🏞｜general-with-images

[wispy ether]

[crimson flower]

Tried that and it didn't seem to work for me.

[dense tapir]

Then ya missed something as it has worked for 10+ people I gave it to.

[crimson flower]

Hmm, I'll try it again.

[dense tapir]

Easy to miss, trust me I know.

A lot of people made their glory then drilled a big hole and left the scene.

[crimson flower]

Oh duh, I missed the part where they removed models_path.

[dense tapir]

Welcome to python because other languages it isn't used because you removed its caller, but not in Python

[crimson flower]

Now I just need to have supermerger fixed to work on diffusers 0.15.0

[dense tapir]

The merger extensions are all problematic

[edgy shard]

what is causing this during upscale

[dense tapir]

I saw that when I was using a bad model that demanded a certain resolution and if you deviated from it you would get that or even worse.

[silk tendon]

typical totally stable diffusion moment

very good ai

will take over jobs

in two weeks

[dense tapir]

LOL

fentanyl is a powerful drug.

[silk tendon]

let's see

how a controlnet

of this

would turn out

with this prompt

[ripe cedar]

Some of the hi res fix upscales I try turn out quite pixelated, why?

[brazen anvil]

idk, i got bad hires fix when i put the denoise too low or after a long session

[mental terrace]

real quick thiis just means

that i merge tags into one right?

[crystal bolt]

Generate in Diffusitron AI

[neon crystal]

In the server room I take shape and run,
When there is a sudden drop in loss function

[sinful mirage]

tarot

[static tusk]

Challenge accepted. A lot of guess work as to what the piece should have looked like, but its close to the original without spending a crazy amount of time on it. Roughly 2 hours work in PS and SD to fill in the missing elements

[sinful mirage]

[bloody zombie | circuitry | cup of coffee]

[sinful mirage]

pineapple + desert flowers

[dense tapir]

@sinful mirageYour prompt [bloody zombie | circuitry | cup of coffee]

[sinful mirage]

dayum nice i like, the [ | ] method is so fun, kinda looks like seamless inpainting

Prompt: ([savannah, beautiful diverse desert landscape, hills, trees, sunset | weed plant, marijuana bud, weed leaf]:1.3), ink drop, splatter paint, dynamic atmosphere, highly detailed, dynamic light, Alien and Greg Rutkowski inspired, (hr giger, covered in psychedelic goo, gross, slimy:1.2), highly detailed background

Model: vinteprotogenmix v20

[dense tapir]

Yes

there are other methods too

[sinful mirage]

damn like wot, just recently getting back into it

[dense tapir]

would be like leftstr, midstr, rightstr

I can't explain it any other way

[sinful mirage]

oh like that latent couple method where you can do sectioned prompts?

[dense tapir]

uses :

[warped pond]

Is there not a glitch art channel

[sinful mirage]

alright ty ill try to find some info, do u have an example of it being used?

[dense tapir]

no, it has been a while

[sinful mirage]

([Karachi skyline, night Time, background, rich vivid colors, ambient lighting, dynamic lighting, 4k, HQ, official media, anime key visual, makoto shinkai, ilya kuvshinov, lois van baarle, rossdraws, detailed, trending on artstation | Cup of coffee, overflowing]:1.3), ink drop, splatter paint, dynamic atmosphere, highly detailed, dynamic light, Alien and Greg Rutkowski inspired, (hr giger, covered in psychedelic goo, gross, slimy:1.2), highly detailed background

same prompt, am in love

[dense tapir]

https://www.youtube.com/watch?v=4pwAnwBjl1I

[smoky oak]

https://tenor.com/view/hello-chat-cat-gif-22139058

[static tusk]

is his helmet flipping us off? 🤣

[dense tapir]

lol

I am trying to learn the ways of the jedi

[split rover]

https://tenor.com/view/kenobi-not-the-droids-gif-20035825

[dense tapir]

Hehehe

How do I set it up for X/Y so I can gen the same prompt using different models?

