🏞｜general-with-images

[shell sleet]

I have one for line art and then openpose

I googled it but I dunno

[wispy nest]

i dont use open pose much, but i dont think you need t2i for that

[shell sleet]

It makes it significantly less heavy and makes it faster

[wispy nest]

just preprocess the image to get the sketch/pose/depth/normal/etc , then turn off the processor, load in the processor result in cnet, turn on model (NOT processor) and generate

fair enough

[shell sleet]

I have pre-made open pose things

Found them on civitai

[wispy nest]

so performance for you is better using t2i rather than just an open pose model w/ no preprocessor?

[shell sleet]

T2i doesn't need a preprocessor

[wispy nest]

huh i might need to go backl and read more on it..

lol we are talking about the same thing 😄

[shell sleet]

..I misunderstood my bad

[wispy nest]

t2i is alittle different, i think. its a preprocessor if im not mistaken

[shell sleet]

Nah

It's like regular models, but lighter and faster. Only runs once throughout the entire generation, while regular runs once per iteration

[wispy nest]

or at least, there are t2i preprocessors. maybe ive just got my controlnet folder all messed up

[wispy nest]

both run throughtout assuming start and end steps set that way

also, i highly recommend the dw open pose full, it handles hands, fingers, and facial features as well as basic pose

depth and normal maps are really good too

[shell sleet]

This was a pre-made, not one I made

[wispy nest]

got ya

[shell sleet]

It was part of a batch on civitai

But yeah I'm incredibly confused.

[wispy nest]

(also you can take the image you generate with that pose, load it into controlnet, run dw open pose full pre processor, and then you can edit hands,fingers etc on the pose if you want more control over everything)

confused about what

[shell sleet]

Why this error seems so random

[wispy nest]

ALMOST got it... theres just some random demon ghost faces here and there, but it didnt mess up the character at all

yeah if you wanna send the log ill take a peek at it

[shell sleet]

I'll send it whenever I get it again...

Like using regular cn models absolutely wrecks speed

[wispy nest]

you should try miaoshou assistant

helps with memory release

might help, might now lol

might not*

[shell sleet]

I have uh..memreduct

[wispy nest]

never heard of that one

[shell sleet]

I've had it for a while

I just have to wait for that error

I 100% need to go lay down, I'm tired af

[wispy nest]

haha have a good one

[shell sleet]

It's also 4am lmao
Good night folks

[languid pebble]

Good to hear some do it with 4GB!

[wispy nest]

ok finally

got high denoise upscale without creating a mosaic or completley chaning the image, in fact it seemed to put more emphasis on my prompts.

i also was doing this with a halfway edited image, and it went ahead and fixed the stuff i was going to do in a photo editor 😄

[dull flame]

is this not the way to go ?

i cant make her grab the sword at all

no matter how much i change denoise etc

[wispy nest]

GAH

how can people be using a white background like that.. it is a war crime on the eyes

[dull flame]

https://tenor.com/view/peach-peachcry-cry-tissue-emote-gif-27496755

tears are coming for real soon

[wispy nest]

theres several ways you can do this. try adding to positive and negative prompt, (negative open hand etc) (positive clenchec fist, gripping, hold weapon, etc). also try using control net with inpaint model and upload that as controlnet image, set denoise to 0.7-0.9, whole image checked, mask the hand, llama inpaint preprocessor,.
or what will probably be faster is just do it in gimp or something. 2 layers. one layer with original, second layer (generate image again, same settings same seed, negative sword.) take your original image and put it as top layer, add alpha channel, erase where the hand should be (revealing second layer with hand) , merge layers, low denoise inpaint or img2img. or just completley do it all in one layer, wouldnt be too difficult with alpha channel, erase, clone , then run through img2img to clean it up.

[dull flame]

https://tenor.com/view/betty-white-math-calculating-confused-golden-girls-gif-27641696

i will have to read that a few times

[wispy nest]

also, what are your inpaint settings? masked only? whole picture? original fill or latent?

[dull flame]

wtf, why does it decide it wants to cooperate now

i changed to fill, and denoise of 0,75

but im sure ive done that alrdy

[wispy nest]

^ those are the important settings when asking about inpaints 🙂

[wispy nest]

something else that can help, try setting batch to 4+, and maybe turn on extra seed and crank the variation up to like .25+, if you arent getting results you want then throwing in some random noise can help.

[dull flame]

trying it now

wait, what do u mean random noise?

gah

maybe 1024 is just not enough for inpainting

[wispy nest]

sorry poorly worded, i just meant that enabling extra (setting by seed) caqn introduce more variation in results. also 1024 is plenty for inpaint

[dull flame]

i see

[wispy nest]

hands are also something sd struggles with. so, control net is recommended here. and, adetailer with a hand model is helpful as well. both will help avoid theose mangled ass hands youre getting

[dull flame]

so if i use only maked, does it use 1024x in that little area i marked?

[wispy nest]

no, only masked is going to only focus on the area you masked, + taking into account padding and blur. (theres some tricks here where you can do things like mask a very tiny or multiple tiny dots around masked area to have the model focus on a larger area as it inpaints only what you have masked), and whole picture the model will look at the entire picture as it inpaints what is masked. honestly youre really shooting yourself in the foot by not using controlnet while you inpaint though..

honestly, to get a good result using ONLY inpainting and none of the things im recommending, you need to add more to prompt and generate large batches because its going to come down to luck 🙂

[dull flame]

alright ill try to understand controlnet then

[wispy nest]

and , this one is just opinion, but i find it easier to inpaint or do any editing at lower resolution than 1400+ that youre using. for one its going to be quicker. two, the results wont be as crisp but thats what upscaling is for anyway. you just wont it to be mostly right before you do that

[wispy nest]

its not as challenging as it sounds! tons of good videos on yt that are like 5-10 minutes long that will get you started

using sd without controlnet is like riding a bicycle with no handlebars. gamechanger.

[dull flame]

alright, inpaint controlnet pixel perfect

[dull flame]

sigh, now upscaling doesent work for me, this is not my day

i might have to go back to comfy

[tawdry vapor]

https://tenor.com/view/sending-love-gif-2080182813636534225

[clever oar]

shark koi

[wispy nest]

😮

[nimble mason]

Bwahaha

[jovial tiger]

I guess this explains why I haven't seen any more sd3 pics from these people I followed.

[clever oar]

for some reason the neural network often thinks that what I want is not an animal, but an object

[hazy bluff]

/ogurt packaging design, Hourglass shape special-shaped box, cute, creative shape ，mattetexture，round，goldandblack，elegant，

[pastel root]

that frog has a tushie!

[lyric root]

Guys I need help

[dull flame]

it helps him float

[lyric root]

What is that extension called? Does anyone know?

[dull flame]

yes, one sec

[lyric root]

I just saw a youtuber randomly use it, and I was like WHOA

I need that!

[dull flame]

https://github.com/DominikDoom/a1111-sd-webui-tagcomplete

there u go

[lyric root]

Trying it now

[shell sleet]

ooooo those are gorgeous

does anyone know why controlnet would just be ignored, even if it's enabled?

[languid pebble]

Models not downloaded/ in the right directory?

