext

[tame grove]

so I found the folder it made when I initially installed it on the 14th

would I run launch.py, webui.py,?

[strange jungle]

there should be a bat in there

[tame grove]

webui.bat or webui-user.bat?

[strange jungle]

but you should first do CMD in that folder. Git Pull to update

[tame grove]

ah. right

[strange jungle]

webui-user.bat

[strange jungle]

actually, you had update on launch checked

it auto pulls the newest version when you launch

[tame grove]

looks like commits were done?

[strange jungle]

right

it updated

[tame grove]

I don't have a lot of git knowledge but I at least have some from my CS class years ago lol

[strange jungle]

sometimes updates break things

[tame grove]

from here, do i do webui-user.bat?

[strange jungle]

there's a daily update to some of the dependencies, which can also break things

[tame grove]

like typing that in cmd?

[strange jungle]

like pytorch

[tame grove]

right

[strange jungle]

double click on the file

you can make a shortcut to it

[tame grove]

I have 3 separate UIs with the same image. paaaiinnnn

Would it be best for me to just uninstall VLAD and Easy Diffusion?

[strange jungle]

I have comfy, a1111, two sdnext, and then the same on another machine

If you think you won't use them

[tame grove]

Or would a file folder be sufficient

[strange jungle]

the models take up the most space

delete the models from the ones you don't use. If you need to use them again, copy the folder

[tame grove]

Alright

May as well stick these all into a FIle folder just labeled "Stable Diffusion UIs"

[strange jungle]

youll likely have to update them next time you want to use them

[tame grove]

right

The folder just has the shortcuts for each UI's version of the webui-user.bat file

[strange jungle]

I have to go. I hope this works for you

[tame grove]

I'll mention you if anything weird happens ig

[strange jungle]

ok

[tame grove]

for some reason I'm not seeing sd_lora popup in the top like I did when changing quicksettings, even after a UI reload

it says add network to prompt instead of add LoRA to prompt

nvm, had to just refrresh

[vocal tide]

Will it solve the problem? Can you explain sir what it happended?

[prisma stone]

How to make wings on her back?

[chilly reef]

why it is showing so much time

is i need to change any of these settings

[opal warren]

cap was spelled perfectly on all of the 10 images i generated. Its definitely an sdxl competitor

[opal warren]

if you can get full sentences right 80% of the time then you are cheating with controlnet

[cyan snow]

well, I'm not. hold up; I'll make stuff with text

[opal warren]

prompt?

[cyan snow]

[sign saying: "WELCOME TO NOWHERE" on it in the middle of the forest]

[opal warren]

idk what you are using but this is what i get with sdxl on playgroundai

12 prompt guidance 50 steps 1024*1024

splittic:

[cyan snow]

I'm running it locally

[opal warren]

lemme try comfyui with the workflow i stole from someone

hopefully my computer survives

I am now trying normal sdxl base and my computer is dying

Its just loading the model

base did an ok job

[late cosmos]

Can anyone reproduce this type of image? It was also generated by an AI

[kind quartz]

It is from civitai?

[late cosmos]

Nop

[kind quartz]

ok then

[boreal falcon]

eventually, didnt try yet

[chilly reef]

Is there any site where I can free prompts for SD deforum

[chilly reef]

Hey

[vocal cloud]

It's called your imagination.

tHeRe ArE nO LimIts, oNly pLAtEaUS

[sly mountain]

Thanks!

[strange jungle]

">' not supported between instances of 'NoneType' and 'int'" What does that mean? If I type a prompt, that's the error. If I image browse, send the image to txt2img, delete the prompt for a new one, it generates

[urban forum]

Hey everyone, I've just introduced some new features to my Deforum Helper tool – an editor and visualization tool designed for Stable Defusion Deforum animations. It's a solution for effortlessly controlling animations within Deforum. These additions include handy frame markers, audio channels, supplementary settings channels, enhanced user interface options, and much more. Give it a try and kindly share your feedback on how I can further enhance your experience. Let's enjoy the process and come together to create some amazing AI art!

https://deforum-helper.vercel.app/

[chilly reef]

Hey

Anyone I have a query

Can we use our own image at the end of the deforum video

[prisma stone]

It won't generate a backward cap :(

Impossible

[obtuse palm]

#✨｜sdxl message

forwarding this post here..

Using denoising strength more than 0 breking the skin.. is there a fix?

[heady anchor]

change upscaler to latent (nearest)

[zealous socket]

Hey guys does anyone know how can I make the tiling effect less visible when using Ultimate SD Upscale + CN tile? I’ve tried messing around with denoise and padding size but can’t seem to get it right

[native tapir]

I've had a lot of success with 0.5 denoise for the seams with a tile size of 256.

Although I have run into issues where no matter what I do the image will have visible seams.

[zealous socket]

Hmm I’ll give that a shot, do you use seams fix as well? Yeah, some images upscale beautifully while others have these ugly tiles lol, it’s so finicky

[native tapir]

Yeah my bad, that's what I meant. Regular denoise for the image will vary, but the denoise for seams fix has only worked well at 0.5 or above for me.

[zealous socket]

Gotcha, is band pass the best seams fix or is there no “one size fits all”?

[inland summit]

Why when I try to generate and image it say hang tight, you're still dreaming! 🙂 That's for More than 30 minutes ?

[native tapir]

I haven't experimented much with the different options, I just leave it on bandpass

[vague haven]

Hi guys, I'm trying to create an image with the prompt "A dark outline of a woman putting on her business suit jacket is seen in a dark room. In front of her there is a square opening that shows a brightly lit stage. The stage has a microphone in the center. In front of the stage there is a crowd. The focus is on the opening that shows the stage. Photorealistic." and negative "cropped, lowres, poorly drawn face, out of frame, poorly drawn hands, blurry, bad art, blurred, text, watermark, disfigured, deformed, closed eyes"

But the results don't fit at all what I had in mind

(this is with SDXL in Automatic)

Can anyone suggest a different prompt that would fit better my idea?