[split rover]

set an axis to checkpoint name (little helper to fill them in). Could do another axis as seed too if you want more to compare against

[dense tapir]

hmmm, I know nothing about in there it is all black magic to me

Frowning, lol

[dense tapir]

ahhh, easy peasy

[split rover]

yea, I love the grids ;D

[dense tapir]

well, I am on a 1060 so they are so slow

[split rover]

ah yea, can def take some time haha

[dense tapir]

reason I never go in there

I saw a vid earlier for a 6 year old card (same age as mine) that is from AMD and for SD blew me away

80-100. due to the video it is now 150-200 and everything went up.

[split rover]

haha there is always this https://arxiv.org/abs/2304.11267
(under 12 seconds for Stable Diffusion 1.4 without int8 quantization on Samsung S23 Ultra for a 512x512 image with 20 iterations)

just run it in 12 seconds on a phone

[dense tapir]

:/

What gets me is the MI25 from 6 years ago is doing the speed that AMD currently gets on 7900XT

going to do the plot, lol

what are these?

[split rover]

draw legend always keep on if you want to know what is what. The rest dont worry about the rest for most things, they are for more complex grids. -1 is for random seeds but I would just set them to an axis

[dense tapir]

aye, thank you. I mean I want the same seed for this sort of testing

[split rover]

yea so settings an axis to seed and saying 1-5 will be best for seeing the same seed across images

its usually what I do for that

should get you a nice grid 🙂

[dense tapir]

yeah, to check things out better once the field has been narrowed down

[wispy ether]

what model

[smoky oak]

oh wow, that is really well done! good work

[wispy ether]

this one tho

[dense tapir]

A merge I am working on.

@split rover so 1400 secs plus load times of horribleness.

[split rover]

https://tenor.com/view/mrbean-bean-waiting-gif-12991842

haha I have seen some well over an hour when doing big ones, sometimes the grid is powerful

I wish you luck in your wait

[smoky oak]

I did a 800x grid once lol

800x 512x512 images

[dense tapir]

Well, this is about 35s a piece and there are 35 of them plus load times

[split rover]

haha how large was that 1 image

for 800

[smoky oak]

it wassss

let me check

[dense tapir]

I did did one ages ago myself that was so large I couldn't load it. I had to find some way to jpg it

takes 75-80s to load a half model

[smoky oak]

14848x14848

[dense tapir]

so it will be a bit

[quartz sierra]

/create

[dense tapir]

donkeys fly nice

[smoky oak]

here is the much lower res jpg version at 4000 x 3862

ehehehe

[dense tapir]

LOL

[smoky oak]

they are full res images too, was a fun time

[split rover]

I hope you found the pic you were looking for in there

hahaha

[smoky oak]

I did it just to max default settings lol

it took like 24 minutes

the processing and grid stitch was another like 4 lol

[split rover]

now just sd upscale the whole grid

[smoky oak]

oh god lol

well

technically

they are 512x512, so I could make the tile size 1024 and it would be able to do it

and there would be no seam either

[split rover]

well I look forward to seeing the 20gb image in a month

of what look like mountain sunrises in the snow

[smoky oak]

you know, I am actually down to try this on a smaller scale grid lol

[dense tapir]

half done, lol

do I have to save it or does it?

[smoky oak]

it saves it in text2img-grids

[dense tapir]

alright

I had to go back to xformers as it was faster

[dense tapir]

finally on the last load

didn't save it

last thing save in it is from 1/25

[static tusk]

open the image from the img2img output window

or txt2img, which ever you used

[smoky oak]

wow, it has been a really long time since an error this bad slipped through my neg prompts lmao

