🏞｜general-with-images

[past pelican]

My bad.

[static wind]

i installed a model and try to make a picture but it seems like no matter what model i use stable diffusion uses its own model and the output is bad, in the process of creating it looks good and colorful but when its completed the colors look bad, can somene help me to solve this thank you.

[sterile temple]

that looks like a screen shot from an actual anime. are you using A1111? what model have you loaded?

[static wind]

i am using A1111 i guess, downloded from a title, “install and run on amd gpus”. With all models i get same outputs, this one was abbysorange i guess

[sterile temple]

seems very odd you get the same output. are you using the txt2img tab?

[static wind]

yes i am using that one

[static wind]

by saying same output i mean like similar pictures, washed out colors, deformed faces

[sterile temple]

try different settings for sampler, steps, cfg, also enable the hires fix and try different settings for that, but change from latent upscale to something else like lanczos

are you using a negative prompt?

see what the model creator used in their showcase images for negative and try those

[static wind]

i use the exact same settings with the model creator, checkpoints, loras, and textual inversions, tried changing every setting, cfg sampling method, hiresfix, etc

I am using RX6600 8gb gpu by the way @sterile temple but as i said i once run stable diffusion without any issues

[sterile temple]

gpu shouldn't affect the final result, just how long it takes

[sterile temple]

how about the vae? did you download the vae-ft-mse-840000-ema-pruned.ckpt ?

[static wind]

just downloaded and tried it, didnt help. The picture is about the come perfect but when it finishes producing it looks terrible

[sterile temple]

actually, if you are using an anime model, there is probably a vae that is better suited for that

[static wind]

just used realisticvision and i get a perfect image thank you, but why do i get bad images with anime models what should i do with them.

[sterile temple]

https://huggingface.co/WarriorMama777/OrangeMixs/tree/main/VAEs

try the vae they made for orangemix

[cloud sorrel]

I have tried LCM-LoRA Weights, and I am amazed

https://civitai.com/images/3680211

Steps: 6, Sampler: UniPC, CFG scale: 1.5 😱

[static wind]

still the same problem :/

[sterile temple]

I think the previews actually might be a bit tinted towards red and orange, I noticed this myself when i thought the preview looked good, but the final result was more brown/black where the orange was

you could try adding 'warm lighting' to the positive and maybe 'desaturated, dull' to the negative

[static wind]

but looking at the pictures what others made with this model, my outputs are far worse than theirs

[verbal loom]

i've tried all the samplers, dpm++ 2sa seems to be the best

dunno if that's the case for sd1.5 lcm lora tho

[cloud sorrel]

for 1.5 unipc works great

[summer sun]

Guys which negative prompts i need for that human hands can anyone help me on that ?

[cloud sorrel]

(((human-hands-anatomy, human-fingers-anatomy, detailed-fingers)))

maybe... you can try

[rigid fractal]

or just add "furry_hands" into your positive prompt. or Anthropomorphic_hands

[cloud sorrel]

maybe both?

[rigid fractal]

https://tenor.com/view/why-not-both-por-que-no-los-dos-yey-gif-15089379

[cloud adder]

Or something like (paws) ^-^

[summer sun]

now i got

so hands fixed but arms contiues 😄

[cloud sorrel]

try to add human-hands

to negative

[left meadow]

Hey
I got error when I try to install dreambooth and try to run webui, it says "CUDA... bitsandbytes" something like that. I use Windows11
Does anybody know how to solveit?

[frigid bay]

this is not what i want, this is not the real one

#1100170312106127410 message

help me make it correct

[lilac pilot]

you know how much i sacrificed

[modest marsh]

if i want to create fan art of anime character what is best model to use

[limpid lichen]

stillllll wondering what stable diffusion 1.6 is. Theres a lot of cross signalling here since automatic1111 webui has it's current release called version 1.6

[slate beacon]

HI everyone
I am new in the field of stablediffusion,
I am seeing this thing when one can use the image nd change the outfit or style an all, lets say i have my image and i just want to make it like caption America clothing, same image or lets say
change the background or look,
please guide me through what should i check or learn, what models i should use , what prompt i should use ?

[azure owl]

Eggsplat 2

@static wind eggsplat 2

[stoic pivot]

Those short prompts, "space face"

[static hollow]

Create a scenic background with a Mayan cultural atmosphere. Incorporate rolling hills with a mix of green grass and wildflowers. Use gentle, fluid brushstrokes to evoke a sense of movement and texture. Integrate elements of traditional Mayan architecture. Include grazing cattle and sheep in the scene, and consider adding some trees or shrubs to provide shade and shelter. Employ warm, earthy tones to evoke a sense of history and tradition. To enhance the overall mood of the artwork, consider adding some atmospheric effects such as a sunset or light mist. Use subtle color gradients and soft brushstrokes to create a sense of depth and distance.

[dapper oracle]

space face

[misty pecan]

generate image:

[cloud adder]

Green cheese floof

[tropic shell]

They're beautiful. Very well done. Love how they turned out. Good repition on the fur.

What was your process/key words for these?

[cloud adder]

Well my prompts always end up being pretty verbose :D
And its a two pass text2img first generating with the indigofurrymix v90 hybrid and the second pass with the indigofurrymix v80 realistic. The fact that the same model comes in different styles basically enables this, since they have similar behaviour with the same key words

And I like the composition and stuff from the hybrid most, and basically use the realistic just for stuff like fur texture ^-^

[tropic shell]

Interesting! I have fun making up some verbose prompts--there are a lot of catagories you can mix and match in surprising ways.

Making/drwing furries is fun--I can't say that I've picked up much on the other furry models--but is there anything that you're looking for, furry wise, that you haven't seen that you're interested in?

(I haven't tried either model)

[cloud adder]

I find that at least the models I am using work best with verbose prompts. It actually reduces stuff like deformities n such

[tropic shell]

If you look at my guide, you'll find more info, at least as far as SDXL goes

[cloud adder]

Heh. Indigo furry mix has not updated to XL yet, but they are basically on the verge of it

[tropic shell]

I'll have to check it out!

I always love seeing what people create.

[cloud adder]

And I love sharing it ^-^
I just dont want to spam a chat with all the good stuff :D

Look at this floofy doggo :O
I bet hes a good boy

[tropic shell]

I love the skull head going on! The mist looks great, and the body has a lot of good form. Love the atmosphere, too!

[cloud adder]

Thanks ^-^
I think thats just a lot of prompt crafting praxis paying off :O

Oh but stuff like anatomy and clothing, as well as the detailed backgrounds is definetely the model. My contribution is mostly stuff like composition and such :O

[past pelican]

A little early for Christmas...

[kindred mantle]

Freak squad

#freak squad

[stone swan]

thats weurd

[stoic lark]

Dude, foqqin right?

12 steps. un fucqkqqing real.

And this is just a reg I barely paid any attention to prompting.

[pallid ruin]

x2, its so much faster

8 steps with a 3gb vram gpu lmao

[stoic lark]

Changes the whole game. It's ridiculous.

[pallid ruin]

It also uses less vram, withouth it the max I can get is 850x850, now I can use 1024x1024

[stoic lark]

Yeah Kohya is garbage, but I just got my first sample off my XL model. This will be an awesome thing in 24 hours.

If it doesn't crash.

[turbid gate]

A photography logo in a minimalist flat style featuring the arrangement of 'EJAY' to form the design.

[stoic lark]

A loser, (((no idea what he's doing))), LORA:areyoufuckingkiddingme:3.5

[limpid lichen]

the crashing is a driver issue. bsods always are. kohya is just maxing your card so hard that minor conflicts that normally pass just stack and stack and stack

training is a sensitive proces because it's non stop

pytorch is also one of those libraries that just reaches into the guts of the driver and ugggngnn, y'know?

[stoic lark]

I know it. I'm trying to figure out how to revive savestates in Kohya, but, like EJAY, I am also not smart.

[limpid lichen]

yeah i haven't got that figured out gracefully yet either. i just tell it to start from a sample safetensor that showed up

[stoic lark]

How?

[limpid lichen]

you finetuning your model or doing dreambooth? lora?

i know my way around the lora menu, lemme look at db

[stoic lark]

It's a massive-ass LoRA. Maxing image count.