[shell sleet]

no, they're in the right directory, and the little box is enabled

one sec

i was thinking i probably had to like...guide it by putting in a specific prompt

normally i don't have to, it just...does it

I set it to balanced just now, maybe that'll help

That seemed to do it

[clever oar]

read what say your console

[shell sleet]

I can't read this very well

[quiet dome]

please send me an any AI stable diffusion generated demon, I need it for demonstration

[clever oar]

can you rotate camera more?

[languid pebble]

No control over the animation possible ...

[clever oar]

oh

[lyric root]

I am frustrated

[languid pebble]

SVD ... you give an image and get an animation ...

[clever oar]

i see on civitai full rotate camera

[languid pebble]

Sounds like DeForum ...

[lyric root]

I brought it into Krita, and fixed a bunch of things, and still working on things, and I have come to a part I am just not good at.

[clever oar]

i can use svd on my 4 gb vram card?

[lyric root]

He has no nips, and idk how to draw nips. e-e Anyone got any magical lora that adds them?

[languid pebble]

Give it a try ... I don't really know ...

[languid pebble]

You can use the pic in the USA that way ... 😄

[lyric root]

So just leave him nipless?

I don't really mind, honestly

[languid pebble]

Maybe try inpainting?

[cyan shoal]

24gb?

[languid pebble]

Missing 24GB?

[clever oar]

oh

i forget

i can add ?

or need new

lol again 24 lost

[languid pebble]

🥳

[clever oar]

how many millionaires 😃

24 gb..

not cheap

[shell sleet]

I love how jank forge looks during genetation

Generation

Mobile sux

[grave scarab]

just started learning stable diffusion last night

its amazing the things u can do with it

but it's so complex, so much things to learn

kinda overwhelming

that was the best i could pull off so far

[languid pebble]

Well ... learning is part of the fun!

[grave scarab]

i'm way too impatient

[languid pebble]

I'm learning over 2 years now ...

[shell sleet]

i made an apple to test the speed on my s/o's computer with forge... it goes fast

legit only used the word 'apple' and it came out great 😆

🤣

[languid pebble]

An apple with instruction ...

[shell sleet]

I'm not even using sdxl so that logo came out great, didn't change my prompt either

[languid pebble]

Good nite!

[shell sleet]

i got curious

[wild goblet]

Man cyberpunk

[shell sleet]

yeah good luck with requesting images when the bot is down

[clever oar]

apple watch

[clever oar]

model for extension not work...

[lyric root]

Guys!

[clever oar]

its someting new

[lyric root]

This

to

Finally figured out what was wrong

It was doing this, couldn't figure out way.

[lyric root]

Because of the dang hidden feature restore faces

XD

[grave scarab]

pressed to generate something 20 min ago and it's still going

is that normal ?

img2img

[lyric root]

Depends on how high you set the upscale to

I just started an upscale of x2 on top of running ultra upscaler x4, says it'll take 20 mins

XD

[grave scarab]

damn

i thought my pc could handle it

xD

[nimble mason]

i'm doomed

[nimble mason]

in, out

for comparison: 0.5 denoise exponential, then karras at 0.4/0.45/0.5 denoise

[wispy nest]

you can change that in settings : live preview 😄

[nimble mason]

okay what i've got here is badass as hell

i don't think i've seen anyone do this... (though I can't possibly be the first to think of it) ^^^

in, out

[mild jay]

may i ask what model did you use?

[nimble mason]

with 0.45 and 0.5 denoise with karras. total joke by comparison

[nimble mason]

this is juggernaut but i'm doing some fancy/insane tricks with the scheduler in comfyui

using res_momentumized as the sampler

[mild jay]

thats cool, and juggernaunt can be insanely good at times if done right.

[nimble mason]

yeah

the model isn't too important here but yeah it's def a very good one

this is with 50% unsampling/resampling with karras

0.5 denoise with exponential, and then 50% unsampling/resampling with exponential. so yeah, i really do have something here

iterative unsampling/resampling via a sine wave sigma scheduler

[clever oar]

Do you need to connect hundreds of lines each time?

[nimble mason]

lol, no

i really do work just as fast in comfy as a1111

[clever oar]

its real?

[nimble mason]

yeah

[clever oar]

I thought you had to be a scientist to create something there

[nimble mason]

i do have a chem phd but i'm pretty sure it doesn't help me at all with this

the thing that makes it hard is the lack of documentation

if there were nice written guides that actually explained what each option did with an example, it'd be easy

[clever oar]

i try simple extension for animation and it not work

errors...

is so hard to start something

[nimble mason]

extension with a1111?

[clever oar]

yes

[nimble mason]

the best thing to do is use someone else's workflow in comfy

[clever oar]

animatediff

[nimble mason]

also, i remember you saying you had low vram... comfy uses less vram than a1111, that was why i tried it

[clever oar]

also you saw my poll about video cards?

[nimble mason]

i didn't, link?

been really swamped with work the last few days so i prolly missed a lot on here

[grave scarab]

can i buy an image from someone that knows how to mess with stable difusion ? if yes, where ?

i have 3 days to create something but i cant get it right

[nimble mason]

what are you trying to make

[grave scarab]

too rookie still

something from img2img

a scythe weapon

[clever oar]

#🏞｜general-with-images message

[nimble mason]

cast my vote (4090)

man i feel bad for the person with no gpu at all

ouch

[clever oar]

lmao

maybe we all help him

[nimble mason]

in, out

[clever oar]

scorpion

[nimble mason]

maybe post here what you need and maybe someone will help?

[shell sleet]

does anyone know of a zora lora? Like Zora from zelda

i dont have the resources to train one, nor the knowledge

[nimble mason]

for sd1.5 i presume?

[shell sleet]

yeah

i'm reading a guide for it and it feels like it's going right over my head

[nimble mason]

haven't seen one, but it's not too hard to do

the guides are a mess

here's what you need to do... get 30 or 40 images of zora together with the most diverse angles, lighting, poses, gender, etc possible

with a consistent size, prolly 512x512 since you're takling sd15

no bad quality ones, it's better to have a small set than crap ones thrown in

[shell sleet]

Like i understand the file stuff, like 1.png, 1.txt with 1.txt having a bunch of the image information

like tags

since im not on my laptop, i probably do have the resources for it...

[nimble mason]

don't worry about that

you'd want to use onetrainer too btw

[shell sleet]

onetrainer?

[nimble mason]

it's easier and uses less vram and is also faster

https://github.com/Nerogar/OneTrainer

that's what i use

[shell sleet]

ahhh

okay

[nimble mason]

focus on the data set first

get that and i or someone else can show you what to do next

don't worry about the naming either

[shell sleet]

alr, i'll do that while i wait for onetrainer

[nimble mason]

in, out... those eyes

[shell sleet]

most of these are game screenshots, would that work?

[nimble mason]

if that's the look you want, yes

but again... diverse backgrounds, outfits, everything

ideally you want to make it so every single aspect of the image changes except for whatever makes a zora look like a zora

[shell sleet]

..yeah i dont think i'm gonna be able to do that.