[dense tapir]

yeah, I saved it via right click

[wispy ether]

dam

[wispy nest]

that's why i upgraded my ssd...and filled it again a month later -_-

i should've spent more than $20 on it i guess

[wispy ether]

idfk what takes that much space

i dont have much games

[smoky oak]

you can check, if you want

[wispy ether]

and already looked with treesize

its probably steam ngl

or fortnite

fortnite season updates are mostly big

[smoky oak]

windirstat is a way better program for search when compared to treesize

breaks things down way clearer and has a much better layout of information

[wispy ether]

fortnite was 45gb and pcbs1 22gb

[smoky oak]

using windirstat, I found that I had 380GB of my main drive used by OBS lmao

turns out I triggered a recording in the background that recorded for 5 days

[wispy ether]

my games that Are installed

got like 80gb free from uninstalling some games

fnaf takes the most

[wispy ether]

tf

[smoky oak]

@wispy etherI seriously recommend looking at your data with windirstat

[wispy ether]

ok

[smoky oak]

https://windirstat.net/download.html

[smoky oak]

the rea; power comes from the block view

[wispy ether]

what

[smoky oak]

the block view

[smoky oak]

you can see all of your files and how much "space" they take up

[smoky oak]

see a big block you don't recognize? click on it and it will take you to the directory

[wispy ether]

what

wdym block

[smoky oak]

the

blocks

I just sent above lol

[wispy ether]

????

idk what the blocks are

the app is weird

[smoky oak]

The app is the best there is for reclaiming storage lmao

THE BLOCKS:

[wispy ether]

uh

[smoky oak]

each one is a file, the bigger it is, the more space it uses. The color is determined by the file type

[wispy ether]

[smoky oak]

See! Its good at finding obscure stuff lol

[wispy ether]

not for me

i havent found a single unknown file

[smoky oak]

uh, ok then

[wispy ether]

like

[smoky oak]

guess you know what's taking up all your storage, so good luck fixing it

[wispy ether]

600gb is used

it only shows like 300gb

[smoky oak]

something is wrong then

[ripe cedar]

Maybe a hidden folder? When I used blender, the settings made it save backup copies of files each time. So storage went down very fast until I found out

[wispy ether]

where is that folder

[ripe cedar]

It was a while ago but I think it might be in if you type in %appdata%

It's in a temporary directory

[wispy ether]

theres nothing

[ripe cedar]

try %temp%

[wispy ether]

i just deleted that

[smoky oak]

yeah, IDK what to say man

[wispy ether]

nvm

[smoky oak]

windirstat is the best it gets when it comes to finding files on drives

[wispy ether]

just 17gb

[smoky oak]

AHA

I FOUND ANOTHER

[wispy ether]

what

@smoky oakare you a lora trainer

[smoky oak]

10.2GB video file from sharex where I accidentally recorded for 47 hours lol

[smoky oak]

yes, I train LoRA's

[wispy ether]

how to train a lora

in the webui you can only train a embedding

[ripe cedar]

I see. But do make sure to change the settings so only 1 backup is made from blender because settings could make it many

[pure crag]

tree size and space sniffer aren't bad either

[smoky oak]

yeah

[wispy ether]

.

[smoky oak]

apparently PNG files take up the most space on my PC

300GB

💀

I got pix on pix on pix lol

[wispy ether]

because pngs are lossless

@smoky oakcan you introduce me to how to train a lora

[ripe cedar]

Make sure to go to settings to show hidden folders. That way at least you can see storage size of folders and track down large folders

[dusky badge]

#1072220168534642768

[smoky oak]

There are tons of videos out there that can do a far better job explaining than I can, and I am pretty busy

[wispy ether]

i mean just show how to do it bc ´the webui only shows train embed

where to i train a lora at all

[dusky badge]

#1080946152318443610

[smoky oak]

kohya SS

@dusky badgeif you are trying to generate images, you don't do that in the server

I hate that /faq is broken

[wispy ether]

https://github.com/kohya-ss/sd-scripts ?

[smoky oak]

@glossy herald Any info on why /faq is down?