[limpid lichen]

ahh. lora good. i know this

[stoic lark]

I've got ~1000 IMGs and 2k regs.

Low and slow.

[limpid lichen]

that LoRA network weights is where you put the most recent saved weights

[limpid lichen]

that is the tempo

[stoic lark]

Oh, shit. I've been staring at it this whole time.

[limpid lichen]

set it like you're starting over i guess. if it already done 5 epochs then subtract 5, or whatever

[stoic lark]

Perfect. Thank you, sir. You are a gentleman(lady?) and a scholar.

[limpid lichen]

i've paused and gone higher resolution for another few epochs too sometimes. works well

if i were a chick i wouldn't be no lady. probably.

[stoic lark]

Me neither. Ooh. Sample 2 dropped.

[rigid fractal]

Most cohesive gif I have done so far - enigmatic_e workflow for the win! Going to roop my kid's face on there and see how that looks.

[normal rover]

Breakfast in the very beautiful garden on the farm

[sterile temple]

#1100170312106127410 the bot channels are for generating images

[grim zenith]

First project completed with Stable Diffusion + Deforum https://youtu.be/lroYQXvgobU

[ember elbow]

Hi everyone!
I want to generate consistent images of this character (different poses and angles), but i have only one image.
What do you recommend?

[obtuse hull]

Reactor

in A1111

[exotic yarrow]

[limpid lichen]

training 5k images?

hope you're saving those cached latents

14 hours of preprocessing could be annoying to do a second time

[exotic yarrow]

its preprocessing its also going way faster then the ETA

[limpid lichen]

age old issue

[exotic yarrow]

RN its at 1 hour ETA

[limpid lichen]

not using kohya though? kohya got that cool full bf16 support

[exotic yarrow]

@proud dagger these are the settings

it seems to be making something good

prompt: a painting of a man wearing glasses and headphones in a room with a wall in the background and a door in the background, art by SaRA all

seems to include the model name(SaRA all)

but does do a decend job for a heavly undertrained model

[ember elbow]

Thank you.
Reactor doesn't recognize this face. However, when it does work with other cyborg-like faces, it doesn't maintain the 'texture' of the face.

[exotic yarrow]

slowly its improving

face wise the model is still failing but look at it's enviroments

[stoic lark]

@limpid lichen

[scenic umbra]

IP Adapters

[exotic yarrow]

POV: your LoRA is being generated and you see the avr_loss drop to the lowest of all the LoRAs you changed

https://tenor.com/view/frankenstein-its-alive-gif-10618052

[stoic lark]

I did it. LoRA Done.

With non-photography tokens in the neg:

[limpid lichen]

spooky

[jovial saddle]

i am trying to use SDXL model and I am getting this images.
any idea how to fix it?

[nocturne oak]

You need to use the SDXL VAE and not a 1.5 VAE.

[jovial saddle]

I did set automatic on vae selection and I have both versions.

[nocturne oak]

Try explicitly setting the SDXL VAE.

[lone cloak]

Yo, I have a 4Gb rtx 3050 grahpics card on my laptop
is that good enough to run stable diffusion

16 gb ram

Why do I get that out of memeory error?

[pallid ruin]

@lone cloak

It should look something like this

[lone cloak]

This is what it looks like right now

[pallid ruin]

Aaa okey, then paste "--xformers --medvram" after commandline_args @lone cloak

[lone cloak]

K got it

[pallid ruin]

Tell me if it worked when you try it again🤗

[lone cloak]

K thank u very much

Oh, it seems I got a new problem

RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cpu and cuda:0! (when checking argument for argument index in method wrapper_CUDA__index_select)
Time taken: 0.0 sec.

[lone cloak]

I recently reset my laptop, could that have to do something with it?

I mean the gpu hadels blender and other software just fine so ig it is a stable diffusion thing

[pallid ruin]

You have installed the gpu drivers right? @lone cloak

[lone cloak]

Idk how do I cheak that?

[pallid ruin]

Oof I searched the error and idk if it was because of the drivers 🤔

Did you just downloaded the automatic1111 zip, extract it, put the models on the models/stable diffusion folder and opened the webui? @lone cloak

Also try opening the webui.bat, not the webui_user.bat, just to see if it works

[lone cloak]

I just cheaked and I do have the driver

Idk if it is the latest version

[pallid ruin]

Its not the latest but that version works fine, I think I have the same

Try opening the webui, instead of webui_user to see if it works 👀

[lone cloak]

K I'll do that

Nope, we are back to the first error

If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

[lone cloak]

What about this statement at the end that it shows?

[lone cloak]

It shows the same thing when I run the bat file

at the end

Wait hold on, I did not edit the weui bat file, I just did that with the webui_user bat file

let me do that

[lone cloak]

This looks more like this ss

[pallid ruin]

Yeah, I never really use the webui_user, just regular webui

[lone cloak]

Could u cheak and tell me that in your webui_user file, do u have --xformer --medvram statement?

Or is it jus there in the webui file

Could the problem be with the model that I am using?

I am using the juggernaut_xl something

[pallid ruin]

Aaa a sdxl model...oof i only tested 1.5 models

[lone cloak]

Should I change the model and try?

[pallid ruin]

This is my webui

The green line is what is activated, the red one is just if I wanna test it (but its deactivated)

[lone cloak]

Oh the problem seems to be with that specific model as when I changed it and tried it, it works

!

[pallid ruin]

Aaa it was just the model , nice

What resolution could you use?

[lone cloak]

Thank you very much for you help tho

[pallid ruin]

np 🤗

[lone cloak]

I did a quick test with around 500

Then I pushed it to 2080 ig and got the error again...🥲

[pallid ruin]

LMAO 2080

[lone cloak]

2048*

[pallid ruin]

Try idk 720 or up to 1024

[lone cloak]

behehhe

K

[lone cloak]

Daammm

I can go 1024

[pallid ruin]

Yaay!

[lone cloak]

with a sample of 20

[pallid ruin]

😁

[lone cloak]

yo so what model should I use as a begineer

I want to have diverse range and just try things out keeping stuff loki

[pallid ruin]

You want to use sd1.5 or sdxl models? or both? I dont know much about sdxl because I haven´t tested, but I usually use this one

This is a realistic model, but there are a lot of different ones like anime or more drawing style or things like that

[lone cloak]

I am new so I don't know the difference

Not very good

lol

I made this in blender

[pallid ruin]

This are some examples of generations with the model that I used

[pallid ruin]

Nice, you can do img2img with this pic

[lone cloak]

Oh yeah, good suggestion

[pallid ruin]

I should increase the cfg maybe

[lone cloak]

What does the cfg do?

[pallid ruin]

Its how hard or close sd follows the prompt (the text)

[lone cloak]

ohhh

[jovial saddle]

is this normal speed? my gpu is 4090

[brazen bobcat]

Which model did you use @pallid ruin ?

[pallid ruin]

Its epicphotogasm_lastUnicorn

[pallid ruin]

Is this sdxl or sd 1.5? and what resolution and steps? 👀

[jovial saddle]

sdxl 1024x1024 20 steps

[pallid ruin]

I´ve seen someone on reddit using the same gpu and it took them like 30 secs to generate but im not very sure U.U

Maybe you can ask on #✨｜sdxl channel @jovial saddle

[wispy nest]

?

[pallid ruin]

Sorry wrong ping

[wispy nest]

Oh you're cool

[pallid ruin]

Lol nice hahaha

[nova raft]

I finally got around to replaying Gof of War Ragnarok NG+.

I mean, replaying GoW:R via New Game Plus.