[zealous socket]

@vague haven that prompt is way too verbose, your best bet is using controlnet canny or lineart to achieve what you want

[vague haven]

Oh, I had understood that with stable diffusion xl the goal was to have long very specific prompts

Was that incorrect?

(or I misunderstood?)

[zealous socket]

Sorry! I’m still on 1.5 so that very well could be the case, my bad

[vague haven]

No worries, may I ask why you're staying on 1.5? From what I've been reading it appears that XL is massively better

[zealous socket]

Mostly because I’m on an RX 580 8gb lol, but even so it’d be tough to use sdxl without ControlNet and all the various loras I use

[pastel moth]

It took a awhile to transition to XL, 1.5 just has so many goodies out there for use.

[round shell]

With SDXL less is often more because it listens way better in general 🙂

try whittling your long prompt down to its most essential aspects (without the negative prompt if you have one), and then start adding things back in slowly so you can see how the tokens you're choosing impact the output

[wicked needle]

sdxl surprises me more every day, he's on his way to being an all in one

[strange jungle]

That civitai image contest...I have a hard time finding the images I submitted. How are others supposed to find it to react?

also, the reactions in the contest collection are separate from the reactions to the image on its own...it's very confusing

[vocal cloud]

That is not surprising and a shame. Civit is a damn mess.

[strange jungle]

yes. Plus, the votes really dont matter because then it will be judged by some panel

[fast talon]

i was trying to turn a crazy drawing into something even crazier

[prisma stone]

It's impossible to generate a cap that have blue brim

Can someone help me plz?

[sterile temple]

did you try red and blue hat? and just run a few and hopefully you get one with the right colours?

[prisma stone]

I'll try, thx

[prisma stone]

a big steam boat in the sea, starry sky

[quiet current]

Aaaah yes wind powered steam boat

Remind me of the chinese wheeled train.

(Obviouly, use negative prompt to remove that sail)

[dense void]

Using A1111 my images have started appearing in the preview ok but then final image dieappears and in the output nfolder they appear like this... anyone help?

[nocturne oak]

Maybe look at what model you're using and what vae you've selected.

[kind quartz]

hope it is o.k.

[sterile temple]

she does look a little bit worried 😄

[kind quartz]

Lot of spy work to do 🙂

[opal warren]

yo guys

I challenge yall to remake this with SDXL or other

a cat with blue eyes with a sign saying "I have blue eyes" in its mouth. I got this with SplitticAI

its amazing with text and following prompt (most of the time)

[stoic pivot]

My new background, it was given to me by SDXL-base in Automatic1111 with:
Prompt "close portrait image of a small anthropomorphic cartoon teddy walking in the forest rain autumn forest BREAK seen from side"
Negative "mid placed design, center placed, right placed, borders"

[opal warren]

I love how you can steal other peoples nodes :>

lol

[cyan snow]

I never said you can't have it.. https://civitai.com/models/133818/tdg8uu-aitworkflow

I'm taking a break for a while now anyways; the new MJ version that is about to release is definitely better than SDXL, so yeah- I'll be gone for now

[opal warren]

new MJ version 👀

[cyan snow]

yeah, fuck that. I don't care; they are better now. perfect excuse for me to go silent for a while

[opal warren]

lol

[wraith karma]

Playing Oldest game with C.ai bot

https://youtu.be/iZOwaeAEIw0?si=Zu2gL8oWEhBvsphl

[prisma stone]

What

the

Fuck

What the fuck did I do wrong?

[heady anchor]

https://tenor.com/view/ez-noob-skill-issue-simply-a-difference-in-skill-gif-20780203

[coral mulch]

guess the celebrity game but images are generated by ai

[loud narwhal]

i was a stable diffusion beta tester last year and i miss the distinct grainy images that that text-to-image model could generate. does anyone know if any models like that exist today?

[rain rover]

AI or nah

[heady anchor]

try with sd version 1.1 , 1.2 ,1.3 or 1.4

[loud narwhal]

ooh where can we try those?

[heady anchor]

https://huggingface.co/CompVis here

[loud narwhal]

!! O: thank you!!!

[dusty bough]

i tricked my friend with a fake fortnite event i generated

[stoic pivot]

Before I got a smaller stroke I was often out to take photos, AI helped me get some creativity back, now I will print this, stick it on my door as a goal for this fall.

[prisma stone]

If you're really skilled, why won't you do it by yourself instead of calling me Skill Issue huh?

Mr.douchebag

[wispy nest]

That looks really good

Was it hard to achieve the 3d game look

[warped geyser]

Hello everyone. I came here to ask if anyone know what one particular model some pixiv authors are using. I like it's style very much, looks very cute, but I can't find anything about what model it is. The only thing I know is that there's not a single person on pixiv who uploads similar arts, there's a plenty of them. Here's an example image. The most distinctive part for me is eyes, the pupils with very soft outline. Maybe someone here knows what model this is or how to achieve same results. It's most probable a chinese model, but even my English is limited, so it's very difficult for me to search it on my own.

[wispy nest]

.> I think I know what model this is

Gimme 5 minutes

Saw one exactly like that today

[heady anchor]

https://civitai.com/models/47893/kawaiimix-niji-v5-cute

[warped geyser]

I found some clues on a chinese website called tusi. There's a GuiltyFlame (罪火 in chinese) model, which looks very similar, but the author said that his model is a poor imitation of "88 key's 草三". So there's one original model, which can be read as Caosan, cao san, or Grass Tree, but I've also seen 草3 variant. The problem is that none of that is searchable. It's like they refer to a model that never existed.

[warped geyser]

Thank you for helping, but it doesn't look like what I'm searching. The style looks different to me.

There's actually 2 versions of GuiltyFlame on Tusi, uploaded by 2 different persons, and the other one says something like "praise caosan" in his description. So it might be kind of very popular or iconic model I guess... Why can't I find it then?..