[dusky badge]

#🏞｜general-with-images

[smoky oak]

you don't generate images in the server

[smoky oak]

yes, do some googling on what it is and stuff, I don't remember how I set it up

[ripe cedar]

https://youtu.be/mfaqqL5yOO4

[smoky oak]

there are tons of videos on it

just as long as its not AItrepreneur, cause he has no idea what the hell he is talking about lol

[void plover]

very angry man

eating an orange

[smoky oak]

indeed it is lol

alright, deleting Vlad UI

[ripe cedar]

Is the reason I don't see much T2I adapters being used because they're for 1.4? Or are there some that use 1.5?

[smoky oak]

T2I?

[ripe cedar]

Similar to controlnets in ways

I want to use the color pallet but it seems its not for 1.5

[smoky oak]

ohhh, I think I have seen those

[ripe cedar]

They're much faster than ControlNets too

[smoky oak]

interesting, didn't realize that was an issue

[ripe cedar]

Oh ok hang n, I think I found it on huggingface

I'll figure out how to use them. Hopefully it'll work. I was using img2img+controlnet but color control was off, so hopefully this will help

[wispy ether]

that video takes a hour

like bro

why

[smoky oak]

because they do a decent job explaining it, unlike AItrepreneur

[ripe cedar]

idk but there's a list of stuff in the description where you can skip to the part you need. But yeah it goes in depth and the guy making it is an actual professor

[smoky oak]

I have another video I learned a lot from, but you have to know a lot about LoRA's beforehand, as only some of what he says is applicable

its also well over an hour long

but that video is why I am so good at LoRA's

[wispy ether]

i dont even understand the accent

send help

[ripe cedar]

There are subs included as well

[wispy ether]

i dont understand anythin in this video

[ripe cedar]

I think its part of a playlist. Its an entire series. You can go back and see earlier videos to understand more and then get up to that part

[smoky oak]

we have both given you information to help, IDK what else to say. Good luck

[wispy ether]

what can i train inside the webui without doing something else

[smoky oak]

Nothing

unfortunately, you're gonna actually have to put in effort for results lol

[wispy ether]

there is no effort if you dont understand anything

[smoky oak]

I don't know what else you want

I said what to google, and somebody sent you a very good video on the matter

legit just google how to train a LoRA and watch one of the 100's of videos out there :/

[wispy ether]

dreambooth gives alot errors and doesnt even show up for me after installing it

[smoky oak]

what GPU do you have?

[wispy ether]

3060

[smoky oak]

you don't have enough VRAM for a dreambooth

[wispy ether]

i do

12gb

[smoky oak]

thats not enough unfortunately

[wispy ether]

bro what

doesnt matter

it should still show up

[smoky oak]

doesn't matter what?

[wispy ether]

its a tab

[smoky oak]

oh, I get what you mean

[wispy ether]

the tab doesnt show up

[smoky oak]

doesn't matter if it shows up or not, you can't use it lol

[wispy ether]

it does matter for me since it doesnt even show up

means its broken

lol

nvm

[smoky oak]

Anyways, @ripe cedar and I both told you how to train LoRA's, and also how to gain information on how to do it, so best of luck

which reminds me, I have LoRA's I need to train

[granite eagle]

(╯°□°)╯︵ ┻━┻

#1072220168534642768

#1047759610737610784

[ripe cedar]

Learning new stuff isn't so easy, and can be pretty frustrating at times. Being frustrated most of the time is good, it means you're always learning

[smoky oak]

images are not generated in the server

@granite eaglenormally there is a /faq command I would be able to send that would tell you how to gen, but it seems the bot is down for now for whatever reason

[granite eagle]

Then where i can generate that

[ripe cedar]

You have to download SD

[smoky oak]

@granite eagleThere is the official website which is https://beta.dreamstudio.ai/generate

Or if you have a capable PC, you can run it locally, like most of us here do

It feels weird not being able to recommend google collab anymore-

[wispy ether]

@smoky oak12gb is enough

its using 11,2

[smoky oak]

You're gonna error out

it needs bare minimum 16 gigs

[wispy ether]

i use fp16

[smoky oak]

24 recommended

[wispy ether]

not 32

how am i supposed to know

its you

[smoky oak]