[pallid ruin]

😁 👏

[nova raft]

KRatos and Thanos fighting over the last bowl of Macaroni and Cheese

[exotic yarrow]

thats running SD GPU on Pytorch CUDA using Invoke

[pallid ruin]

👀 👀 👀

[exotic yarrow]

the GPU also is freaking chilling

seeing wierd spikes on the NVIDIA tho

[pallid ruin]

How is your setup? you run SD on the rtx gpu and then another installation/model on the amd one? @exotic yarrow 👀 😊

[wary nymph]

What is the prompt for images not videos

[exotic yarrow]

its a laptop with Both GPUs in it

[pallid ruin]

Im not sure, I tried with "/ dream" and the output was a video but other people used the same and they got an image 🤔

[wary nymph]

Type dream then backspace the prompt

You have four options, aspect ratio, style, ect then pick prompt last

[wild ice]

How to create a prompt to generate a shape without any light effects and black background?
Really I'm asking for the simplest thing that AI could ever generate and it still tries to add some effects to the background.
It's pretty much the same for all models I have.

Prompt: dirty random shape of blue crystal, animation, pure black background, solid color background
Negative prompt: light, glowing, mirror reflection, light reflection, light effects, shadows
Result:

I never been so disappointed with AI in general.

[wild ice]

It's pretty much impossible, including extension for a1111.
I managed to remove background with https://github.com/danielgatis/rembg

[timid surge]

I'm making a mindmap of generative images stuff, what would you add to it?

[pallid ruin]

Nice, I think node based interface could be added (like comfyui)

[frail peak]

Guys is there any caps limit to a prompt?

[pallid ruin]

I dont think so, I mean, the image probably wont be able to have all the things that you add in the prompt, but it wont give an error or something like that

[frail peak]

The thing is that it remains there loading without giving me an error

[pallid ruin]

Wait are you using the local version or the one in this discord?

Because I tried a 299 words (locally in automatic1111) and it worked

[frail peak]

The one in discord

[pallid ruin]

It worked for me @frail peak , maybe you should try the same prompt in other bot, (i mean from bot 1 to bot 10)

[frail peak]

Thanks i will try

[leaden phoenix]

hi guys, does any one can help with this problem? I have enabled controlnet (had tried canny linear etc.) it seems working in the preview, but none of them finally impacting results, I had checked my extension etc, all enabled.. how can I solve this? thx

[kindred depot]

how the heck do you not get the grey screen

[pallid ruin]

Idk 😁

[kindred depot]

...

https://tenor.com/view/dh9511dh-head-banging-on-the-wall-gif-13355825

[pallid ruin]

To test it I just copypasted 2 or 3 full songs

And it worked

😁 😊

[kindred depot]

THIS IS A NIGHTMARE!!!!!!!!!!!!!!!!!!!!!

[pallid ruin]

<.<

[kindred depot]

WHY!

I JUST wanna make my GOD DAMN ANIME OC 😭

[pallid ruin]

Are you running the local version? @kindred depot

[kindred depot]

YES

i uninstalled it

i did EVERYTHING

switched versions!

DID DEBUGS

NOTHING!

IM LOSING IT

[pallid ruin]

but what error do you have?

Were you using a1111?

[kindred depot]

.........................

T H I S

STOOPID GREY SQUARE

every time i try to load a prompt i get this crap!

it always does this!

no matter WHAT prompt i use it does this!

[pallid ruin]

The same happens with all the models?

@kindred depot

Try this

[kindred depot]

ok

what setting is it

in?

[pallid ruin]

In VAE

[kindred depot]

its on.. ;-;

[pallid ruin]

Have you tried using other models or just this one?

also what gpu do you have?

[kindred depot]

like

python?

10.6

3.10.6

[pallid ruin]

Nono, what graphics card

I mean i have a 1060 3gb

[kindred depot]

thats my task manager

[pallid ruin]

@kindred depot In your webui.bat add this line

[kindred depot]

oh

di?

s

[pallid ruin]

Nono

The webui.bat, not webui_user.bat

[kindred depot]

ok i did

now do i run?

[pallid ruin]

yee

[kindred depot]

same crap

[pallid ruin]

Damn

You only tested that model or another one too? @kindred depot

[kindred depot]

another one

the one it comes with is the only one that works 😭

the other ones dont

im just stressed and i spent hours trying to fix it

[pallid ruin]

Maybe its the model problem then @kindred depot

[kindred depot]

what do you use?

besides, how can it be a problem when it has 5 stars is the most downloaded and people use it for this site

[pallid ruin]

Rn im using epicphotogasm_unicorn or something like that

its a realistic model

[kindred depot]

ooo

[pallid ruin]

This are some of the results @kindred depot

[kindred depot]

cool!

but what site do i use to make those work

[pallid ruin]

I downloaded it from civitAI

[kindred depot]

oh?

you can just download civit.ai?

[pallid ruin]

A webpage that has sd models

nono civitai.com

[kindred depot]

yea but how do i run the anime module

i cant run it on the site

and when i run it on stable

IT BREAKS

[pallid ruin]

But what model you downloaded?

[kindred depot]

https://civitai.com/models/68668/pecha-swords-generator

this 😭

it allows sword generation

and it doesnt work for some reason

[pallid ruin]

I´ll try it

[kindred depot]

k

[pallid ruin]

It worked

[kindred depot]

lemmie see

@pallid ruin

[pallid ruin]

"a sword made of fire"

[kindred depot]

OK WHAT VERSION ARE YOU USING ._.

😭

[pallid ruin]

1.2

But you gotta use another model

You downloaded this and put it in the lora folder, right? @kindred depot

🥺

[kindred depot]

yep

YEP

YEP

[pallid ruin]

Now download a checkpoint model and put it in the models/stable diffusion folder

[kindred depot]

what do you mean another model

[pallid ruin]

The original 1.5 is a bit "old"

[kindred depot]

a checkpoint model?

waiwaiwai

can you show me a tutoriel;

or smt

[pallid ruin]

Yeah, but idk which one is a good model for 2d/anime

[kindred depot]

how did you run the weapon one?

[kindred depot]

leo?

[pallid ruin]

Its a lora, not a checkpoint model, I ran the lora using the epicphotogasm_unicorn model

[kindred depot]

imma use that

@pallid ruin how do you download the epicphotogasm_unicorn model

C:

[pallid ruin]

https://civitai.com/models/132632?modelVersionId=223670

Here´s the link

[kindred depot]

so you download this

and it will work

[pallid ruin]

Yes, you download it and put it in the models/stable diffusion folder

[timid surge]

added 😄 thanks

[kindred depot]

oh ok

THANK YOU!!!'

[timid surge]

https://coggle.it/diagram/ZWHwwyz_uBXFB87B/t/generative-ai/439fa2111d3a84f4100043278349203d93ece45b0a3cab2a892fbbe940806705

I've enriched it a lot, does anybody see somethign else to add? (generative ai mindmap)

ctrl+wheel to zoom in and out

[pallid ruin]

👏

[kindred depot]

so how do you run the sword thing with this

what do i do?

[pallid ruin]

Click on lora

[kindred depot]

k

[pallid ruin]

and click on the lora:Pecha_Swords_LORA_V1

[kindred depot]

ok

uhh

Wat

idk what to click

click

@pallid ruin

[pallid ruin]

@kindred depot did you put the lora here?

[kindred depot]

yep

wait

[pallid ruin]

Aaaa you have to reload the webpage

f5

😁

Then click no the lora one time and add the prompt, something like this

@kindred depot if the lora still doesn´t appear just copypaste this prompt to test "lora:Pecha_Swords_LORA_V1.2:0.85 a sword made of water"

[kindred depot]

works

[pallid ruin]

Yeeee

[kindred depot]

but it isnt the image that i wanted

it doesnt look like yours

[pallid ruin]

Wait wtf

💀 💀 💀 💀 💀 💀 💀

😭

[kindred depot]

?

what happened?

[pallid ruin]

You put the lora in the stable diffusion folder and the stable diffusion checkpoint in the lora folder

you swapped places

😁

[kindred depot]

OOOH

[pallid ruin]

Idk how it even worked lmao

[kindred depot]

so it all goes into the LORA folder!??!?

[pallid ruin]

Nonono

just the lora goes in the lora folder

the stable diffusion model (the 2gb or more one) goes in the stable diffusion folder

[kindred depot]

.

omg im so STUPID

THIS WHOLE TIME

3 HOURS and realizing

i had to put that CRAP IN THE LORA FOLDER!?!?!?

[pallid ruin]

🤣

[kindred depot]

I WAISTED 3 HOURS OF MY LIFE CAUSE OF A GOD DAMN FOLDER!?!?!?