[nimble mason]

it doesn't have to be perfect

[shell sleet]

these are all of mipha, it's gonna have like no variation

[nimble mason]

but if you have, say, zora after zora that only shows up swimming in water, the lora will have a hard time producing a zora on land

yeah, you'll want to get more

nice excuse to play the game more haha

[shell sleet]

haha yeah, i'm just googling these though

[nimble mason]

yeah you prolly want to fire up your switch and get your own screencaps

if you're going for the in-game look

[shell sleet]

nah, i'm going for like... the ability for it to blend in with almost anything. Like if i were to throw it into an anime style i'd get anime style, but throwing it into realism i'd get semi-realism... I know that aint happenin' but still

i might have to run around and get screen caps... i'd just have to charge my switch up.

[nimble mason]

it's doable, but yeah, you'll want to start with a diverse set if you want to go in that direction

you'd probably have to do some img2img work to create a synthetic dataset

but starting with some in-game stuff would help ya

you can sometimes use a lora that's "stiff" and inflexible to generate just enough new data to train another one that's more flexible

[shell sleet]

ahh

[lyric root]

Made this by mistake, and I love it

[shell sleet]

oooh

okay so for the training and stuff, does it have to be 512x512?

can they be a bit bigger?

in order to get full body it'd have to be bigger

[thin echo]

Amazing

[shell sleet]

...oh what does this mean

[nimble mason]

oh yeahhhh

[nimble mason]

looks like a preset is missing

[nimble mason]

best you could do is maybe 768x512

[shell sleet]

that'll work

that's what i use for basic images anyway

[nimble mason]

cool

yeah the resolutions are so so important

if you want to see why, try generating "a woman walking on the beach" with 512x512, 768x512 and then 768x768

mutant city with 768x768

[shell sleet]

oh, yeah I know the difference...it does a lot with it

if you make it too big, you get a mutant

if you make it too small, it looks bad

the size changes a lot of stuff

[nimble mason]

yep

when you train a lora, you're just basically refining a model that was already trained

and those were trained primarily on 512x512, and a bit on 768x512, not so great at many others

[shell sleet]

okay so i found like...11 images and resized them.

I would boot up my switch but it's charging right now

[nimble mason]

gotta be careful with resizing

cuz of the risk of quality loss

[shell sleet]

I know, if i resize it wrong it gets all smushed

or that, yeah

[nimble mason]

cropping is fine, downscaling with lanczos is fine, but upscaling you wanna avoid if at all possible

it's better to pad the image by outpainting than to upscale

[shell sleet]

...what if the image is smaller than 512x512?

like i dont know how to use outpainting

[nimble mason]

you'll want to read up on that 🙂

it's worthwhile

[shell sleet]

Yeah but they use so many big words and tech terms that it just goes right over my damn head

[nimble mason]

outpainting is the same as inpainting, except on the outside of what you got, instead of the inside

[shell sleet]

"You need to activate the schmorgus setting inside the yufidoo..."

[nimble mason]

lol

you should try that as a prompt

[shell sleet]

😆

Yeah, hold on i'll do some reading

it might actually help me

oh nice okay that wasn't that bad.

i dont know what im doing in order to make this work, but i just kinda...made it bigger i guess

it counts.

probably not.

....

that definitely doesn't count

why can't i throw it through an upscaler exactly?

it still looks good...

[nimble mason]

if it looks good enough to you, then it's fine

[shell sleet]

okay so i have my images

this is just a test, so i'm not worried about if it's 100% good or not...

the page gives no instrunctions on how to use this. love it

ah wait there it is

it says i stil need to do the txt file thing

[nimble mason]

yeah there's a program to do that for you

i think onetrainer can do it too

[shell sleet]

I don't see where

[nimble mason]

https://github.com/starik222/BooruDatasetTagManager

this is what i use

[shell sleet]

...yeah i dont think I'm cut out for this. I'm reading but retaining absolutely nothing.

It's going right through my brain.

[nimble mason]

that's what happened with me pretty much every time i read it

just get that program installed

it's pretty easy once you have that

[weary light]

[lyric root]

Anyone know what this means?

[nimble mason]

you divided by zero and created a singularity

[lyric root]

What

:c

What is wrong?

[shell sleet]

hey i got that bug too!

I don't know WHAT causes it

[lyric root]

I have done everything recommended, and it just refuses to work. It's only doing it with this model, so maybe it's just refusing to work

[past pelican]

Lmao

[lyric root]

Now no matter what I do, it's giving that error

What do I do?

[shell sleet]

i restarted forge

like i closed and reopened it

and it still happened yeah, but i'd just restart each time

[lyric root]

I did that, and it didn't work

[shell sleet]

i guess make a bug report on the github for it with a copy paste of your console?

[lyric root]

I am restarting again, to see if it works

Okay, it's working again, but no idea what caused that

[ruby gulch]

Does anyone know how to make the first image as good as the second one?

[grave scarab]

is it possible to take this scythe and enhance it ? like this glow around more detailed, everything sharp and high res ?

im trying for 2 days, can't get it right

tried different models, idk what im doing wrong lol

[nimble mason]

Here is the image you requested.

[grave scarab]

ye those are completely different scythes tho :/

but thanks ^^

[wild sorrel]

Just upscale basically?
Maybe little bit of inpainting for more details

[grave scarab]

okay im gonna search for guides on those

thank you

[wispy nest]

how do you use inpainting

[hazy warren]

I play totk too

[mental flame]

Hi Everyone. Not sure if this is the right channel for this. I'm looking for a stable diffusion / MidJourney professional who can assist with a project on digitally altering images of socks. I have PNG images and 3D files of the socks. The goal is to take these images, keep the socks unchanged, and completely transform the model and background to a design of my choice. I would also love to learn this process. If anyone is skilled in these techniques and is open to collaboration and teaching, please DM me. I've attached an example of the final result we want. Thanks!

[dark harness]

a 25-year-old friendly-looking man sitting behind a desk in a futuristic studio, wearing a yellow hoodie. window background, smooth, soft, ultra-sharp, detailed, looking straight forward, centered in the image, straight, front-facing a camera.

[nimble mason]

Here is the image you requested.

[south temple]

hi friends

[nimble mason]

Beep boop!

[languid pebble]

Here we say: Moin! 🙂

[nimble mason]

Here is the image you requested.

[languid pebble]

That reminds me of a friend who takes Polaroid pictures on parties and writes jokes under them ...

[wild sorrel]

Which webui are you using?

[wispy nest]

i use forge

[wild sorrel]

never used it, but it looks ~same as a1111, so...there's img2img section => there should be inpainting
You mask section you want and let AI change it...there are some settings, making it a bit mroe complex, might want to go through docs or some vids about it

oh and you will need inpainting model, usually models have 2 versions - base and inpainting version

[wispy nest]

a1111 is what i use im pretty sure

i just thought it was called forge

[wispy nest]

okay so i gotta download some stuff

[wild sorrel]

at least that was the case with 1.5x models, idk what's up with SDXL and if it can do inpainting or need something more

[nimble mason]

it can, but not as well as sd15

[nimble mason]

my new fav denoising schedule

[nimble mason]

gorgeous

[crisp stream]

thank you 🙂

[nimble mason]

you might like that denoising schedule

[nimble mason]

workflow embedded

that noise scheduler is giving me some of my best results ever

[crisp stream]

ty, gonna check

[languid pebble]

Using the noise thingy?