[nocturne oak]

https://giphy.com/gifs/spongebob-spongebob-squarepants-season-8-xUPJPhanYiB7z9xlCg

[wispy nest]

My 5 minutes takes into account everyone's 5 minutes

[nocturne oak]

https://giphy.com/gifs/math-lady-meme-WRQBXSCnEFJIuxktnw

[wispy nest]

Can you give me this picture link

[warped geyser]

Sure, here you go.
https://www.pixiv.net/en/artworks/108452042
There's more from this author, I think he uses the same model every time, but there's probably different loras involved. I've tried to ask him directly - no response.

[wispy nest]

@warped geyser check dm

[dusty bough]

i just included fortnite in my prompt XD

[wispy nest]

Which model was it ?

[dusty bough]

xl

[wispy nest]

xl is too powerful

[stiff jetty]

哈哈

[nocturne oak]

https://giphy.com/gifs/starwars-star-wars-episode-3-3o84sq21TxDH6PyYms

[dusty bough]

i trained a lura with my friend in it now im making him getting arrested

im trying to get him holding a coca cola bottle of its making his among us hoodie do it

[ripe wraith]

hello

guys do you know any game made heavily with AI by one developper or a small team, that is not a visual novel ?

like a small simple game

[stoic pivot]

Thank you SDXL for giving me yet another thing to have nightmares about.

[pastel moth]

https://tenor.com/view/robert-redford-jeremiah-johnson-nodding-yes-nod-of-approval-gif-21066931

[dawn seal]

I have no idea what happened, but suddenly SD is just giving me shit for images:
Any idea what's happening?

[wispy nest]

Why you doing 200x1200 images D:

Have they worked before?

And with that model?

[fleet goblet]

It seems Ideogram.ai just popped out of nowhere.

Another text to image generator.

And it's also quite good.

[dawn seal]

I'm trying to e a banner for a website.

[fleet goblet]

Like, insanely good.

[fleet goblet]

AI image gens are exploding, man.

[wispy nest]

And that model is for xl
Can you use xl on a1111

[fleet goblet]

SDXL only just recently released and now others are following up

[dawn seal]

I've been using it with no problems. I can try another model.

[wispy nest]

Also models aren't usually trained for 200 resolution

[fleet goblet]

Then just generate an image at 1024x1024 reslution

then downsample to 256x256

🤷‍♂️

[wispy nest]

Try something bigger height yeah

[fleet goblet]

This is not SDXL btw

I'm looking at a website mattvidpro just posted about.

[dawn seal]

I was doing 512x512 yesterday, and getting similar results. I thought it was just that SD didn't know what a chair looked like, top down. I've rebooted my system, let me try a normal sized image.

[fleet goblet]

If you are using SDXL at 512x512

you aren't supposed to.

SDXL was designed for 1024x1024.

[fleet goblet]

This model's no joke.

Damn.

Luigi from super Mario brothers if he was actually a real plumber in Brooklyn. Analog Photo.

[fleet goblet]

Lol

[wispy nest]

x3

[wispy nest]

Cat judges you unworthy

[fleet goblet]

I have been sentenced

[dawn seal]

Ok, so where can I find out what models are used for which resolutions?

[fleet goblet]

Everything before SDXL is 512.

SDXL is 1024.

Lmao

[dawn seal]

Ok. I'm completely new to this. Trying to figure it out

[wispy nest]

https://tenor.com/view/squish-that-cat-squish-cat-pet-gif-16708169

You'll have to squish the image

For banner purposes

[dawn seal]

Yeah, or I can create elements on a single color backgrounds, and then composite them in Photoshop.

Ok, 1024x1024 is giving something more expected. I had no idea that resolution would screw the image up so much.

[zealous socket]

My gf and I made using 3DAnimationDiffusion + controlnet + img2img

[heady anchor]

https://tenor.com/view/quagmire-real-meme-gif-26389588

[stoic pivot]

I installed ComfyUI and it is almost a bit to advance for me, I saw text_g and text_l and had to test, but I cannot grasp it, this image is "range cat sitting in the shade in the garden a summed day in Sweden" in text_g; and "oil painting, Carl Larsson" in text_l.

Do some here understand what G and L is and how to best use them, or is is always best to use same text in both of them?

[stoic pivot]

ComfyUI was like magic where I can throw random spells and things happen even I have no idea what I am doing, but I liked this image.

[nimble geyser]

https://tenor.com/view/no-one-knows-what-theyre-doing-filmjoy-movies-with-mikey-gif-22973486

[marble wasp]

Made with Easy Diffusion 3.0.3 Beta, epiCRealism modeler, 2 loras for lighting, CodeFormer for face correction, and Gimp for editing and touch ups

[prisma vessel]

Excuse me, how a prompt for generate image? Thank you.

[sterile temple]

#1100170312106127410

[prisma vessel]

#1100170312106127410 red hair girl

[split valve]

when you take the example jellyfish from clipdrop but just add HR giger

[delicate tiger]

best one I could make (CPU 720x720)
took 13 min. 26.5 sec.

it was pain

[shut sinew]

2.0/2.1 were 768

[wispy nest]

A very hard working girl!

[delicate tiger]

i need optimizations

[wispy nest]

Do you have all those --sub --lowvram

Arguments?

I don't even know if they work on cpu

[delicate tiger]

i have 16GB ram

ryzen 7 5800HS (8 cores/16 threads) 16GB ram

[wispy nest]

Have you looked into Google Collab

Without a GPU sounds too much of a pain

[delicate tiger]

is collab any better than this?

[wispy nest]

Collab you use goggles PCs to gen for you

Like discord bots

You have limited use but it's free

It's like 20 hours a week

I'm not sure

[delicate tiger]

i'll look into it

[wispy spindle]

https://discord.gg/czzw5JZn?event=1145771898421121125

starting soon!

[nova harbor]

I think we got it, please stop. I wasn’t interested the first 5 times, I won’t be interested for the next 5.