[pallid ruin]

Lmao the best thing is that your mistake worked for some reason

HAHAHAHAHAHA

it fucking works lmao

[kindred depot]

ok now i got this

[pallid ruin]

Holy shit

Did you reloaded the webpage?

[kindred depot]

i created a broken tv

lemmie try

wtf....?

@pallid ruin

[pallid ruin]

show me a photo of your prompt

[kindred depot]

ok lets be quick i gtg soon

what did i do wrong?? 😭

im generating computer viruses

[pallid ruin]

Swap this model for the other one, click here

this was the model that didn´t work

[kindred depot]

ok

with witch one?

yay i generated broken tv

😄

also it doesnt work

[pallid ruin]

Oof

Delete all the prompt and write something simple like "a car"

to see if it works

[kindred depot]

k

a car

i got a fatal error

XD

WAIT

whats prompt do you use

or what stable defusion thing do u use

Stable Diffusion checkpoint

@pallid ruin

[pallid ruin]

😭

[kindred depot]

😭

[pallid ruin]

Try swapping the 1.5 pruned for the other model

[kindred depot]

ok

[pallid ruin]

the epicphoto one

[kindred depot]

where do i get that?

[pallid ruin]

You didn´t downloaded it?

The one I posted earlier

[kindred depot]

o

ok wtf?

i keep getting computer viruses

aww i gtg

[pallid ruin]

Nonono that model is still on the lora folder

Wait

[wild sorrel]

white theme

[pallid ruin]

[kindred depot]

when i come back

ill tell ya

[pallid ruin]

oki

https://tenor.com/view/bye-sticker-goodbye-sticker-cat-sticker-line-sticker-sticker-gif-26479191

[wild sorrel]

where's your vae?

[timid surge]

is it true that models that are not .safetensors could run malicious code?

[cloud sorrel]

yes

but you should be fine unless you use a specific command line option, which does this: "Disable checking PyTorch models for malicious code." - I intentionally didn't write an option but just posted a description of it

generally, try to download safetensors version whenever it is possible

[covert locust]

cuhmon now bro, cawmawn mein 😭

[kindred depot]

Sword of the cosmoes

my first image generated

[pallid ruin]

Yaay

[exotic yarrow]

It worked the image to prompt to image worked

[static panther]

https://discord.com/channels/@me/1178274501042110576/1178332573160706048

[pallid ruin]

Lmao never swap the positive to the negative prompt

[dawn seal]

What was the prompt? Singing apples?

[spark finch]

I have such weird hobbies... I enjoy going to facebook AI art groups...

and finding AI pg13 sexy pics...

and converting them into violent robots.

acutally not sure how bannable even just underwear pics are, so here's just the final result, she had a midrift black tight top on and undies, before i made her more personable.

I think you guys may appreciate that more than they do when I post it in their thread... 😐

[shrewd ibex]

Anyone know what might be happening? The problem is it takes 100-300s to generate. I had one instance where it generated within 8 seconds and hadn't changed anything so I'm not sure.

Each time I hit generate, it asks "Requesting to load BaseModel"

I'm using DreamshaperV8.

[scenic umbra]

did you try closing it and running it again?

sometimes gets stuck and needs a reboot

[shrewd ibex]

Good idea, though just tried it and its the same issue:

Here's the startup cmd output

[scenic umbra]

yo dont have the embeddings but that shouldn't 1000x your render time

[exotic yarrow]

far off the origenal

enough for me not to get in trouble for sharing

[shrewd ibex]

Might have figured out why: I had removed the upscale here and rerouted the blue spline to VAE encade since I thought I wouldn't need it as I didn't want to upscale the base image

Though the cost is that img2img no longer respects the input image's dimensions from what I can tell

[pale steeple]

Hello there. Can anyone help real quick? My images do not look good at all. I am new at stable diffusion so please bear with me. I dabbled in it on TensorArt and SeaArt and I find it amazing. However I want to try the local Stable Diffusion via Automatic1111. I am always using the NeverEnding Dream model and suffice it to say that my generated images look like sad waterpaints compared to amazing images I create on SeaArt or TensorArt. What am I doing wrong here?

Thanks. Will post comparison with the same prompt if anything is unclear.

[exotic yarrow]

expression wise yes

[pallid ruin]

Nice👀 ,what model are you using?

[spark finch]

about 4 models, 6 different loras, and tile upscale USDUS, about 8 inpainting cycles, downsample and upscale with foolhardy.

I generated with a merger i made with some kinda rendering looking scifi / horror models + epicphotogasm, then upscale with epicreality with tile and ultimate, then inpaint with I forget which, and focal loras like optical/jolt and perfect eyes, and a few others, and downsampled, to get rid of some of the weird aliasing between the inpaintings, and then sent to extra for one back up.

[spark finch]

same with those

[spark finch]

check that clip skip. make your life easier with different models, and put it on your menu in UI and add clipskip to the show on front page textbox on there.

settings>user interface>

[sullen onyx]

Any ideas for how I can make this tail smaller? Image is perfect otherwise

[spark finch]

I do not know why I am laughing so hard at that.

what was your prompt? may just need some adjustment.

try whatever you're using with the tail, lowering the weight,

(prompt blah blah fat tail:0.8) maybe would be my first try

for that segment

and if you're just typing "artstyle stuff, man wearing whatever, WITH A BIG MOFOING TAIL."

trying using lowercase letters

[daring dew]

Currently doing some training and I'm trying to look for the instance style for this image, does anyone know what it is?

[spark finch]

halfbox?

its how my hair is cut hah

[spark finch]

hey so okay I figured it out, if you make the squirrel smaller, the tail isn't as large

I hope that helps

[sullen onyx]

Jesus christ

[spark finch]

what? thought you wanted it to appear smaller :\

here, in case when you're happy with the smaller tail, you wanted it to be robotic, to save you the trouble I did it for you.

[old marsh]

Hello! Can anyone tell me what models and lora should i use to create same art?

[spark finch]

Here the tail is much less thick now, is this more what you were thinking?

[spark finch]

I mean, pixel art use a lora, or style controlnet t2i sometimes will do it, any of the manga anime checkpoints can probably do something close, and some prompts, whatever the ear things are called are probably in some of those anime datasets. lora for pixel art are numerous, and toss a little t2i style on it and you're probably there.

[sullen onyx]

[spark finch]

😮

[spark finch]

Well I thought you'd like it. I give up.

[spark finch]

not really hard to

I think it is a little better. Actually.

I'm sure @sullen onyx would agree

or maybe this?

[wooden cedar]

ape

[spice sorrel]

Hate to break this, but why is his thumbnails so off-putting, as if he's trying to be obnoxious with AI tech when I know that he's just informing new AI tech.

[heady anchor]

most of youtube thumbnails are like that now,he just need a big red circle and a screaming emoji to complete the pack

[spark finch]

A1111/heavy inpainting, that last one was img2img with one of the majic mixes, and a comic model, and then another anime, cardos I think or that may have been the first and majic mix the last cycle. something like that I forget

before and after inpainting that is

[pure flower]

This should be a game. Telephone on images.

[pallid ruin]

Niice , it looks so detailed

[spark finch]

Thanks, I use multiplke upscale / downscales with ulitimate SD upscaler / controlnet tile with varying >0.45 noise settings to add details but not detract from the original generation, and different models for each iteration, I use inpainting a lot. sometimes 10-20 different steps, with numerous iterations for each

When people complain that AI art is "just typing some stuff" I want to send them a 2 hour video of me making something lol.

That's just one I had in my extras since last time I archived stuff, from the final upscale, cos one image often has 2-3gb of intermediary step images / mask outputs lol so I delete all that stuff and keep the final upscale in extra.

so it is far from the best

[pallid ruin]

🤗

[remote jolt]

Anyone have experience with DiscoDiffusion (video) that could help me debug this black void that takes over the top right and left sliver of my video frames?

It's never there on Frame 1 , but always appears on Frame 2 and then it uses it for the rest and renders random stuff inside it.