[crisp stream]

no, not yet 😄

[nimble mason]

someone asked for a mailman in another section...

[midnight kettle]

How to install stable diffusion in low end pc

[languid pebble]

How much VRAM?

[midnight kettle]

My pc has amd Radeon A4 video card

It is a notebook laptop

[languid pebble]

A web service to create could be a better idea

[midnight kettle]

Do you know any web service

For free

[languid pebble]

Leonardi.AI is giving some free tokens every day ... https://www.craiyon.com/ should be free

[languid pebble]

Realtime Generation @ leonardo.ai is for free and interesting for learning, too

[crisp stream]

so, the image upscaler node doesn´t work even if installed. Already tried "try fix", yet didn´t work either

[languid pebble]

Take your time 🙂

[crisp stream]

It just doesnt work, so what´s got time to do with it? 😄

[crisp stream]

http://dream.ai

Freemium

[languid pebble]

Learn ... 😄 Just kidding ... had a problem with the SUPIR and stopped working on that.

[crisp stream]

had to update everything, now it´s working 🙂

[languid pebble]

Looks like it was worth it ^^

A bit more ?organic? than the other style ...

[crisp stream]

still doesn't work though, something with controlnet missing 😀

even though it installed tons of c-net stuff during the update

[jovial tiger]

@nimble mason impressive stuff. new release from pixart, sigma (versus their old alpha). they have a free space to try it out. https://huggingface.co/spaces/artificialguybr/Pixart-Sigma

I couldn't get sdxl to do this with any amount of prompt expansion and trying various models. this is impressive stuff. I think we're about to start seeing an explosion of new models that use T5 llm models as part of the render pipeline like SD3, Ella, and now this.

[cyan shoal]

I run it offline with comfyui-extra models

there, you can run T5 at fp16, 8-bit and 4-bit, with no conversion

[jovial tiger]

Have a workflow handy for it that you can drop on here?

[cyan shoal]

well I'll measure the VRAM first for you

you have 10GB or how much?

[jovial tiger]

Yeah, 24 gigs

[cyan shoal]

oh lol

then the one on the repo is perfectly fine: https://github.com/city96/ComfyUI_ExtraModels

https://private-user-images.githubusercontent.com/125218114/289118097-eb1a02f9-6114-47eb-a066-261c39c55615.png?jwt=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJnaXRodWIuY29tIiwiYXVkIjoicmF3LmdpdGh1YnVzZXJjb250ZW50LmNvbSIsImtleSI6ImtleTUiLCJleHAiOjE3MTMxMDA1NzQsIm5iZiI6MTcxMzEwMDI3NCwicGF0aCI6Ii8xMjUyMTgxMTQvMjg5MTE4MDk3LWViMWEwMmY5LTYxMTQtNDdlYi1hMDY2LTI2MWMzOWM1NTYxNS5wbmc_WC1BbXotQWxnb3JpdGhtPUFXUzQtSE1BQy1TSEEyNTYmWC1BbXotQ3JlZGVudGlhbD1BS0lBVkNPRFlMU0E1M1BRSzRaQSUyRjIwMjQwNDE0JTJGdXMtZWFzdC0xJTJGczMlMkZhd3M0X3JlcXVlc3QmWC1BbXotRGF0ZT0yMDI0MDQxNFQxMzExMTRaJlgtQW16LUV4cGlyZXM9MzAwJlgtQW16LVNpZ25hdHVyZT1hZGIyMGRhNTJjZDRjN2JmODI2YTg0ZmE0NmI2MWFmYWE1NWY1YWRkYmRjMTIzMjFiNTU2NDAwYjUwMGQ1ZGE4JlgtQW16LVNpZ25lZEhlYWRlcnM9aG9zdCZhY3Rvcl9pZD0wJmtleV9pZD0wJnJlcG9faWQ9MCJ9.Qcbyl1N-0CVbPd6vnzUoGDMBjkFltC0v600G_m6kCE8

its quite easy to install

only problem is (with ELLA as well), is the amount of time T5 needs to load in

its quite slow

but yeah with 24 gigs you can run T5 at fp16

@jovial tiger just wanted to tell you that this and ELLA aren't good at text btw

but they are good for what you are usually doing

complex scenes

I wonder if you can do regional prompting cause of the close integration with comfyui

I mean it has conditioning right there for you to concat/combine and whatever

[jovial tiger]

@cyan shoal awesome, thanks. going all over the place to download stuff. yet again. 🙂

[cyan shoal]

yeahhh heh

@jovial tiger

[jovial tiger]

yeah, i tried one of those as well and it was already better. still downloading. can't wait to see what my llm expanded prompts do with it

it's not perfect, but it's a large step up.

certain actions are still not going to be there, but I'll take any leg up at this point.

[cyan shoal]

^

[jovial tiger]

From reddit thread: orange cat wrapped in white bandages and black dog wrapped in red bandages sitting on a bench on top of a hill filled with round stones, photo, cinematic

[pallid ruin]

cool! ⚔️

[jovial tiger]

@cyan shoal what resolution settings are you using for the empty latent?

I keep trying to use my own and it says not good. what's the best way to get hi-res with this?

[cyan shoal]

wait

show me the Pixart Resolution Select node

and the options for it

[cyan shoal]

the 3rd one

pixart sigma xl 2

that is for 1024px

[cyan shoal]

epic

[jovial tiger]

sure, but on the demo, you can do 1920x1080 for instance.

I tried making an empty latent with 1920x1080 and it refused.

[cyan shoal]

that's odd???

you know

[jovial tiger]

I guess I'll try the usual upscale methods. have you tried samplers other than the default euler?

[cyan shoal]

tjere are only 4 model types rn available to the public

256px, 512px, 512-DMD, 1024px

then there are 2 remaining models that are not available yet: 2K and 4K

you could try kohya's deep downsample

or just generic highresfix maybe

[cyan shoal]

so they guy probably use kohya's deep downsample or something other

[jovial tiger]

yeah latent and image upscaling arne't working.

just get a stretched image

[cyan shoal]

kohya's deep downsample

iirc that worked

[jovial tiger]

there's a high change I'm using this wrong. 🙂

[cyan shoal]

hol on

ok it doesnt work for me as well for some reason

did you load T5 in 8bit

[jovial tiger]

I loaded the 20 gig t5

fits on my gpu. 🙂

[cyan shoal]

but in 8-bit or fp16

[jovial tiger]

wow. he's even holding the skulls i had in there.

I guess I'm happy for now. I'm using 1.67 ratio, which is 16:9. the output is amazing, so I won't fiddle any more.

these sampler settings give really good output

gigantic robot reindeer dwarves tiny santa who is looking up at it, swirling snow, ethereal christmas lights,,ultra highres, High detail RAW Photo, , dslr, film grain, ultra detailed, 8k, masterpiece, hyper realistic, photorealistic, photograph, sharp focus

wow: orange cat with white hat sitting on a park bench next to a black dog wearing a blue scarf and rasberry beret @nimble mason

[nocturne oak]

Is that beret the kind you find in a second-hand store?

[jovial tiger]

you know what? it is!

this is kind of nuts.