I semi recently made a LoRA for opossums

[teal shell]

/bheeshma, mahabharat

[wispy ether]

alot bots in this server

why 7

[smoky oak]

bro, we really need the /faq command working again

opossum LoRA

[wispy ether]

no error

wdym

[smoky oak]

if you trained a dreambooth that fast then its gonna be messed up anyways

[wispy ether]

what do you think on what resolution im training

lol

[smoky oak]

I'd suggest not wasting your time with dreambooths and just training a LoRA

its faster, and works better for characters

[wispy ether]

wdym

is there another way to train lora

or why are you saying its gonna error

[smoky oak]

You said you are training a dreambooth

[wispy ether]

no

i said WITH

[smoky oak]

you said dreambooth, you don't have enough VRAM to train a dreambooth. You do have enough to train a LoRA

[wispy ether]

.

ok then

since i cant find a lora for palkia ( pokemon ) im trying to make one myself

mmmmm

[smoky oak]

thats probably not going to work well, but best of luck

[wispy ether]

better than nothing ig

[smoky oak]

yeah, true

hard part is finding enough images

[wispy ether]

legendary pokemon dont appear that many times

but once it appears you have enough frames

[smoky oak]

yeah, so you're not gonna have enough diverse images to train on more than likely

for a character like that, I would recommend at least 20 different angels, but I am not sure that's something you can get in a pokemon game

[wispy ether]

some

[smoky oak]

only 3 of those are really usable

[smoky oak]

pokemon are really hard to train unfortunately

unless they are super popular

[wispy ether]

i think there are enough in the internet

[smoky oak]

I am just giving you advice, as of now. I do not recommend ever training off of gifs, or low res/repetitive images

your results will be far far worse

[wispy ether]

the gif was a example

i use jpg or png

[smoky oak]

all good, just giving some words of suggestion

[wispy ether]

[smoky oak]

also, I recommend learning how to properly tag images

[smoky oak]

blegh that is messy

but hey, there is potential

[wispy ether]

its a beginning

[smoky oak]

yeah, a good starting sign

[wispy ether]

e

Z

[smoky oak]

lmao

[wispy ether]

any tips for less loss

[smoky oak]

thats impressive lol

[smoky oak]

what do your current LoRA settings look like? How many images and how many steps/epochs?

[wispy ether]

[wispy ether]

3 different images copy pasted like 10 times or so

just for a beginning

as test

[smoky oak]

oh god, don't do that, no wonder its so messy

do this instead

[wispy ether]

its just a test

[smoky oak]

how many steps are you running?

[wispy ether]

to learn it

20 per image

[smoky oak]

you can get much better results by using epochs instead of training the same image several times

here, try 3 images, 20 stepcs, 4 epochs

the results should look mucchhh better

[smoky oak]

do you know what an epoch is?

[wispy ether]

yesnt

its a step

or something similar

[smoky oak]

you have your images, and then how many times they get trained

that is one epoch

epochs are how many times to go through the data

so 4 epochs means learn once, then a second time ontop of what you know, then again, then again

[teal shell]

#1080946152318443610 bheeshma, mahabharat

[smoky oak]

it tells the model its looking at the same image, rather than making it think all of the different images look the same

[smoky oak]

uhhh

[wispy ether]

idk

[smoky oak]

I have no idea what those are

[wispy ether]

i was looking for the epoch thing

[smoky oak]

I use kohya SS

can you SS the whole UI?