[long blade]

Was gonna post the same screenshot

[nova harbor]

I mean you can now LMFAO

only like 9-10 notifications I think not everyone got it can we have like 17 more

[long blade]

https://tenor.com/view/discord-torpedo-stop-spamming-ren-gif-22208739

[fleet goblet]

I kind've ignored 2.1.

No offense to SD.

[radiant rune]

just started messing around with SD last weekend, maybe 5/6hrs total

I have never been good at art and now I am

[vocal cloud]

you mean stable diffusion is good at art??

[manic bear]

Anybody how to use the image-image models to copy coarse details, e.g.: (style, mood, colors, medium) without copying the fine details (main subject, overall geometry, background features) ? So far my experience with image-to-image only allows me make slightly different variants of the base image.

I'm making AI generated stories and need the scenes to have the same colors and artistic style, but the content in each scene is quite different, but the image-to-image models make all my images look the same.

BTW I have tried editing the image_strength parameter but it doesn't do what I'm asking for really.

1. Raven and Phoenix, their faces illuminated by the glow of their computer screens, their fingers flying across the keyboard as they work together to hack into the vault. The scene is intense and focused, with a sense of urgency and determination. 

2. A split-screen image showing Raven and Phoenix on opposite sides, their faces filled with shock and disbelief as they read each other's messages. The tension in the air is electric as they realize the double-cross. 

3. The vault's server room, filled with rows of blinking lights and humming computers. Raven and Phoenix, connected to the server, their faces bathed in the glow of the screens. The atmosphere is charged with anticipation and excitement. 

4. A close-up of Raven's screen, displaying the message "50/50 split?" in a bold, green font. The emotions of hope and trust are evident in Raven's eyes as they await Phoenix's response. ```

[radiant rune]

I am pretty good at prompting.. but I like to think if I am close to what I sketch or see in my minds eye I am good at mixed media at lest!

[radiant rune]

Chilly react cuz what's cooler than cool?

[pastel moth]

Img2img won’t do what your looking for.

You’d need a reliable prompt for the style or a LoRA for consistency.

Img2img will retain the composition of the image, it doesn’t see the style.

[wispy nest]

Maybe style2image can help

It can generate stuff in the style of another image

You'll need controlnet

[obtuse palm]

Can somebody give me example and of tile images generated on auto1111?

[tulip stump]

#🏞｜general-with-images a young man with undercut hair is selfie at Tokyo

[flat remnant]

#1100170312106127410 devil cat

[loud notch]

finished about 50% of the training, had it running for the last 17 hours

hope the ai will stop duping the gun barrel and make the foliage on the tank a little more realistic looking

[versed swallow]

The Monkey GOD

[cyan dagger]

shits bout to get real

[covert hill]

How do I seek help? Been generating images non-stop for the last week.

[lean dawn]

How did this person achieve this?

I have the same model as them

[lean dawn]

Bro how do I achieve this

Or this for example

[bold mist]

Something about super low CFG just looks neat to me

[forest furnace]

controlnet img2img with a lower diffusion rate (0.25-0.5)

upscalers would also help

[prisma stone]

Why they put "()" stuff?

What does it do?

[prisma stone]

It's impossible to generate two colored eyes

Nothing can do this, I double doubt.

[wispy nest]

Emphasizes the prompt

wet might not make the image wet in some way
(wet) you'll probably get it more wet
(wet:1.3) people will be drenched probably
(wet:2.0) bro you're underwater

Kinda of how it works

(xxx:0.5) also works if you don't want something as strong affecting your prompt but still want it, I use it on smiles and blush mostly

[nocturne oak]

What, you don't like it when every image you generate with the word smile makes the subject look like they've been afflicted by The Joker's gas?

[wispy nest]

Try heterochromia

[wispy nest]

It's so funny sometimes kek

[nocturne oak]

Literally went through all this with them, btw:
#📝｜prompting-help message

Figured you wouldn't need to retread all that ground. 🙂

[wispy nest]

Oh

people who talk in multiple channels

[nocturne oak]

amirite?

[wispy nest]

I'll too!

[wispy nest]

That reminds me of when I was making Pokemons smiling very hard

Sending them to my friend saying "Would you capture this"

And it looks like a psycho in the woods

[heavy moth]

ohh this is soo badass. Mind sharing which model and the prompt?

[pallid ruin]

Wow 👀

[wispy nest]

wood man

[pallid ruin]

What model is that? Maybe that user is using controlnet btw 🤔

[plain kite]

Is something like this achievable with stable diffusion? If yes, Where should i start at?

[icy jackal]

doable with ip2p controlnet

[plain kite]

Any tutorials or resources where I can learn that from?

[icy jackal]

this random guide looks instructive https://stable-diffusion-art.com/instruct-pix2pix/ but it's really out of date
this is the official controlnet page problably have more info there https://github.com/Mikubill/sd-webui-controlnet

[plain kite]

Thank you 🙏

[mossy night]

Hi there! Not sure if this is the right place but I can't seem to find the reason anywhere

My question is, why do all my generations looks so... painty?

They all look like they're painted or something. I'm not sure how to describe it, but basically they all look like they have this weird painted texture. Even across different models. If you zoom in you can see it more easily.

And even with simple prompts like " a car" without any other description

Does anyone know what could be the reason? Thanks!

[bronze crescent]

Make sure you are choosing a style, and describing photo realism, some properties of cameras etc..

Big foot combined with alien monster, photo realistic, high quality, hdr, V-RAPTOR 8K, HDR, f1.8 shot by Frans Lanting

[mossy night]

That's the thing, I'm being quite descriptive I think, which is why I don't get it

[mossy night]

The prompt for that one is:

hybrid between a rock golem and a magic plant golem, skull, terrifying, intricate beautiful details, ambient occlusion, earth magic, detailed textures, sharp focus, ultra high pixel detail, cinematic, intricate, cinematic light, concept art, illustration, art station, unreal engine, realistic, movie scene, cinematic, high quality, craig mullins and wayne barlowe style, full colors, incredibly detailed, 4k, 8k

The negative is:
body out of frame, cartoon, low quality, low poly, poor quality, poor graphics

[bronze crescent]

You are basically asking for drawing. with concept art, illustration, art station

[mossy night]

Yes, but with the exact same prompt I've gotten images like this one

Though in that case it started with "hybrid between a goblin and a fungus" but the rest was the same

So I don't understand

And the weird thing is it's constant between models too

[mossy night]

K, let me try hyperrealistic, you're probably right

[bronze crescent]

Ya just leaving out the illustration stuff, and adding a couple realistic camera items

[dry crow]

Your just missing a vae file for that model

Its for color correction

[mossy night]

Ohh. How do I get that?