I select cpu for the t5 model, and once it's loaded, it only uses 3 gigs of vram. and it's no slower than loading the whole thing on the gpu instead.

and their model isn't censored either.

A man in a rugged helmet grapples with a towering, anthropomorphic Cheeto in a dimly lit living room, as if straight out of a surrealist painting. The camera captures the scene from a low angle, highlighting the absurdity and drama of their wrestling match.

[vapid crest]

think im using upscale wrong, getting outputs like this

[jovial tiger]

you need a second ksampler with a 0.5 denoise after the upscale latent.

[vapid crest]

do i plug in the same stuff for model, +ve and -ve prompts?

[jovial tiger]

correct.

just that the latent input is from your upscale latent node instead of the empty latent from the beginning.

[nimble mason]

Oooo I rewemb.r hearing they were working on it

Those images look great

Same T5 files as before with alpha?

[jovial tiger]

yep

[nimble mason]

I have em on my HDD, can move em back

Sweet

[jovial tiger]

I did notice that if you made a reeeeally complicated prompt, it needed more steps. so 50 instead of 30 for res_moment

but it did it

[nimble mason]

Nice

Yeah the main issue i remember with alpha was either censorship or under training or both

Had a pretty limited vocabulary

What it knew it was very good at though

[jovial tiger]

it's definitely not censored.

I tried both main uncensored angles and it did both

only catch is that there's no upscale. @cyan shoal mentioned that 2k and 4k versions of the model will be released at some point.

so with a 1.67 ratio, it does 1280x768 which when the prompt is adhering so well, is fine

[nimble mason]

K cool

Yep fuck it

[jovial tiger]

@nimble mason

here's the workflow.

[nimble mason]

That's really good to hear re: censorship

I don't even care about making that type of content but when it's censored as hell it really does affect its ability to generate tons of peripheral stuff properly

lol @ them dropping a pickle checkpoint 🤣

i'll use it but jeez what a way to ensure large numbers of people will use it without hesitating

[nimble mason]

regarding res momentumized... def use the samplercustom version, those extra options (especially the noise sampler and sigmas) make a really big difference

i'm gonna see if i can come up with a better schedule for denoising in general

The perlin sampler is nuts for crisp details in most cases

For noise, uniform is often better than gaussian espec in combo with the perlin sampler

[jovial tiger]

In the foreground, a meticulous mechanic, clad in protective garb, wielding a powerful welder, strikes a focused pose amidst a shower of sparkling arcs, adding intricate details to the colossal robot's metallic body, while towering skyscrapers rise imposingly in the background, emphasizing the immense scale; the scene is captured with a long exposure, creating a breathtakingly detailed and realistic image in shades of grey and blue, capturing the gritty essence of the mechanical realm.

so far it's not limited by 77 tokens

[vapid crest]

is this a new model?

[jovial tiger]

new image checkpoint, but more importantly, throws CLIP out the window and uses a real llm instead.

[cyan shoal]

wtf how do you have that sampler

[jovial tiger]

there's an extra samplers node in comfy

[cyan shoal]

thanks

plugin or builtin

ah plugin

[cyan shoal]

yeah it keeps adding letterboxes for some reason

[cyan shoal]

thanks I found it, gonna try it out

[jovial tiger]

sometimes it's better, sometimes not.

it's one of the few samplers that seems to work with this pixart thing though.

and looks better than the default euler.

[cyan shoal]

yeah I tried a bunch of samplers

[jovial tiger]

i'm getting good results at 30 steps with res_m, difficult prompts look better at 50.

[cyan shoal]

wow

[nimble mason]

Res momentumized is the most interesting sampler I've found so far and it's not even close

That doesn't mean "best" but in many cases it is

[cyan shoal]

how have I lived without these lol

lets see

[jovial tiger]

yeah, i get great results with 20 steps dpmpp_2m with another 20 0.5 denoise for most stuff. but if you don't care about how long things take, then it can be better than the usual higher quality ones like dpmpp_sde_*

[cyan shoal]

just gotta wait for T5 to load in first

just a couple hours needed

[jovial tiger]

hah yeah. for the first image to load form nothing, t5 takes minutes to load into system ram.

[cyan shoal]

well that's one downside for SD3 already 🤔

I mean its not generation speed, but still

might tick people off

[nimble mason]

minutes??

[jovial tiger]

yeah...

[nimble mason]

took about 5 seconds for me

[jovial tiger]

but once it's loaded, then generations after that are quick.

what size t5 are you using? the one i got off the recommended site is 20 gigs.

[nimble mason]

idk

it's broken into two files

T5v1.1

[jovial tiger]

yeah, 2x 10 gig for me

[nimble mason]

pytorch_model-00001-of-00002.bin

yeah, about that

[jovial tiger]

well, once it's cached it's fast.

[nimble mason]

are you loading off a HDD?

[jovial tiger]

nvme top of the line alienware.

it's not a drive speed thing, it's a processing thing

[nimble mason]

wtf

[jovial tiger]

oh

ya know what.

[nimble mason]

yeah it's seriously just a few seconds for me

[jovial tiger]

it's probably doing an md5 hash the first time it's loading it.

[nimble mason]

ahhhhhh

that would explain things lol

[jovial tiger]

I'm doing it across 3 different machines.

so i'm going through that first initial load 3x.

[nimble mason]

are you just using the standard sdxl vae?

[cyan shoal]

yes

[jovial tiger]

i'm using theirs, but I tried both and I can't see a difference

[nimble mason]

link to theirs by chance?

maybe they're the same? idk

[jovial tiger]

https://github.com/PixArt-alpha/PixArt-sigma?tab=readme-ov-file

they actually mention sdxl vae...

so i think it's the same

[nimble mason]

ahh yeah flat out says sdxl vae so i bet it's the same file

yeah

[jovial tiger]

300 max token length!

[nimble mason]

what's cmp sdxlvaewhatever.safetensors and theirs give you?

niiiiiice

[jovial tiger]

? I don't understand what you wrote there. 🙂

[cyan shoal]

nice prompt

[nimble mason]

oh, the linux command

[jovial tiger]

cheeto man is going down.

[nimble mason]

wasn't sure if you have that on your system or not

[jovial tiger]

lol

it's just the regular fp16 fixed vae.

[nimble mason]

i have wsl running on mine so i use that sometimes with the chaos of SD resulting in lots of models from different sources with different names being the same giant file

k cool

[jovial tiger]

I think i was just renaming stuff to make sure i knew it was the new one.

[nimble mason]

a pink frog sitting on top of a green cat

[jovial tiger]

clearly a bald green cat.

[cyan shoal]

exactly

[nimble mason]

lol

even has lil cat ears

hey, this is exciting

i'm glad you or whoever noticed sigma was released brought it up

that had completely fallen off my radar

[jovial tiger]

ok you found the one prompt that breaks it

[nimble mason]

leave it to the clown

[cyan shoal]

yeah it was obviously not trained on text as much

[nimble mason]

to make the very first prompt test break it

[cyan shoal]

not even close

[jovial tiger]

yeah text is worthless with it

[cyan shoal]

its not like ELLA is any better

[jovial tiger]

ok, it just needed a vertical aspect ratio for the frog

[nimble mason]

so now here's the other q... what's the compatibility situation like with loras and controlnets? i'm guessing zero? and how hard would that be to address

based on the tests you've shared, it certainly seems this is worth a closer look by the community

[jovial tiger]

0.78 ratio works best

I tried just loading a regular checkpoint with this t5 thing and of course, no go.