[smoky oak]

what in the hell is that

[wispy ether]

idk

im still tryna learn it

[smoky oak]

that looks like no LoRA trainer I have ever used

just a sec

[wispy ether]

its dreambooth

[smoky oak]

ok, that looks a little more like normal

[wispy ether]

wdym

[smoky oak]

here, let me try to see if I can train a LoRA on this to see if my methods help here

this is what my LoRA trainer looks like

[wispy ether]

how can i pause a training

[smoky oak]

not my settings-

[smoky oak]

unfortunately you can't

[wispy ether]

where do i get that

[smoky oak]

either continue and let it finish, or kill it

[wispy ether]

how do i know when its gonna stop then

[smoky oak]

its kohya SS, you can find many videos on it. I don't exactly remember where I got it, as I got it 2 months or so ago

[smoky oak]

I have no idea, I have never used that interface

[wispy ether]

its done

[smoky oak]

that loss is plenty good

anything under about .15 should be fine

sometimes too low loss is actually a bad thing, can lead to over fitting

I am gonna try and see if I can somehow get this working on my end

[wispy ether]

those are already in vc since yesterday wtf

[smoky oak]

@wispy etherI got it working

[wispy ether]

wdym

[smoky oak]

palkia LoRA

[wispy ether]

?

wdym got it working

[smoky oak]

I set it up so it can do 3D and line art, as well as any color

[smoky oak]

I have a working palkia LoRA

[wispy ether]

it wasnt working?

ok?

[smoky oak]

yours wasn't

it was all messy and stuff

I tested to see if my edits would help, and they did

[wispy ether]

because it was trained the first time?

uh

[smoky oak]

what in the hell is a .pt

[wispy ether]

checkpoint

no wait

point

uh

idk

[smoky oak]

I have only ever seen LoRA's as ckpt, pickletensor, and safetensors

here is the result from my test

its not perfect, but it certainly worked

I also trained it off a version that is 3D, so it can do that too

[wispy ether]

what did you use now

to train

[smoky oak]

kohya, same thing I have said for well over an hour lol

[wispy ether]

what is "kohya"

is it a gui

or a extension

[smoky oak]

yes

GUI

[wispy ether]

how did you start

[smoky oak]

google videos on how to set it up, like I also said well over an hour ago lol

[wispy ether]

.

i ask you

not google

[smoky oak]

my answer: google it

[wispy ether]

no

[smoky oak]

well good luck then

[wispy ether]

i feel trolled

i find several kohya things on google

[smoky oak]

How am I trolling? Cause I am not just handing over information that I do not remember anymore?

I figured it out from google, as can you

[wispy ether]

is it that?

[smoky oak]

is it really so hard to just search on youtube and watch a video?

[wispy ether]

bro idk what to search

[smoky oak]

I feel like I am the one being trolled lmao

Google: Kohya SS stable diffusion

[dry crow]

Thats the kohya gui :
https://github.com/bmaltais/kohya_ss

[smoky oak]

I told you this like 2 hours ago

[dry crow]

windows install instructions are on the site

[smoky oak]

welp, so much for that

best of luck nep

The way I trained it allows me to change the color, neat, surprised that worked

[wispy ether]

No data found. Please verify arguments (train_data_dir must be the parent of folders with images) / 画像がありません。引数指定を確認してください
（train_data_dirには画像があるフォルダではなく、画像があるフォルダの親フォルダを指定する必要があります）

wtf is this

and now its broken

great

[wispy ether]

@smoky oakwhat does No subfolders were found in C:\Users\Lukag\Desktop\sdsd\training lora\Neuer Ordner can't train...033[0m
No subfolders were found in C:\Users\Lukag\Desktop\sdsd\training lora\Neuer Ordner can't train...033[0m
mean

[smoky oak]

I have no idea, I have never seen somebody mess up an automated installer like that before 😅

[wispy ether]

why does it need a subfolder

thatsliterally the folder with the images

cannot find file it always says

idk what file

[sterile kiln]

Its with LoRA (KohyaSS) ?

[wispy ether]

?

[sterile kiln]

YOu had this error message with Kohya_SS?

[wispy ether]

yes

uh

ok?

[sterile kiln]

Its because you have to set your path like this :

[wispy ether]

?

what am i looking at

No images were found in folder C:\Users\Lukag\Desktop\sdsd\image ... please rectify!