[dry crow]

Base vae for 1.5 models is the 84000-mse vae

[dry crow]

Here the direct link:
https://huggingface.co/stabilityai/sd-vae-ft-mse-original/resolve/main/vae-ft-mse-840000-ema-pruned.ckpt

Put that file into the models/vae folder

[mossy night]

Wow, thanks! Do I just add that into my models folder?

[mossy night]

Oh lol

[dry crow]

Then go into Settings>User Interface>Quicksettings and there add sd_vae
Then hit apply and restart

You then have a vae dropdown in the webui where you can select the VAE

[mossy night]

Omg thank youuuu

That worked really well

Thanks!

[dry crow]

No problem 🙂

[jagged garden]

Hi, I'm learning to create models. The old on the left i made like 3 weeks ago and the new one today. With the exact same settings. same model, same vae, same prompts, same seed, same after detailer settings...
But why it looks different. I notice that alot. The picture on the left has like more blue eyes and the one on the right, more green. Even tho i have in the prompt blue eyes, it makes me more pictures with green then blue eyes. and she seems have more makeup..?
For example on another model i have freckles on 0.8. Around a month ago, it was light freckles, now my model looks like morgan freeman. Even tho i change freckles to 0.2, nothing really happens.
Why is that so? Did something update in the background?
I think only the after detailer is a newer version now, but i don't have anything written down there on the eye color. Just more detailed lips...
Can anyone help me?
PS: Even she didn't had a nose piercing before

[prisma stone]

Now I understand why ComfyUI is completely very fast more than Auto1111

xformers on Auto1111 ain't do nothing tbh

[wraith karma]

Chat-bot bullying

[wispy nest]

you must be using outdated verison of A1111

[heady anchor]

thats for sdxl,he was using 1.5

[prisma stone]

I don't use SDXL sorry

It's very slow

I use SD 1.5 to 2.0

[heady anchor]

skill issue then since 1.5 is very fast in both auto1111 and comfy

[prisma stone]

You always call me a skill issue

What's wrong with you?

[wispy nest]

no issues with sd 1.5 and auto 1111 for sluggishness in my experience

[prisma stone]

Keep calling me that and I hope you get a moderation strike

[wispy nest]

low end gpu or too much addons?

[heady anchor]

lmao what for,we arent fighting here

[prisma stone]

Nah just only 1 addon tho

[wispy nest]

some of the LoRAs can be intensive

[prisma stone]

Maybe I forgot to use xformer

[wispy nest]

im actually having fun with SDXL you should try it out

altho not many cool community mods out yet

finger rendering could improve still

[wispy nest]

basic XL Base and Refiner

[dry crow]

Xformers does a lot in auto

It increases performance and lowers the vram usage nearly down to 60%

Its very important for low vram GPUs but helps bigger GPUs too

[pastel moth]

I found Xformers a bit tricky to enable properly the first time.

In AUTO anyways

I am mostly on Comfy these days.

[subtle condor]

lora:add-detail-xl:1wooden man standing on a hill his flesh like wood, fantasy, dungeons and dragons, metal plates, metal face
Negative prompt: (worst quality, low quality:1.4),,(EasyNegative:1), (monochrome:1.1), (ugly, jpeg, jpeg artifacts:2.0), (worst quality, low quality, lowres, poorly drawn:2.0), out of frame, six fingers,six digits,grainy,kids,kids face,anime,pixelated, BadDream, UnrealisticDream, 3 arms, multiple limbs, 3 hands, ,2d art, 3d art, ((illustration)), cartoon,blur,blurry,text,words,figurine,base,no base,platform,
Steps: 57, Sampler: DPM++ 2M SDE Karras, CFG scale: 6.5, Seed: 2160939781, Size: 440x728, Model hash: 165581bc60, Model: sdvn7Realartxl_beta2, VAE hash: 63aeecb90f, VAE: sdxl_vae.safetensors, Denoising strength: 0.72, Hires upscale: 2, Hires upscaler: None, Lora hashes: "add-detail-xl: 9c783c8ce46c", Version: v1.6.0

lora:add-detail-xl:1.5photorealistic, volumetric, wooden man iron face big jaw standing on a hill flesh like wooden vines covered in iron plates,fantasy, dungeons and dragons, striking pose, mahogany,iron, three fingers, two toes,no nose,glowing blue eyes,
Negative prompt: text,words,figurine,base,no base,platform,nose,mouth,ears,flesh,hair,skin,blue skin,human,(worst quality, low quality:1.4),bad-hands-5, badhandv4,(EasyNegative:1), (monochrome:1.1), (ugly, jpeg, jpeg artifacts:2.0), (worst quality, low quality, lowres, poorly drawn:2.0), out of frame, six fingers,six digits,grainy,kids,kids face,anime,pixelated, BadDream, UnrealisticDream, 3 arms, multiple limbs, 3 hands, ,2d art, 3d art, ((illustration)), cartoon,blur,blurry,
Steps: 61, Sampler: DPM++ 2M SDE Karras, CFG scale: 7, Seed: 3460269647, Size: 440x728, Model hash: 11168f6cc3, Model: xl6HEPHAISTOSSD10XLSFW_v33BakedVAE, VAE hash: 63aeecb90f, VAE: sdxl_vae.safetensors, Denoising strength: 0.72, Hires upscale: 2, Hires upscaler: None, Lora hashes: "add-detail-xl: 9c783c8ce46c", Version: v1.6.0

[pastel moth]

Some refinement:

[heavy moth]

Some results from initial, direct derivative prompt from yours, using realistic vision.