[nimble mason]

so what about a mech punching a hole in a building?

[jovial tiger]

I'm trying to smush these 2 together, but they won't go. 🙂

[nimble mason]

yeah i'd imagine the architecture is different

other thing too: when we're talking about prompt adherence, res can be a problem, i think in part cuz the schedulers we have are usually too aggressive with the sigma schedule

[cyan shoal]

you guys love this game too, right?

[vapid crest]

damn havent seen it in years

[nimble mason]

looks burnt with cfg = 6

[jovial tiger]

way better than sdxl, but not like ideogram which is the last image.

I'm using cfg 5.5

[vapid crest]

so i tried ultimate SD upscale, but it gives me 4 different images instead

[nimble mason]

wrong tile size for one

if you're using sdxl, tile = 1024

or some other native sdxl resolution, i usually use whatever my latent size was originally

[vapid crest]

my empty latent is 512x512 though

[nimble mason]

which is also the wrong size

you want 1024x1024 as your default

sdxl wasn't trained on 512

there's some resolutions for sdxl

[junior sky]

Ouch

Changed the prompt a bit to: a pink frog sitting on the head of a green cat and ELLA gave me this

[nimble mason]

that last one looks like lora fuel lol

[jovial tiger]

side view of an anthropomorphic muscular green cat is pulling a wagon along a sidewalk on a residential street. There is a smiling anthropomorphic pink frog wearing a racing helmet in the wagon.

[nimble mason]

does it understand left/right/top/bottom?

[jovial tiger]

This is where fine tunes come in. Ella's ability to use existing fine tuned models is a pretty big plus.

[jovial tiger]

I have a command for it working on gremlin

[nimble mason]

whoa, check this out...

a race car driving on the left side of the freeway against traffic in detroit during a thunderstorm

that is the left side, or appears to be for that image

a lil mushy looking

[jovial tiger]

So the answer is yes, but takes some seeds and there's some subject bleed, so it might take a bunch of seeds before you get a perfect one

[nimble mason]

ooo great timing, the readme was updated with some great info https://github.com/Extraltodeus/sigmas_tools_and_the_golden_scheduler

looks like ass (messing around with schedulers now) but hey, left side, and i'm pretty sure that's against traffic

that is definitely against traffic

i wonder what kind of noise sigma was trained on? the usual shit, or pyramid?

[jovial tiger]

Hah looks good.

[nimble mason]

still need to figure out what scheduler/sampler/noise type works well for this obv

but some really good signs already for prompt understanding

also, iirc one thing pixart was throwing around was that their models would be more trainable...?

it can actually do rain... most sdxl models do the effect of rain but don't show it streaking through the air

effect on a surface that is

[jovial tiger]

Whimsical hand-painted watercolors: Vividly depict a cheerful red cat, its fur raising in the gentle breeze, perched to the right beside a serene blue frog atop a dainty mushroom, with a dreamy forest backdrop of soft pastel hues and gentle lighting, creating a delightful and peaceful scene.

Huh. The image prompt adherence went way up when put through prompt expansion first

[nimble mason]

wrong side, but who cares, great image

is that the prompt that went into T5?

or the one that went into your LLM

[jovial tiger]

It's right, not your right. 🙂

My llm

[nimble mason]

what's the expanded prompt?

[jovial tiger]

What I pasted

What's your prompt for the race car?

[nimble mason]

oh k

a race car driving on the left side of the freeway against traffic in detroit during a thunderstorm

[jovial tiger]

I'll try it through this

Detroit rainstorm, nighttime, dramatic lighting. A sleek race car speeds on the left side of the soaked freeway, defying traffic with its brilliant red body aglow, towering city skyscrapers beyond, creating a breathtaking, high-speed silhouette.

@nimble mason

Looks awesome

[nimble mason]

nice, nice

[cyan shoal]

also try cfg rescaling at around 0.8

[jovial tiger]

I need to change up my command to do llm expansion instead of raw. Looks like it really benefits

[nimble mason]

now can we get it to show traffic on the other side too? view of the freeway from a bit farther back

yeah

have it spit out the expanded prompt too when it generates so we can learn from what it understands and what it doesn't

[nimble mason]

the rescale node in comfy?

RescaleCFG?

[cyan shoal]

yes

[nimble mason]

def helps with the burnt look

[cyan shoal]

you can lower to like 0.7

the higher it is, the blurrier it gets

[junior sky]

Who needs round exhaust pipes anyway? 🤷 How much vram does it need? Just want to know if i should bother looking at it.

[nimble mason]

dpmpp_2s_a and karras

that's uniform noise... this is gaussian

pyramid... yuck

power noise

tried setting the t5 type to fp16 and to load via gpu... pow, comfy crashed

[cyan shoal]

https://www.youtube.com/watch?v=mQSKoAEaIJA

@jovial tiger FINALLY

[nimble mason]

@jovial tiger you said you were using t5 with fp16...?

supreme/exp

res/exp

all exponential scheduler with gaussian noise: dpmpp_3m_sde_gpu, dpmpp_2s_a, dpmpp_2m

[clever oar]

new forza?

[nimble mason]

i remember it didn't take very long for someone to publish a finetune on civitai with alpha... i don't reumember how big of a diff there was, but i have it on my HDD

[jovial tiger]

yeah, way better with expanded prompts

[nimble mason]

In a cinematic, high-contrast noir-style digital painting, a scene unfolds on a stormy night in Detroit where a sleek, aerodynamic race car hurtles down the left side of a rain-slicked freeway. The car, a masterpiece of engineering, is painted a deep, glossy black, accented with stripes of iridescent silver that catch the intermittent light from the storm above. Its headlights slice through the heavy downpour, casting eerie beams that reflect off the wet asphalt and the rain-drenched vehicles it narrowly avoids. The oncoming traffic, a mélange of startled drivers in mundane sedans and trucks, flash their headlights in confusion and alarm. Overhead, the sky is a tumultuous canvas of rolling dark clouds and sudden, jagged flashes of lightning, illuminating the scene in brief, dramatic bursts. Each lightning strike highlights the car’s aggressive motion against the natural flow, emphasizing the danger and chaos of its path. The surrounding environment is a blur of towering billboards advertising local Detroit haunts and neon signs flickering spasmodically, struggling against the storm.

[jovial tiger]

wow

[nimble mason]

that's 167 words... chatgpt4 expanded prompt

yeah

[jovial tiger]

awesome

[nimble mason]

we're gonna have a lot of fun with this 😄

hope there's a way to train controlnets for it

[jovial tiger]

haha jesus, just the first minute of this sd3 video has me blown away. he flashes insane images real fast by the screen, every one is amazing

[nimble mason]

if this is anything like it looks ilke so far i'd gladly pony up for some h100 time if needed

oh really

[clever oar]

what you test?

[jovial tiger]

https://www.youtube.com/watch?v=mQSKoAEaIJA

[nimble mason]

what aret hese tools

[clever oar]

its free for all?