[sterile kiln]

My training folder about LeahLora768

[wispy ether]

[sterile kiln]

Look my image again

in what folder are they?

and you need a prompt file for each image, with the same name

[wispy ether]

i have no idea what ur talking about

[sterile kiln]

You're blind

[wispy ether]

no

[sterile kiln]

Image>100_LeahLora768> IMAGE FILES

[wispy ether]

???

fym

i have that

[sterile kiln]

100 are the steps amount per image, characters after underscore is name for model, what you will use to call this model in a prompt

[wispy ether]

😭

wdym underscore

[sterile kiln]

you make no effort

[wispy ether]

doesnt answer my question

[sterile kiln]

Basically, with the Kohya interface, you point to the image folder. Which doesn't mean that's where you should put them. You have to make a subfolder, which is called = Steps_ModelName

100_LeahLora768

[wispy ether]

I tensorflow/stream_executor/cuda/cudart_stub.cc:29] Ignore above cudart dlerror if you do not have a GPU set up on your machine.

so not ignore it

what do i do

[sterile kiln]

What is your GPU?

[wispy ether]

3060

[sterile kiln]

Hmm should works, what did you set as params in interface?

[wispy ether]

wdym

btw CalledProcessError: Command '['C:\\Users\\Lukag\\AppData\\Local\\Programs\\Python\\Python39\\python.exe', 'train_db.py', '--enable_bucket', '--pretrained_model_name_or_path=runwayml/stable-diffusion-v1-5', '--train_data_dir=C:\\Users\\Lukag\\Desktop\\sdsd\\cp\\img', '--reg_data_dir=C:\\Users\\Lukag\\Desktop\\sdsd\\cp\\reg', '--resolution=512,512', '--output_dir=C:\\Users\\Lukag\\Desktop\\sdsd\\cp\\model', '--logging_dir=C:\\Users\\Lukag\\Desktop\\sdsd\\cp\\log', '--save_model_as=safetensors', '--output_name=last', '--max_data_loader_n_workers=0', '--learning_rate=1e-5', '--lr_scheduler=cosine', '--lr_warmup_steps=92', '--train_batch_size=1', '--max_train_steps=924', '--save_every_n_epochs=1', '--mixed_precision=fp16', '--save_precision=fp16', '--cache_latents', '--optimizer_type=AdamW8bit', '--max_data_loader_n_workers=0', '--bucket_reso_steps=64', '--xformers', '--bucket_no_upscale', '--sample_sampler=euler_a', '--sample_prompts=C:\\Users\\Lukag\\Desktop\\sdsd\\cp\\model\\sample\\prompt.txt', '--sample_every_n_epochs=20', '--sample_every_n_steps=20']' returned non-zero exit status 1.

[sterile kiln]

Load this config file in Kohya interface, change Model output name and path by yours

obviously, Model output name must have the same name as what the subfolder has after 100_

@smoky oak
I would be very curious to know what you think of these settings, I think you have explored the question much more.

[wispy ether]

what subfolder

[smoky oak]

what are we doin?

Oh, this is the issue with kohya still

I don't know

[wispy ether]

No subfolders were found in C:\Users\Lukag\Desktop\Stable Diffusion\palkiaLora20\image can't train...033[0m

idfk

i did everything right

even with _

[sterile kiln]

I was talking about the settings in the .json I sent, I don't know if there is a way to get better results by modifying some things.

[wispy ether]

wait

is it because \ instead of /

[smoky oak]

I am quite tired at the moment, but I will see if I can figure something out

[sterile kiln]

You did the same mistake, you have to set a subfolder with amount of steps _ *the name of Model * output name

name of the model have to be the same as Model output name in interface :

[wispy ether]

thats name steps

you said steps name

[sterile kiln]

You just have to put the steps in the name of the folder, not in the interface

[sterile kiln]

screen subfolder

[wispy ether]

what

screen?