Want to get use to ComfyUI, I don't have a lot of Loras ready nor controlnet.

Can't get controlnet running on ComfyUI yet.

The poses and background are boring. No idea on how to add proper foreground yet.

But yeah learning as I go.

Went to deviate to female fantasy stuff. Still boring poses. But I do love the stoic look of the first generated image

[prisma stone]

Yeah but even If I typed in --xformers in webui-user-.bat

Nothing work

It's just come up with "proceeding without it"

Comfyui is easier to setup more than Automatic1111

[dry crow]

You also need to start the webui-user.bat and not the webui.bat

[mint perch]

hey guys,
can someone help me build a prompt for these type of illustrations

[kind quartz]

tarot cards, there is i believe LORA for it.

[fervent silo]

im having a problem using refiner. in comfyUI i start getting a okay image when i want to run it through a refiner it just adds noise?

[subtle condor]

I had good results with these SDXL FaeTastic: https://civitai.com/models/129681/sdxl-faetastic ProtoVision XL - High Fidelity 3D / Photorealism / Anime / hyperrealism - No Refiner Needed: https://civitai.com/models/125703/protovision-xl-high-fidelity-3d-photorealism-anime-hyperrealism-no-refiner-needed XL6 - HEPHAISTOS SD 1.0XL (SFW&NSFW) : https://civitai.com/models/119279/xl6-hephaistos-sd-10xl-sfwandnsfw I cant seem to find the sdvn7Realartxl_beta2 download now. hmmm.

[delicate tiger]

AOM3
I wanted a blank sky for a backdrop but this was found in the batch

a ((blue)) sky with a large sharp cloud, sunny day
Negative prompt: buildings, sun, oil painting, green, land, sea, (blurry, grainy, jpeg artifacts, noise, tiling, lowres, low quality, worst quality, monochrome, out of focus), text, logo, signature
Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 2422694690, Size: 512x512, Model hash: d124fc18f0, Model: AbyssOrangeMix3, VAE hash: ab33ad55c6, VAE: ClearVAE [v2.3].safetensors, Refiner: AbyssOrangeMix3 [d124fc18f0], Refiner switch at: 0.8, Version: 1.6.0

[heavy moth]

Sweeeet! Thank you very much for the pointer!

I haven't venture to SDXL though. Was using RealisticVision 5.1

Not sure I would generate fantasy settings much, as my main stablediffusion need is realism

But it's fun from time to time to make these majestic creatures!

[wispy nest]

My old plan is to found a workflow in comfy that can process complex regional prompts like A11 nicely. I'm making good progress, still working on it, no lora or controlnet needed, just the multiarea node, prompts and samplers. For simple pictures it's good, for complicated ones 1-2 out of 4 can be used. If anyone has any ideas on how to improve it or wanna test the workflows, let me know in pm.

[covert hill]

Any tips on how to do dark rooms where the model doesn't still look like they are on a spotlight?

[scenic umbra]

this is the sampler, did you use the dpm gpus with the normal scheduler? sometimes does that, try changing the scheduler to karras or some other

[fervent silo]

i just swapped the sampler out to the sdxl eff. from efficient nodes and havent had the same problem yet

[kind quartz]

@covert hill dim light? It is true i tried it just with toys, but could work for peple as well

[prisma stone]

Is it possible to generate images with only just few seconds?

My laptop setup is seem poor

I have seen Pixai can generate 512x512 with only few seconds and it surprised me

I use ComfyUI btw

[sterile temple]

which sampler are you using? the SDE samplers can be twice as slow as other samplers

[prisma stone]

DPM++ 2M Karras

[sterile temple]

how many steps?

[prisma stone]

For 30 seconds - 20 steps and 512x512
For 68 average seconds - 30 steps and 512x768

Are there Token Merging in Comfyui? How to get it?

[balmy arrow]

so the model's images looks awesome, but i only get this shit show, using the exact same prompt, settings and models... what am i missing? (i'm using a1111)

[prisma stone]

So what do you think?

[sterile temple]

I think maybe you can't get much faster unfortunately

[prisma stone]

Really? :0

[heady anchor]

u can with money

[prisma stone]

I'm broke lmao

[heady anchor]

yea the 16xx series cards are kinda slow on SD

[sterile temple]

Surely there's some rich kids at college. Just make friends with them 😛

[prisma stone]

What kind of series can generate images in just few seconds tho?

Maybe I can save my money lol

[heady anchor]

3090 and 4090 like 10 secs

[prisma stone]

Oke, Imma stick with Stable Diffusion around 1.5 to 2.0

I don't wanna go for SDXL, that's for high tier GPU

[loud notch]

hmmm

thats more like it

[weak loom]

Regarding my question in #💬｜general-chat
How can I use low denoise with inpainting to tell the model where to start from instead of middle gray?

(image might be confusing, but I'm trying to regenerate the background without the deer, not add it in)

[balmy arrow]

so the model's images looks awesome, but i only get this shit show, using the exact same prompt, settings and models... what am i missing? (i'm using a1111) (editado)

model: Samaritan 3d Cartoon
lora: Samaritan 3d Cartoon SDXL
prompt: 3D cinematic film.(caricature:0.2). 4k, highly detailed, Smiling woman taking selfie with boyfriend through smart phone, backpackers,nose ring, nature background
negative: drawing, painting, crayon, sketch, graphite, impressionist, noisy, blurry, soft, deformed, ugly. young. long neck. (cross eyed:1.5)
CFG scale: 7
Steps: 30
Sampler: Euler a
res: 1024x1024

the first image uses this exact settings: https://civitai.com/images/2117544?modelVersionId=144566&prioritizedUserIds=1501222&period=AllTime&sort=Most+Reactions&limit=20

what am i missing?

[sterile temple]

try removing the weights

like (cross eyed:1.5)

just make it cross eyed

[balmy arrow]

same bad result

isn't there a setting or another thing i need to checkout?