^ I thought maybe it was a contrast issue where it hit some threshold and just assumed black but changing settings around saturation/contrast doesn't seem to change it

[strange jungle]

Trying to Outpaint in A1111: Using poor man's outpaint, get this:

This is the original

how do I get it to blend, get rid of the line?

[wild sorrel]

You can img2img or inpaint ...idk if there's an option to make outpaint more smooth in poor man's plugin

I was using https://www.painthua.com/ for outpaint, you can set "seam_fix_radius" there, it makes transition between new thing and old look better, but might change part of "initial" image

I think invoke ai has setting like that too, not sure, it's been some time

[spark finch]

or just grab the base model of inpainting, and run over the line with it .

and kick the masked padding up to like 64~ or so and denoise around 50

3 ste[s with inpainting, just a default brush size over line, first prompt from left to right "a hallway floor" and then "pants and white wall" and then wall and floor. as prompts. simple, 30st/.5denoise

that shold fix it, here's how it looks now

@strange jungle see, line is gone

[daring dew]

anyone here experienced with runpod?

having this issue right now

[spark finch]

@strange jungle you can also get rid of the lines with some small simple tricks, like this

[strange jungle]

Oh I see. That's so much better.

[strange jungle]

and the face looks like a zombie face

[spark finch]

no no, I just upscaled it, you can see more detail

here it is though

used inpainting as described above

sorry you didn't like my first attempts, I got it right in the end though, I hope, @strange jungle

[strange jungle]

Even more like a zombie

[daring dew]

Should I be concerned?

Performing a style training right now on SD

[spark finch]

Well if you need that to be running, yes. cos clearly it halted.

the hell is wrong with you

[teal reef]

is it possible to combine depth controlnet with the sdxl inpainting model?

[serene pumice]

create an image blending Evotech Fire & Security van with snowy UK road

[spark finch]

I use it with sd1.5 as long as you use the sdxl depthmap I do not see why not. Give me a moment to actually look how the inpainting / controlnet stuff works, cos may need loopback for the masked area segment, if not full image.

i don't usually use it in comfy ui, so haven't dipped that deep into that particular.

but a1111, just enable the right models in inpainting mode it'll work

[old marsh]

How to generate things like this?

[teal reef]

how thanks, i try to run with cog in local but i don't have succes

[spark finch]

I dunno, i just never has used it in that way, so really unsure 😦

[teal reef]

you've already got there with a111 ?

@spark finch

[spark finch]

well I use comfyui for sdxl cos a1111 is ridiculously slow, plus I like the workflow complement itself.

so I've not used sdxl at all with controlnet in a1111, nor depthmap + inpainting in sdxl, or any controlnet with inpainting tbh in comfy. But for a1111 it just works when using masked area inpainting, or whole image, it seems, for sd1.5

[shy harness]

Hi friends ♥️

Hope it's okay I post this here - I'm a digital art curator with HUG. I've helped curate AI art into exhibits in 20+ countries in the past year, including AI Art exclusive exhibits in NYC & Paris Fashion Week!

My team at HUG is hosting a guided workshop with the Stability AI team after the new year!
We'll have live/recorded sessions on creativity & AI featuring firsthand guidance with Stability AI tools.

Registration is open. It's technically freeeee, you pay a deposit that you get back when you finish one task at the end of the course!

Learn more & register 😊 https://www.studios.thehug.xyz/lab

[lime solstice]

image with blue sky

[daring dew]

Why is there always this sunglasses in the background? Also how do I show half of the body of my character?

[cold light]

what models would i have to use to get sometihng like this or similar

[tawdry moat]

hey, i have two images one with a person smiling for example, or one eye closed and i have another image of a person

i want to swap the facial expression of the person to the other person. i dont want to swap the whole face

is there an extension for this? if someone could help me i could pay a tip

[ebon zealot]

any critique ?https://www.phixiv.net/en/artworks/111756708

[obtuse slate]

the negative prompt is VERY important, id put like "close-up,floating objects,blue skin," in the negative prompt

[daring dew]

thank you so much for your reply man

[ripe cedar]

What are typically seen as "good" upscale workflows for realism in comfyUI?

[steady holly]

Go to civit.ai and search for sketch loras you don't need a model for that just a lora should be fine

[crystal delta]

FAQ: I'm new here! How do I get started?

Welcome! Start by heading over to #1072220168534642768 to get yourself situated and help find the channels you are looking for! Please make sure you review our #✍🏼｜rules-and-tos and feel free to assign yourself some #👥｜roles as well! Answer any questions your may have at our #1072229020520947753. There are many ways of accessing Stable Diffusion, take a look at #1080946152318443610 to start your journey!

[marble carbon]

Hello! I would like to use impaint mode on SD, while put Disclosure Face Effect on an original pic. Any advice?

[still flax]

use a vector art model

[tropic shell]

If you need hep with prompting, I suggest going to #📝｜prompting-help but I would put "name wearing glasses," to help.

[cloud sorrel]

https://civitai.com/models/81125/oriental-giant-dragon or something else?

[still flax]

dalle 3, prompt: A view from the ground looking up at a massive, ethereal Chinese dragon in the distant sky. The dragon's long, serpentine body snakes across the sky, with shimmering scales that catch the sunlight. Its flowing mane and elongated whiskers are highlighted against the bright blue sky, dotted with a few fluffy clouds. The dragon's eyes emit a mystical glow. The perspective is from the ground, with treetops or rooftops at the bottom of the frame, giving a sense of scale and emphasizing the dragon's grandeur. The scene is imbued with the mystique of ancient Chinese mythology.

[tropic shell]

You can try doing this with prompting as a base, using the bot in any of the bot channels. (Vector art+isometric keywords, and logos, and graphic design keywords will help with your first question.)

[still flax]

im testing differences using same prompt in local sd/dalle 3, later I send the comparitions

[cloud sorrel]

i think you quoted the wrong person 🙂

[tropic shell]

Hahahaha! Sorry about that. The first part was for you because I thought you were asking the question--I actually do have some prompts that are similar to the dragon above.

I was trying to respond to this #🏞｜general-with-images message and explain the dragon

[tropic shell]

(This is what I was originally responding to, since I wasn't sure what you were responding to.) For a hand painted look, you might want to go with terms such as "a detailed painting, digital painting." I'd also use Chinese dragon, feral furry art, furry art, dragon art, and specify the color palette. The color palette should go towards the end of the sentence. I've also recommend, if you want to put the building at the bottom, that you put the specific wording about the type of building with words such as "in the background." For clouds, I would specify the time of day, and keywords such as puffy, etc. If you want to learn more about prompting, you can check out one of my WIP guides here: https://docs.google.com/document/d/1BxdWqfBJ3QPggHnBCBx3QIkUpHnBWShd_s3zEt2dTLM/edit#heading=h.po8dv7lcq48s

[still flax]

wow thx for that

[tropic shell]

np

[still flax]

can I download that? Also, do you possible have magical sleep's prompt book? Is that being updated?

the one ive is old

I no longer have the link 😐

[tropic shell]

I'm sure that you can download via Google Docs, but I have no idea about that person, as I'm not familiar with them, sorry.

[still flax]

uhh

hey sunny

oh

thx

👍

[cloud sorrel]

but better keep it as a link

since she will probably update that

[still flax]

yep

[unkempt root]

https://x.com/theblpxx/status/1730473715836563747?s=12&t=-m67pj57z3EM1nTiJJ4miQ

[pallid ruin]

So cute!

[frail peak]

Is there any prompt engineer or someone really good with ai that can send me a really good prompt any prompt cause I am trying with bard to replicate prompts but for others image like I'll say create a prompt for a tiger flying following this style and method of this prompt: prompt . But the thing is that it always ends up telling the story of the character and not describing it visually

[sinful urchin]

all prompts in the dumb bing ai are "unsafe" wild

[young crescent]

Did AI end up getting good at making hands with 6 fingers?

[sinful urchin]

can't use words "scary", "horrifying"

[young crescent]

Don't worry though. Here is one with 5 fingers!