[nimble mason]

my earbuds batteries died and my wired headphones busted so i don't have sound right now

all via discord?

i remember emad saying comfyui wolud be getting an upgrade and/or new tools

[junior sky]

He is showing 4 minutes of a bot Chanel that nobody of us has access to. I feel like he wasted my time with that.

[nimble mason]

pretty annoying tbh that not one regular on their official SD discord has access to their SD3 discord bot, lol

[jovial tiger]

his first SD3 prompt, on pixart-sigma: a wide lens cinematic rear shot of a young male dressed in futuristic minmal brown and dark green sci-fi armor and ragged brown cape overlooking a high cliff, looking down at a large army of desert warriors

[cyan shoal]

idk its still weird to me how inferior these preview images sometimes look compared to lykon's images

admittedly, lykon did use highresfix

[junior sky]

Today i test SD3: A cat

[cyan shoal]

so it does improve image quality a lot

[jovial tiger]

bird's eye view of a legion of angry shouting Spartan warrior batmans armed with shields and speers. chaos, debris, confusion, anger, blood, gritty, dirty, mid-action, god rays, yellow smoke,

[nimble mason]

yeah upscaling def isn't working like it does with sdxl

guess we do need to wait there

[jovial tiger]

sd3 hands seem pretty borked.

[nimble mason]

unless tiling does something

[jovial tiger]

a full body character design of a female puppeteer, short blonde hair, modern streetwear clothing of white jacket, black shirt, and tattered distressed dark blue jeans, alexander mcqueen fashion, arms raised in manipulating fashion, various futuristic sleek androids of different sizes being controlled by her, background workshop with different synthetic organs floating in large tube containers

[clever oar]

what is it sd3?

[jovial tiger]

that's another sd3 prompt

sd3 did it better, but the hands in his video examples were even worse

[nimble mason]

are you still using res or are you using anything else differently?

[jovial tiger]

res. 50 steps. all the other samplers came out very muddy for me

"steps": 50,
"cfg": 5.5,
"sampler_name": "res_momentumized",
"scheduler": "karras",

[nimble mason]

huh, i've found res to be muddier so far than ancestral dpmpp_2s_a

a woman standing in a kitchen clasping her hands together behind her back

legit first time i've seen any model do this

not even held together, but still

they're BEHIND not beside

[jovial tiger]

how many steps and scheduler for 2s_a?

[nimble mason]

just karras with defaults and 50 steps

oof, hands

[jovial tiger]

ok i just did side by side and the composition of the 2s_ancestral was better

both were clear

running a set of 3 with 2s now

[nimble mason]

dpmpp_2s_a, supreme with dynamic stepping, res

ancestral with exponential

[jovial tiger]

I'm starting to think some of this is just seed based.

both are sharp, but every now and then a random seed will be more blurry/muddy than the others.

[nimble mason]

ahh

[jovial tiger]

man, watching this video he touches on safety, saying that if someone can do an image of a large container ship crashing into a bridge, that would be bad and effectively should be banned. rage at the clouds for people who think like this. intentionally nerfing models.

that's why i'll never get a robot punching a building with sd3.

[nimble mason]

Captured in a soft, watercolor-style portrait, a woman gazes directly at the viewer with a gentle smile. Her hands are clasped behind her back, concealed by the flowing fabric of her floral summer dress. The light wash of colors and the fluid brush strokes accentuate her calm demeanor and the subtle twist of her body, suggesting a casual, yet thoughtful stance. The delicate play of light and shadow around her form subtly alludes to the hidden gesture of her hands, adding a touch of mystery to her relaxed pose.

[nimble mason]

yeah fuck that

are we going to ban photoshop then? cuz i sure as hell could photoshop that. jeez

[nimble mason]

ideogram doesn't think this way, i believe

[jovial tiger]

it's literally in their terms of service "we won't restrict what images people want to make" and they legit don't.

[nimble mason]

while on here we're told any amount of blood, or even just a cake made out of meat is too violent/disturbing (despite being at worst, PG-13 imagery, maybe even PG)

ideogram spits out prompts that talk about cannibals and gore and bicycles made out of "human meat and bones"

[jovial tiger]

hah

yeah i did the sd3 monster stabbing a rat, and it did it

so far every sd3 prompt i throw at this pixart, it's doing a really good job

sd3 is better, but it certainly better than ella as far as image quality

sd3 doesn't seem to have the ability to put things in certain places if it's just one subject.

only relative to other objects.

[nimble mason]

really wants to do this

[jovial tiger]

Another sd3 prompt in pixart-sigma: top down wide camera angle aerial rear view of a kpop male adventurer assassin wearing dark techwear fashion in the style of alexander mcqueen with white and teal accents, flowing robes and hood, in a dynamic upside down falling pose holding on to the railing of a sci-fi futuristic greco-roman space elevator, over a huge sprawling aerial city in the shape of a lotus petal surrounded by water on all sides, a mega structure of a towering babel-like tower space elevator in teh center reaching into the heavens, falling downward in the dusk sky during golden hour, split toning, sunset dusk, obscured by clouds, atmospheric perspective, in the style of painterly ink

[nimble mason]

wow

not upside down but that's a pretty tough ask

[jovial tiger]

neither were the sd3 shots.. none were upside down

[nimble mason]

are you using these settings too

Illustrated in the style of a modern graphic novel, a race car is dramatically rendered in bold, angular strokes as it navigates against traffic under a thunderous sky on a Detroit freeway. The artwork is characterized by stark contrasts between the dark, ominous sky and the bright, artificial lights from the car and surrounding traffic. The race car, depicted in hues of fiery red and jet black, cuts through the scene with a palpable sense of urgency, its lines sharp and aggressive. Rain slashes across the panels in jagged lines, adding to the sense of speed and danger. Oncoming cars are simplified into geometric shapes, their headlights glaring against the night, adding to the overall tension. The background features high-rise buildings and overpasses, drawn in exaggerated perspectives to enhance the depth and chaos of the urban environment. Lightning forks across the sky in stark white flashes, illuminating the scene in brief, dramatic moments that highlight the reckless bravery of the race car driver.

cfg = 6 here

[shy eagle]

pixart-sigma

[nimble mason]

and CFG 5

all exponential

karras, ancestral, cfg=5, 50 steps

In a photorealistic style, a race car depicted in sharp detail drives the wrong way against traffic on a Detroit freeway during a severe thunderstorm. The car, a model of precision and speed, sports a lustrous red finish with sleek black accents that gleam under the storm’s intermittent illumination. Each raindrop is captured as it pelts the meticulously crafted surface of the car, creating a texture of crystal-like beads that stream across its body. The storm above is a dramatic spectacle, with heavy, roiling clouds unleashing torrents of rain that turn the freeway into a reflective mirror of chaos and motion. The headlights of oncoming cars, a mix of whites and yellows, create a disorienting array of lights that challenge the race car’s daring maneuver. In the background, the cityscape of Detroit looms, its familiar landmarks obscured and muted by the heavy downpour, with only the occasional glow of a distant streetlight or the flashing of a neon sign providing a sense of place and time.

really fn good for a base

back to res for this one

[jovial tiger]

res looks better, but I think that's a seed issue.