[sterile temple]

can't think of anything apart from the weighted prompt

[balmy arrow]

alright, ty

[subtle condor]

try (asymmetrical eyes:1) @balmy arrow in the prompt and put it toward the begining

[balmy arrow]

the eyes are not the problem, it's this poor quality

[livid ermine]

WarkaWater.com

Inspired by them. These are just for testing, the idea is to apply the same technology to skyscraper facades in tropical metropolitan areas.

They collect water, dry up the airand generate electricity. Kinda like windmills but way cooler.

[wary swan]

how many screencaps do you need for this type of result?

[prisma stone]

How to downgrade my xformers?

[prisma stone]

Is this how to use 2 Loras at the same time?

[dawn seal]

Ok, I'm not sure what's happening. Everything looks great, until the final few percent, then it all just falls apart.

I'm colmpletely new to this, so I'm sure it's something I'm doing.

[kind quartz]

if in last stage problems could be missing proper VAE @dawn seal

yes you have vae SDXL but using 1,5 model?

[wispy nest]

How much would it cost to train a model like the ones minevision has? Of pixel arts of guis, minecraft characters and renders?

[rain gazelle]

They're still of negligible value frankly. We need some new workflows.. something that creates ready-to-go game animations or something.

[thin sable]

Hello!

So I trained with a face and I wish to ask why the resulted dreambooth model doesn't have the same style as the model used as a base for training?

This picture is generated with the base model (without training - aniverse).

This picture is generated after training with a person's face for 1000 steps (10 photos) with dreambooth (thelastben colab) having the base model aniverse.

Used the same prompt, the same settings.
Why doesn't the 2nd picture have the same style (cartoonish lines) as the first one?
What should I do to get the same style even on the newest trained face?

[next arrow]

this might pose helpful, i spend time ordering the samplers based on testing over and over, now i can do this

minus the more or less bad ones

these i consider all purpose currently

[heady anchor]

heun my beloved

[next arrow]

I picked unipc sorry

it was between them for those 2

[heady anchor]

😔

[next arrow]

testing all day unipc vs heun unipc held more coherence as i kept testing new prompts

[next arrow]

these are essentially my current final pick of "works at low step, high step, low cfg, high cfg, any model etc etc"

the all purpose bundle

[obtuse palm]

same seed, same everything and i was trying to re generate the first image but came up with different ones nowhere near close to the first image

[obtuse palm]

@quiet current check this out please

[quiet current]

different ENSD, different negative and positive prompts

[obtuse palm]

does the spacing after a comma (,) matters?

[quiet current]

it might yes

[obtuse palm]

oops heck lol

i think some extensions changed the way prompts are being input

i'll try with a disable all extensions

[wispy nest]

If my prompt isn't like: one, two, three

I go insane

[quiet current]

set similar ENSD first

[obtuse palm]

there is no ENSD info given in the original data

[quiet current]

yup

but you are using one

set yours to 0

[obtuse palm]

the last image was without using ENSD

as 0 yes

ENSD 0 here

it's all the same with the comma and spacing or no spacing

tags like <Neg_Facelift768:0.8> , <bad_prompt_version2:0.8> are embeddings, correct? they change the output as well in presence of the embedding files present?

[quiet current]

yes those are embeddings, why they used them in between <> characters, no idea.

My guess is they don't know how to use embeddings :p

still, install those embeddings

if that can make you feel better, I get the same output as you without their embeddings and styles

so basically you're using the model correctly, don't waste too much time trying to get the exact same output as others from incomplete informations.

[obtuse palm]

still it remains as a mystery why I can't make my generations stand out like those civit ai shared examples

[quiet current]

different hardware, different auto1111 version, different styles, etc

bunch of settings not saved in the metadata

etc etc

[obtuse palm]

literally nothing i could make as perfect as them with a generation

[obtuse palm]

makes sense

[quiet current]

try to recreate different examples (not ones using styles for instance)

(you can see in the preview that I'm getting slightly different but still similar results)

(probably due to xformers+different hardware)

[obtuse palm]

are you on automatic 1.6.0?

[quiet current]

yes

[obtuse palm]

did you also generate that image that i shared?

[astral hinge]

anybody can tell me if its normal that my gpu only use 10% when using deforum? can i make it faster or is it just like that ?

[quiet current]

got the same results as you

[obtuse palm]

can you share a generation? i want to see if there's any difference compared to 1.5.2

i won't bother using 1.5.2 if it doesn't come with the same outputs as shared examples

sdxl on 1.6.0 works faster

[quiet current]

Like I said, I get similar results

[obtuse palm]

wow exactly similar to 1.5.2

so i guess automatic version doesn't matter much.. what do you think?

can you also show me your settings here? i think there's something about it

when i switched the ETA for ancestral samplers to 0 i got this output which was even closer to original [see the color pattern in her dress]

at least the shorter hair comes with the ETA reduction

[quiet current]

That's definitely the kind of setting not embeded in the metadata.
That's why I don't encourage you to waste time trying to reproduce the exact same results

[obtuse palm]

yea lol i won't bother it anymore

just go with my own ways

[quiet current]

unless you want to tweak every single setting... good luck with that

[nimble geyser]

aren't those just settings for the noise?

[nimble geyser]

nsfw?

[faint skiff]

I've been checking a bunch of Loras from the creator, Kappa_Neuro, and ... unless I'm doing something wrong ... they seem to just be artists that are already well trained in SDXL Base. I'm not sure they actually do much of anything.

[dawn seal]

Ok, I'll look at that. Thanks.

[wispy nest]

how dificult would be to create a blender addon to export a 2D render that each pixel has metadata of what each pixel means, based on the 3D scene context. and then use said file as input for a difussion model?
like a 2.5D texture
or a 2D texture + metadata of each pixel

[green grove]

Never have had this happen before and been using this since it almost first came out? Any ideas

[kind quartz]

i think it is clear message. Vote for which picture is better if A or B to remove it. You probably generating too much, not helping with voting.