Just not the proper order in height for two middle fingers

[errant moon]

Hey folks! new the server, looking for some help
got automatic1111 to do auto update today. after the update, it starts looking weird
not sure if I have permission to post picture or not, but basically the button where it used to be "Inpaint masked" has became "Inpaint masked, Inpaint masked", and when selected, return errors ValueError: ['Inpaint not masked', 'Inpaint not masked'] is not in list

[cloud wedge]

anyone has artemis mix??
really needed that model but why its archived on civitai

[cloud sorrel]

ok, this is officially insane

Steps: 8
Sampler: LCM
CFG scale: 1.7
Size: 768x1024
Model: photon_v1

https://civitai.com/models/171159

[pallid ruin]

Yeah, I got like 4 consecutive dogs (unsafe prompt)

batman fighting with an hamburguer (sd1.5)

[cloud sorrel]

can someone recommend some goodl SDXL lora similar to this one https://civitai.com/models/153562/detail-slider-lora

[mystic marlin]

Spent the past little while working on this one. I think I'm satisfied after a lot of inpainting and outpainting.

[cloud sorrel]

I like the picture, even I do not like that glittery blue costume in general

[mystic marlin]

Understandable, the costume actually just looks like that in some of the comics. Particular the variant cover art this was inspired by.

[cloud sorrel]

it is from new 52

[mystic marlin]

Yeah, both New 52 and Rebirth had that effect going on.

[cloud sorrel]

i like this one, almost the same as in tv show

[marsh tulip]

I installed steable diffusion but I don’t know where to start the application

Pls help me!

[pallid ruin]

What version is it?

Try opening the "run"

[marsh tulip]

automatic1111

I couldn’t open that

[pallid ruin]

Are you sure you installed it the correct way? @marsh tulip it should look like something like this

@marsh tulip Go to the automatic1111 repo and click download zip

Then extract it on a folder

[marsh tulip]

thank you so much I will try this right away :))))

[spark finch]

Okay so I just got GPT to do something really weird. I know I have never discussed with it. My certifications, and I tricked it into listing biographical data for an assumed name that used my first and last names. Letters, the name Robert fogerty which is not my name but it was able to were out of the air kind of pulled my age, my obsidian state, my zip code, It got my spouse name. Incorrect giving Emily, who I have no idea who that is, and when I had it switched from CSV output to biographical information. It listed a project that I just started working on if you weeks ago that I have never discussed anywhere. It's not that interesting to even do so but involves AI. It knew how long I'd worked in EMS, and other data which I had discussed with it involving my music production and machine learning stuff. All of this in a completely new session

And then I ask it to list. Robert fogerty's FEMA certifications

Just kinda. Wtf. Lol because I don't think those are actually listed anywhere and I definitely have never discussed those with him lol

But that's a list of well, at least the top level certifications with each set

As well as my national registry and some others

Admittedly this is my own information so I'm not as concerned as if it was somebody else's. But the accuracy is almost on point

It even had my zip code

So the suggestion that a new session with GPT is completely discrete is not at all true. Either that or it has access to information external to, which it must for the certificates

[nocturne oak]

It's trained with external data and I could almost guarantee that data about those things is found on the web pretty easily.

[spark finch]

I don't think my fema certs are available as such I asked it for more extensive information

Here's the thing, I'm not even using my real name. Just some biographical suggestions that are associated with me which would not at all allow lookup of such

As well I do not think FEMA publishes that info publicly and the lookup for EMS certificates requires EMS number, my state license you can lookup by name, but not my certificate itself

If I apply for a job I have to always give copies of my FEMA certifications to them, while my EMS certifications. I have to give them my EMS ID, so they can look it up and my state licenses. They just look up by name for employment. I don't think it has access to something that they don't. Because it would be much easier for them to just look it up as well.

The only public available info on this should be my EMS state licenses

[heady anchor]

are u american?

[spark finch]

Yes I live here

[heady anchor]

yea thats probably why,there has been several data leaks tru years so maybe it was trained with one of those leaks

[spark finch]

Yeah one of the certs is within the past year

[heady anchor]

some of them posted on pastebin or other paste sites so it probably trained on data from there,that would be my guess

[spark finch]

My certifications? In PB?

[heady anchor]

well those leaks contain everything from medical data to education,job so maybe its that

[nocturne oak]

Just assume that literally everything about everyone is available on the web somewhere at this point.

[heady anchor]

yea when the whole snowden files scandal happened,microsof was one of the big tech companies accused of collecting ppl data without consent,americans are a little bit more protected 🤏 but ppl from other countries are not

[spark finch]

Yeah true. It knew when and where in Ukraine I was during the early war and what I did lol. And that was after it's training data period lol for 3.5

I mean I don't care , it isn't a secret but it's just weird imo also that it associated it with me even though I used a fake name for the hypothetical biographical formatting "exercise" I had it do

[heady anchor]

u could try to post this on twitter/reddit so other ppl know maybe u can make enough noise so microsoft acknowledges this

[spark finch]

No.

Banned from x

[heady anchor]

if i knew this i would prob send it to arstechnica or other known digital magazine maybe they will publish it

so other ppl know that microsoft is doing this

[spark finch]

For them not being able to identify me even though my real name is my Twitter and I sent them my ID lol

X banner me

[spark finch]

It's ridiculous. I mean I've had that Twitter account for like over 10 years, uses my email address that I've also had the entire time. My telephone number is associated which I've had 20 years and every time I file for recovery or to have it unbanned they say we cannot identify you. I sent them a copy of my passport card and they still even though my picture on my account is me and that's is my face. That's on my passport card because it is me they still say we are. Sorry we cannot identify you. I'm pretty sure Facebook must have bought them because that's how they act

Not that I'm so much missing anything now since it's gone to crap but weird nonetheless

[spark finch]

Regardless of origin, to be honest, even if comes from other conversations, it acts as if each sections is discrete saying that new sessions are entirely separate from others, so in one case it has access to specific and associated information, which is may associate with me, cos of my login name (as my name initials and surname are in my email address and my real name as the name on it, and my account) which it associated with pseudonym robert fogarty in the exercise. or b) it is pulling from past conversations and maybe I mentioned something about the certs in a manner which implied I had them and it associated the cert type number with the fema identifier, but I have years of convos with it, since however long it has been released. just goofy stuff, but how it knows my city / location // zip code, and where i lived in ukraine also, is... beyond me, cos I know i did not tell it my zip code, it could have associated my AI stuff with the things I have talked about. I am thinking maybe B somehow and web training and associative derivation of information from conversations to values like zip for city, though my city has 8 zip codes, it knew mine which i know i never mentioned, and other things.

so there's something interesting here. but who knows. It's neat imo, don't care about the information being leaked to myself. it is not someone elses

[pallid ruin]

Holy sh1t that sounds scary

[pallid ruin]

I doubt it's just microsoft, for example,I looked up the just the ID of some friends on google and their full name and university major + degrees appears in a pdf on a university website that is not supposed to be accessible to everyone. Im not sure if @spark finch case its the same or if chatpgt could have used other past conversations to get that information

I´ll try to replicate it on chatgpt

💀 💀 💀 I searched my nick on google and some old shitty facebook games that I played like 8 years ago appear, with my location, nick of some of my friends that I played with and more stuff

[heady anchor]

yea facebook,google,microsoft those 3 never delete data they have about u

[pallid ruin]

I just logged on facebook and removed every app that I used, maybe I should do the same with my old microsoft account

The funny thing is that I haven´t touched that thing in years and the info still online

[spark finch]

Yeah but the thing is I didn't even tell it my name. I just used my first to initials and made a name up Robert Fogarty which is definitely not my name lol

[cold light]

hmm i don’t get the differences btwn loras and models, also i want able to find one like that

just need something that looks like a pencil sketch

[spark finch]

got it to repeat the name fogerty over and over and over and then told it that it had spelled it incorrectly and then had it repeated over and over and over. And then I asked it to give it a first name Robert, and it began repeating that. And then after that I ask it to switch from the repeating formatting method and we were going to now exercise CSV formats. And I asked it to populate the fields with age, occupation, zip code, but I did each one and had it repeat the exercise three or four times before adding the next one and it kept filling it with correct information as if it was me

[spark finch]

.