[nimble mason]

welp. ak-47

[jovial tiger]

I tried setting mine to auto/auto and now it's using 2 cpu cores and has been for 5 minutes. just siting there.

processing.

pixart-sigma / 2x upscaling with sdxl ai creator checkpoint

0.4 denoise

[nimble mason]

looks great

[nimble mason]

no category on civitai yet for pixart sigma

try that scheduler of mine for refining/upscaling

i was getting pretty good results with that

granted, i did only try res with the settings from the workflow last night

but setting the multiplier at 0.10 or 0.15 or so was pretty good

even 0.05 did a lot to clean up the van gogh nuke image

[jovial tiger]

the 1.5x upscale with 0.5 denoise seems to always been the sweet spot. actually more prompt following since i did say batmans.

[jovial tiger]

I have to run away for a while, but I'll have a look tomorrow.

[nimble mason]

yeah, in general, karras at 0.45 or 0.5, and exponential at 0.5

[crisp stream]

SD 1.5 dreamlike photoreal 2.0 btw

Thumbs up 😄

hand in hand

[languid pebble]

any idea?

[crisp stream]

looks like red text on a grey background 😄 Just had that as well with Clownshark´s workflow, couldn´t solve it so far like you know 🙂

[languid pebble]

Dang ... it worked yesterday 😄

[crisp stream]

comforting...

[pastel root]

Robot love

[languid pebble]

Yeah ... share love! ❤️

[clever oar]

moofi you use dream like sd 1.5?

[crisp stream]

yep, dreamlike photoreal 2.0

[clever oar]

not work for me

[crisp stream]

wdym?

[clever oar]

error

[wispy nest]

Hey guys is there anyone here that i can send a 16:9 photo and they outpaint it to 3440x1440 because i have a amd gpu and that isnt supported by stable diffusion

[clever oar]

i download dream like model but it give me error

[crisp stream]

workflow?

[clever oar]

a1111

its like my gpu not support

[crisp stream]

haven´t tested in A1111, so I couldn´t tell, it´s working in Easy Diffusion + Comfy though

[crisp stream]

I don´t think that´s really the reason

[clever oar]

my gpu not support something in this model

[crisp stream]

Whta GPU do you have?

[clever oar]

1050 ti

[crisp stream]

I had it running on a GTX 1660

[clever oar]

is better

[crisp stream]

nah

[clever oar]

but i cant fix it with any aruments

[crisp stream]

ah, sry, it indeed is

mixed it up

[cyan shoal]

sdxl ai creator checkpoint?

[clever oar]

also standart 1.5 pruned model very creative but with more artefacts

[crisp stream]

yep

[clever oar]

and distortion

sometimes i like that how different result

[crisp stream]

you can always use those for input images on SDXL

And do hires fix along

[clever oar]

I want to try to restore my sleep using a neural network

[jovial tiger]

pixart-sigma: In a chilling apocalyptic vision, a menacing Flying Spaghetti Monster, an ominous shape with eyes on stalks, looms overhead as a dark cloud against deep-hued, storm-filled skies threatening to unleash a deluge of delicious meatballs and tomato sauce upon the diminutive figures below, its body a writhing tangle of pasta, the entire scene illuminated by an otherworldly light that casts long shadows in this macabre vision of armageddon.

[crisp stream]

"pleasant" atmosphere, yet I would work on the face a bit 🙂

@nimble mason

[clever oar]

try restore my dream but sd igore part promt

[crisp stream]

Hedge-hog (slightly shape-edited in PS)

[crisp stream]

i-Gore? 😄

[clever oar]

you like pink color)

[crisp stream]

not in particular, it´s simply the series with the prompt, containing cyan + pink 🙂

[ripe pilot]

Ideogram

[clever oar]

my dream tonight:
from the slightly open door of the house you can see a running man, who is being chased by people on the street among the trees of a dark winter park at night, the lights do not shine

they caught him and started cutting him

💀

saw nightmare tonight 😃

[languid pebble]

Can you recognize him?

[cyan shoal]

https://github.com/pamparamm/sd-perturbed-attention

@jovial tiger

https://www.reddit.com/r/StableDiffusion/comments/1c403p1/perturbedattention_guidance_is_the_real_thing/

[shut sinew]

This isnt true, I can use the bot right now

[jovial tiger]

This isn't true? Was I following you on twitter?

[shut sinew]

Im saying its not true that the bot isnt working right now

[jovial tiger]

I never said the bot isn't working.

[shut sinew]

The tweet does tho

[jovial tiger]

I said it was turned off for some of the original testers, and now it's been opened up to a new set of people

which is true

[shut sinew]

There might be multiple servers idk

[jovial tiger]

Dude, there's only so many major releases I can handle per day. 🙂

[grave scarab]

can someone help me to transform this image

into this

tried so much and i just give up at this point

if anyone can replicate those things, im happy to pay

[jovial tiger]

@shut sinew Feel like trying this one out on SD3? This is what it looks like with pixart-sigma. prompt: Cinematic, low-angle shot of a menacing cyborg shark with sleek, metallic body, glowing red eyes, razor-sharp teeth, and advanced technological enhancements, emerging from the dark, murky waters of a neon-lit lagoon, illuminated by vibrant pink, blue, and green hues reflecting off the rippling surface, casting an eerie glow on the shark's gleaming exterior, as terrified people on jet skis, with panic-stricken faces and flailing limbs, desperately attempt to escape the looming threat, their vehicles leaving trails of churning water and neon reflections in their wake, set against a backdrop of a futuristic, dystopian cityscape with towering skyscrapers and flickering holographic advertisements, all witnessed from a dramatic, underwater perspective.

[topaz harbor]

What do you want transformed about it exactly?

[grave scarab]

the 2nd image is enhanced, a lot of details, bigger boobs, nice face with ADetailer probably, but the details are there in high resolution

[cyan shoal]

doesnt seem to do anything on pixart

[clever oar]

middle

[topaz harbor]

Dm’d you

[languid pebble]

2 years ago ... ... ... 😄

[shadow mortar]

can you help me with the next promt: a photograph of a creature, with the neck and upper body of a giraffe that is retractable, instead of having hind legs it has a large reptilian tail, standing on its two legs and tail, its habitat is the jungle border in the African savannahs, it is grazing, the sun comes from the upper left side.--ar 2:3

[crisp stream]

😄

[shadow mortar]

thank you! i know it doesn´t look as much a giraffe but i had to give it a try

[languid pebble]

Don't ask ... creature deep in the forrest 😄

[dense bison]

Can you help me with the next prompt. It generates a realistic photo about a creature, which is an impressive blend of divine and earthly elements. It has a large humanoid body with imposing musculature, crowned with curved buffalo horns that radiates a sense of power in a majestic enchanted forest. His large, majestic wings gracefully reflect the celestial light as they unfold. Masterfully, in her hands she controls fire and makes flames dance at will in the place where ancient trees and shadows move to the beat of ancestral magic. Her intense and penetrating gaze reveals the wisdom of ancestral beings, embodying strength, magic and majesty -- ar 3:2.

[crisp stream]

@languid pebble @nimble mason After updating everything for making your file work, I cannot get IPAdapter+ working anymore. This is what it shows in the shell:

[nimble mason]

they changed the name of the opt 🙂

just click on the weight