[green grove]

You know how many times I need to vote to clear it? I've voted on at least 100 of them.

[heavy moth]

Servers need cost to run. The discord version is free to use. And yet you still complain about waiting for 2 minutes or voting for more

If you dont want to wait or vote just run SD locally. Use your own hardware and resources.

I've left my pc since yesterday to generate thousands of images. No restriction, no wait, no votes. Ez.

Oh also I can generate tiddies

[green grove]

😆 Well if I had a PC I would... but it just came as a shock that this has never happened to me before and I've been running it since like the beginning of the year...

[umbral wraith]

Hi guys, I just deployed SDXL refiner model on Google cloud vertex ai, I know it is a text-2-image model but can I pass in “image” as input with prompt (with the KWARGS) in order to achieve good results like what I got from dream studio?

Looking forward to any possible help here🙏🏻

[wicked yew]

I have developed an open-source discord robot based on stable differentiation web ui that supports most of the SD web ui's functions and extensions. For ease of use, almost all options are supported for selection as shown in the following figure

[obtuse palm]

ok so another question to ask is.. xformers makes a run faster also changes the output but does it change to make things better or worse?

[obtuse palm]

in all pictures 1st one is with xformers and 2nd one without it

[wispy nest]

makes sense regarding generating images in 1.43:1 aspect ratio

[heavy moth]

Dont see much difference. Only the yellow building behind, the roof. But probably because I'm on mobile rn 🤔

Like, can't you just compare them yourself and see whether the compromised output worth it or not, based off your need and taste

[obtuse palm]

well, the difference is barely noticeable but also those were generated with Euler A sampler so there was not much details to compare
maybe i'll try with a different sampler like DPM++ 2M Karras to find out how much changes happens using xformers or no xformers

now check out these examples.. no xformers used
all same inputs using DPM++ 2M karras for the above example
left one came from 1.5.2, right one 1.6.0
which looks more authentic? 🤔

this was the generation from civit.ai from which the data were used:

https://civitai.com/images/2083442?period=AllTime&periodMode=published&sort=Newest&view=categories&modelVersionId=142477&modelId=129922&postId=511493

i know my graphics card is kinda poor at this so my generations come out as the same 🥲

[obtuse palm]

but one thing to notice here.. there was visible rain falling in the prompt given which didn't follow in this example but it was certainly visible in mines 😄

i'm sorry didn't notice there were different seeds used in those versions

i'll redo things

[obtuse palm]

re-posting here:

Results from 1.5.2 and 1.6.0 come out exactly same
idk why i've seen people were talking about changes

the but yea one thing i'm certain about now.. the sampling method used shows it all.. you can clearly see difference from Euler A and DPM++ 2M Karras

noob to pro level difference 😂

[quiet current]

Probably from me. I said there could be change. Under certain circumstances

Feel free to consult (and bookmark) this page for more information about potential change from one version to another : https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Seed-breaking-changes

[nimble geyser]

that guy is back? he's been on the radar before. I seem to remember someone posting that he'd been banned. but I guess that did not happen

[faint skiff]

Yeah, I'm not sure to be honest. I'm certainly no expert at this. Just a collector of loras. He's very prolific and is putting out many loras per day. I've been downloading many of his and just recently started testing them. But what I was realizing was that the artists seem to already be trained in the base model and there is very little difference when applying the LORA. But it's totally possible that I'm just not testing correctly

[nimble geyser]

if you get the chance compare the file sizes of his loras

[faint skiff]

Here's a few of them. What are you looking for?

[nimble geyser]

why would they all be precisely the same size?

I don't think they're actual fleshed out loras

probably just emphasizes the artist or style it's focused on

[faint skiff]

Yeah, that I can't say. I'm not very knowledgeable about how training works. Maybe he's got a really finetuned system for training and sort of automates it. I can say, though, that I'm getting rid of most of his Loras. I just don't see much effect

[nimble geyser]

well quite a few people were looking at them and from what they were saying, at least then, they weren't really legit loras

I don't think you could really fine tune a system to make them that close in size

if you were training them on external data

unless all settings, images, etc, were precisely the same

[urban badger]

/imagine,imagine a book cover for this --v4

[rocky quail]

i think this looks very cool

[sturdy pewter]

https://twitter.com/GozukaraFurkan/status/1700514506231672988

Furkan Gözükara (@GozukaraFurkan)

6 SDXL DreamBooth trainings started

Surprisingly the speed is same as SDXL LoRA training

I am training both Text Encoder and U-NET

Using Kohya GUI latest version

Using RTX 3090 24 GB GPUs

Currently trying to find best hyper parameters

Please keep supporting me on Patreon

Twitter · Sep 9, 2023, 2:20 PM

[bronze bramble]

what is going on? why is the other chats keep getting deleted when I click them??

[blazing gust]

Help!!!!!

[wispy nest]

guys what does it want?

i have an AMD gpu

[daring kernel]

Click on "SDXL BETA BOT"

[bronze bramble]

omg thank you so much it works! I was getting desesperate

[daring kernel]

I'm glad it helped you

[fallen dock]

Converted from a 3D blender scene I modeled.

SD just helps with color, shading, etc.

[lapis quarry]

Question.

How can I have SD expand on the drawing.

Like completing her hair.

[heady anchor]

outpainting

[lapis quarry]

?

[sudden jackal]

How everyone can draw the desired scale structure as fast as possible

It took me a long time to get this composition

[wispy nest]

https://civitai.com/models/112552/weight-slider-lora

ok who made this

fess up

[thick turret]

#🏞｜general-with-images car

[heavy moth]

this is actually really gooood.

I'm currently using it in my workflow

[wispy nest]

When TF is the 7950 xtx coming out

Should I just get the 7900 xtx

[sterile kiln]

Me as master jedi, masterpiece ! ❤️

[cobalt pollen]

some issues you should address, there is lazer inside your hands where it shouldn't be, the lightsaber sword is too long, i would inpaint the hilt to end much sooner.