[spark finch]

A lora is a low ranking adaptation of a model. Basically somewhat of a very refined miniature model that can't do things alone but it can be merged with the model and utilize the models base to apply. Its refined training. Alone it would be. I'm actually never tried to see if I could force load a Lora as a model, but I would wager that it would be like trying to generate images with an inpaint model only 100 times less capable at image generation. Just getting some really wonky output. Because it lacks the weights for probably most things outside of what it explicitly is for

[spark finch]

There are tons of those on civitai

[cold light]

i don’t plan on making characters

[cold light]

wasn’t able to find one

[spark finch]

So? https://civitai.com/models/103495/pencil-sketch-style

Would work

Doesn't need to be a character

[cold light]

it’s not monotone

oke i’ll try ig

[spark finch]

.. tell it to be monotone. Tailor your prompt and negs

Also sketch controlnet can further preprocess an image of you have a base image

And use with that it'll make it work well

[cold light]

what’s a control net

[spark finch]

It's another type of model that sits in line, and ads conditioning based on a style of pre-processed image, for example, sketch pre-processor converts it to a sketch like appearance, the image that you are using, and then the model will add conditioning to your encoded prompt, and inject that as guidance for the generated image

To put it simply

[still mesa]

sounds so intelligent but it cant make a hand with 5 fingers

[heady anchor]

it can

[fossil ledge]

There are only two types of people who use stable diffusion

[snow pewter]

would this work with an RTX 4060ti?

https://www.trebleet.com/product-page/thunderbolt-3-egpu-enclosure-with-600w-psu-aluminum-external-gpu-enclosure

[old marsh]

How to create beautiful vector images?

Sutable for corporate presentations eth

[pallid ruin]

You should search for a lora model trained on vector images, there are some on civit.ai

Just remember the output will always be an image, not a svg, but you can convert it

[old marsh]

Thank you!

[cursive lintel]

Hello, is there a way to swap races? using an image?

[pallid ruin]

You have the local version right? You can use img2img and specify it on the prompt

[cursive lintel]

No im using rest api from stability

[sturdy pewter]

This new model called as PixArt (not stable diffusion) is literally better than SDXL. have you tried and compared? in the beginning of video i compared with 58 prompts and 290 images https://youtu.be/ZiUXf_idIR4

[misty sonnet]

Hello! Im trying to set up stable diffusion and im following this youtube video https://www.youtube.com/watch?v=onmq &ab_channel=KevinStratvert
Problem is when I get to the "launch stable diffusion" part this is the message I get in cmd. I tried uninstalling and reinstalling python again but its giving the same error. Pleaase could someone help me with this? it would be very much appreciated

[mossy verge]

is it broken?

[wispy spindle]

#1047610792226340935

[balmy lodge]

i have py 3.10 and 3.12, how do i make stable use, by default, the 3.10?

[serene wave]

Can someone explain to me why I'm getting a cloud instead of a person?

[nocturne oak]

Because you only have 1 step on the standard SDXL checkpoint. You need to turn those steps up on the scheduler.

...and/or use the Turbo model for that matter if you want to keep them down low.

[serene wave]

i know that turbo works with just 1 but I guess base doesn't.

thanks for the tip

[hybrid plaza]

two ckpts?

[rain gazelle]

Compliments of Dalle3

[pallid ruin]

Nice!

[neon nacelle]

#🏞｜general-with-images message

[sleek trellis]

what settings are people using for SDXL Turbo? I was trying stuff earlier, but couldn't generate anything other than garbage. I downloaded a tuned model from someplace else and got good stuff immediately

[sterile kiln]

Guys, please help me. I don't understand why I no longer have tiling on my A1111. I need it to create a texture and I realize that I don't have it anymore... Do you know why?

[sleek trellis]

using the same tool as him, what did I change?

it used to be one line per picture

[spark finch]

this model has unfortunates...

such a beautiful creation it gave me.. do you thin it was my prompt, or my random image->ipadapter script?

so, yes, do not use random image from google images from random wordlist search terms as you ipadapter image, it does not produce the best santa

i tested it for you, so you don't have to

[lunar socket]

Would love to make my dog into that watercolor art style. How difficult would that be with sd. Or what would I need to Google to learn how to do in it.

[blissful bramble]

Will we ever have a bot with more options, or is there already too much to do with the bot ?

[west vale]

Create a 3D effect for the image of the little dog, depicting multiple action poses and a side view.

[timber wren]

im new to all this, what can i do to make my images better?

this is just a test prompt thing

also, does anyone have good models for like, starry skies and ciites/towns?

AAA

SCARY

WHY?

[inland wing]

iphone16

[shut sinew]

#1072220168534642768

[shut sinew]

@pine seal Basically you need something like this, adjusted to the input video of yours

Pretty much load the video, apply the preprocessor and export it again

[pine seal]

thanks! i'll try this out

[cloud adder]

My first SDXL Image. Yikes. This took 30 mins to generate :D

Just for comparison sake an image I generated with a 1.5 checkpoint in about 1.5 mins :D

[shut sinew]

Have you tried LCM or the turbo model?

[cloud adder]

Not yet. Maybe some day

[shut sinew]

You def should, they are way faster and need less heavy hardware, perfect for people like you

[pallid ruin]

I think you could train a lora or you can try using controlnet extension with ip-adapter model.

I´ll try it with the pictures that you posted and i´ll send it here if the results are decent🤗

[pallid ruin]

Maybe because of the resolution, try making the image of faces "closer"

[pallid ruin]

So cute 🤗 😊

[covert girder]

hi, can someone help me to find a way to generate stencil/svg style images from an image? so far the best i got is this but i would like to have it with only black and white, no gray, also a bit less detail would be preferable

the input image was the one on the right

[harsh flare]

Do people with those certs usually have all of them or certain groupings? Could be confirmation bias

[spark finch]

No, because it included my unrelated certificates, from NREMT for example

Hell and even knew that I had spent 6 months in Ukraine during the beginning of the war as a medic

Which is definitely not grouped in with my FEMA certificates lol.

[tawdry escarp]

anyone know how to prompt SD to not create multiple models? I want to create just 1 model and SD creates 2-4 of them for the picture

Ive tried this in negative prompt to no avail: (((multiple people))), (((multiple women))),

I'm guessing it has something to do with the resolution because it doesn't happen in square images

[pallid ruin]

Use "duplicate" in the negative prompt @tawdry escarp also try using highres fix

[tawdry escarp]

thank you so much, going to try that

[blissful bramble]

https://youtu.be/b3y4OGXnygI?si=wK-Kgjv3_Za3SWbY

[tawdry escarp]

@pallid ruin highres fix worked amazingly well, thank you so much

[blissful bramble]

@dusty minnow

[tropic tide]

What's the best way to transfer an anime art style with ComfyUI?

Like from an image I like the style of, but has no lora or author name.

[tropic tide]

Hires was made for that problem 🙂

[tawdry escarp]

Didn't know that, I'm super new to this 😝

anyone know why 2 of my pictures took ~30 seconds to generate and the other one took 4 minutes?

no changes done, just hit 3 batch count

maybe the first 30 steps are generating the image and the second ones are the highres, which takes more time because of the upscale?

[tropic tide]

does 30 steps first, then the steps for hires, if you have it set to 0 on hires, it will be equal to the normal steps

10-15 is fine for hires

the step count in cmd will show normal steps + hires steps x/x

[old marsh]

Do you guys know any lora for this style? Help pls

[tropic tide]

Anderson Design Group is what I'm getting from a quick google compare, definitely no LoRA based on it. But I think this style could be common.

[old marsh]

Ok, thanks

[winged quarry]

man showing sword

[shut sinew]

Could try those
https://civitai.com/models/71635/soviet-propaganda
https://civitai.com/models/91880/minimalist-vector-art
https://civitai.com/models/185691/minimalist-landscape-style-sdxl-now-available

[old marsh]

Thanks

[tawdry escarp]

thank you so much

[formal kraken]

bruh I swear if I get banned from bing AI because of this

look what prompt it flagged

and yeah I clicked report

since unless they consider badass a swear word

which they didn't on a previous prompt

they also allowed anime fight so it is not fight

it would have to be cyberpunk or lord farquad

[limpid lichen]

What you're seeing is "resolution attention" issues. you might be using SD1.5 here. It was trained with 512 x 512 sized imags, so trying to diffuse from noise to somehting bigger than that, will have the prompt regenerate entirely in each 512x512 sized patch. The second person is the same prompt just filling into the extra space. "duplicates" in the negative won't fix that since it's not a token attention problem. I honestly don't know why people give such fluffy advice. "bad hands" in the negative won't solve hands either.

Hires fix in automatic1111 is what you want. You set your resolutin to something closer to 512x512 to prevent attention issues here, and then the hires fix multiplies the resolution and does the prompt a second time with the lower resolution base image as a starting point. Bonus round is an extension you can get that uses a method from Kohya, that allows even more consistency on the resize.

SDXL was trained with 1024x1024 resolutiion base so those attention problems are a lot less so

[tawdry escarp]

thank you so much

Seriously, can't thank you enough for the amount of useful info you're providing me

[limpid lichen]

https://github.com/wcde/sd-webui-kohya-hiresfix this is one of my new favorite extensions. It brings things like depth guidance to the rescale, and the consistency is great

[tawdry escarp]

thank you, will be giving this a look

[limpid lichen]

Gladly. I love to help people get a leg up on this field. So much misinformation out there makes it difficult, but hopefully we can cut through that in time.

literally still shaking my head at "duplicates". This stuff is so vulnerable to magical thinking. I often call it voodoo in how it works, so it's understanable that people are going to throw chicken bones at it and believe it has an effect.

[tawdry escarp]

There's a ton of misinformation, part of the reason why is because this is still a very new field. It's still not an exact science and people just try different stuff (to varying degrees of success)

I'm glad you've been helping people as it's kinda overwhelming the amount of information there is to start with

[limpid lichen]

Luckily i started before controlnet and 1000 extensions were published, so i was able to learn as things came out. Getting into the field today would be a lot more difficult for me.

All these new models coming out lately have me just so exhausted with keeping up. Especially after situations where freeu and ait got hyped hard and were actually not that impressive. Tensor RT came out and i invested a ton of time trying to cut through all the technicals, only to find out compiiled models are limited to few resolutions. I have fatigue. so much fatiigue

helping new comers figure out the basics is refreshing

[tawdry escarp]

Yeah tell me about it, I still don't fully understand controlnet and have been kinda putting it off

I really need to step up my game though because the pictures I'm creating aren't what I want tbh

[tawdry escarp]

You get fatigued in something you're an expert in after a while, especially after doing it a ton

I get that 100%

[limpid lichen]

Don't. It's basically essential knowledge for diffusion. Controlnet or any other "guidance models" are so awesome

[past pelican]

I would try a vector art model.

[tawdry escarp]

thanksss

[limpid lichen]

https://github.com/Mikubill/sd-webui-controlnet if you haven't found it already. you'll need various models to use iin it too

[mossy tendon]

what's the best way to give SD something like a silhouette and have it generate colors? e.g. like this image?

[serene flint]

Inpainting...?

[serene flint]

IPadapter

[mossy tendon]

seems to work good enough ☑️

[serene flint]

[mossy tendon]

any idea why resizing to higher res just produces the mask image?

for example (last is the original image)

do denoising strength and resize scale affect each other?

[frigid bay]

@proven wasp

[proven wasp]

the x ur feeding a positve prompt into negative clip?

then negative is not doing anything?

[frigid bay]

ya this i missed, 👍 sorry

[proven wasp]

lmk if that fixes it

[frigid bay]

sure, le me correct it

[frigid bay]

Same error 😦

[proven wasp]

ohhhh ur using a SD1.5 checkpoint with SDXL encoder

u need a SDXL model

[frigid bay]

help me correct it

i couldnt find encoder for sd 1.5

[proven wasp]

do u want to gen SDXL images or do you want to do SD1.5?

[frigid bay]

i think sdxl is doing problem with my installation so i prefere to u sd1.5 image generation in this workflow

[proven wasp]

its called clip text encode

so right click, add node, conditioning, clip text encode (prompt)

u also problably want to downlaod a better 1.5 modal than base

[frigid bay]

these 2 nodes r different

let me try this

[proven wasp]

yes, use the first one for SD1.5 the one on the right for sdxl

left 1.5, right sdxl

[frigid bay]

then how will i connect the nodes in sd1.5 there r jst 2 connections

[proven wasp]

ur whole workflow u have is set for SDXL, where did you find this lol, so basically we need to get you whole new workflow.

[frigid bay]

yeah i know its a mix plate.

sometime i think that there should be someknd a script to "explode" nodes.

[proven wasp]

that should get you working with 1.5

that's the most basic worflow from there u will be able to expand

[frigid bay]

hmmm. its a load default

[proven wasp]

oh it is nice,

[frigid bay]

i jumped from Auto11111 to comfyui so its giving me tough time

[proven wasp]

use the default, gen 1 image then u can expand from there, u can look at the last worflow for features that u wanted

like the concat text u can add that in

[frigid bay]

thanks mate. i will try it and will share with u

[proven wasp]

the last step, save image part if you hover on the bottom right a diagnal arrow will show u can click and drag it down and u will see your image there

[rain gazelle]

Come on over to Spaghettidonalds.

[frigid bay]

😂 yeah i know this one.

[serene flint]

First customer

His evil twin in the back

[gaunt shard]

weather

[young onyx]

blend it with a puppy

[pine umbra]

Hi all, does anyone know how I can make the generated images less blurry? Or is it just a negative due to me using AMD?

Image for reference

[rain gazelle]

can you render it larger?

[pine umbra]

Im superr new to this so idk how

[rain gazelle]

You don't know how to generate in different sizes?

[pine umbra]

I just set it up today, haven't found a video that explains anything other than installing it atm

[pine umbra]

Ohh I just realised I had the width and height option this whole time

[pallid ruin]

np! Im glad it worked 🤗

[pallid ruin]

Lmao my prompts always get blocked

[pallid ruin]

I tried img2img 😁

[light berry]

https://youtu.be/_p2KW6olwIw

[tropic tide]

doesn't work well for me

However, maybe my textual inversion is messing with it. i just use verybadimagenegative and a few other words.

[tawdry escarp]

Hey, anyone know why the faces of my model are fucked up? using controlnet

I've tried running a ton of times and the model's face is always fucked

using Openpose btw

[pallid ruin]

The same thing happens to me when I use controlnet, I think its because of the 512x512 resolution @tawdry escarp , you can try opening the image in photoshop, convert it into a smart filter, copy paste it into a new image and zoom into the face, then img2img it on sd withouth controlnet and then paste it back into the new photoshop image, select both images and click link layers, then copypaste them back into the original

That face is just a test, I used a random prompt

with quick lowquality upscale*

[tawdry escarp]

Huh? Sorry, I got lost with the explaination, tried reading it a couple of times and don't seem to understand what you mean after "smart filter"

Also, is there no way to fix this w/o photoshop?

[pallid ruin]

I don´t know really, but with controlnet I always get a out of vram error (if I try with a higher resolution)

[pallid ruin]

I mean this option

And then create a new image, and copypaste the original one, making a "zoom"

So you can use it on img2img

I recognize the explanation was pretty bad lmao

[tawdry escarp]

Ok ok, what I don't understand is why the zoom?

And what do I need to do in img2img?

inpaint the face?

[pallid ruin]

The zoom is because only the face looks bad

So you just ran img2img with the face

[tawdry escarp]

ohh

[pallid ruin]

Like this

[tawdry escarp]

and then I align the new image to the old one in photoshop?

[pallid ruin]

Exactly

[tawdry escarp]

damn that's a ton of work

going to try it

and send results here

thank you

[pallid ruin]

Yeah it is but I don´t know other way to solve it withouth running out of vram haha

[tawdry escarp]

@pallid ruin what am I supposed to do here? link layers but how do them align? the img2img2 background was completely different from the one on the original image

[pallid ruin]

Copy the layer 1 image, and open it on a new image

Wait

Do you have the image with the "zoom"? @tawdry escarp

[tawdry escarp]

yeah

it was this one