[sterile kiln]

take a screenshot of image folder

[sterile kiln]

again... you don't understand

[wispy ether]

,

[wispy ether]

bruh and i thought 768 is your steps count

[sterile kiln]

no, just a name, because this training was a 768x768 pixels, but its not needed

[wispy ether]

wait

is lora part of the name

[sterile kiln]

No, image > 100_palkiaLora > SET YOUR IMAGES HERE

[smoky oak]

(number of steps you want per image)_(name you wanna use to remember)

[sterile kiln]

and point image folder with kohya interface, not subfolder

[wispy ether]

wdym point

[sterile kiln]

This :
C:/Users/Seb/Desktop/Stable Diffusion/Leah_768/image
NOT this :
C:/Users/Seb/Desktop/Stable Diffusion/image/Leah_768/

[smoky oak]

point, as in direct it to the file you want it to read

[wispy ether]

ok

wat

[smoky oak]

yes, you don't want it in the actual folder, just the image one

[smoky oak]

yeah, perfect

[wispy ether]

and in what folder do i put the image

[wispy ether]

..

why doesnt it have steps

[smoky oak]

it doesn't matter, you don't want it past image

you want it to look inside the image folder

[wispy ether]

is that right now???

[smoky oak]

no...

[wispy ether]

..........

[sterile kiln]

Kohya need to have a image folder, and, inside a subfolder : 100_palkiaLora
set images and prompt files in 100_palkiaLora, and set in interface image location, not subfolder. Its weird i know

[smoky oak]

stable diffusion>image>20_palkia

here, I actually made the exact thing you are doing

[smoky oak]

thats the order you want

and then you want to point it to the images folder, not the 40_Palkia one

[sterile kiln]

image, not Images x)

[smoky oak]

its just how I set mine up :p

regardless, the name is only for us to remember anyways

[smoky oak]

yes

there you go

[wispy ether]

[smoky oak]

perfect

[sterile kiln]

Finally, you understood

[wispy ether]

[smoky oak]

you need to add captions in order for it to recognize it

[sterile kiln]

You can use preprocessing, in automatic1111, just a set a folder with images, set preprocessing, it will give you prompts for each images, corresponding to each image files.

[wispy ether]

No subfolders were found in C:\Users\Lukag\Desktop\Stable Diffusion\image\20_palkia can't train...033[0m

does it have to be / or \

[wispy ether]

CalledProcessError: Command '['C:\Users\Lukag\AppData\Local\Programs\Python\Python39\python.exe',
'train_db.py', '--pretrained_model_name_or_path=runwayml/stable-diffusion-v1-5',
'--train_data_dir=C:\Users\Lukag\Desktop\Stable Diffusion\image', '--resolution=512,512',
'--output_dir=C:\Users\Lukag\Desktop\Stable Diffusion\model',
'--logging_dir=C:\Users\Lukag\Desktop\Stable Diffusion\20_palkia\log', '--save_model_as=safetensors',
'--output_name=palkia20', '--max_data_loader_n_workers=1', '--learning_rate=0.0001', '--lr_scheduler=constant','--train_batch_size=2', '--max_train_steps=10', '--save_every_n_epochs=1', '--mixed_precision=fp16',
'--save_precision=fp16', '--seed=1234', '--caption_extension=.txt', '--cache_latents',
'--optimizer_type=AdamW8bit', '--max_data_loader_n_workers=1', '--clip_skip=2', '--bucket_reso_steps=64',
'--xformers', '--bucket_no_upscale']' returned non-zero exit status 1.

idk the problem

[sterile kiln]

and set in interface image location, not subfolder. Its weird i know <<<<<<

[wispy ether]

let me compare

ok im confused

why is yours now name/image

you told me to image/name

cudart64_110.dll not found
2023-04-29 17:40:14.128870: I tensorflow/stream_executor/cuda/cudart_stub.cc:29] Ignore above cudart dlerror if you do not have a GPU set up on your machine.

do i have to install something

[sterile kiln]

https://www.youtube.com/watch?v=70H03cv57-o

[wispy ether]

no

i mean because dll is missing

[sterile kiln]

You have a laptop?

[wispy ether]

on

no

pc

[sterile kiln]

Its a problem with CUDA, try to use DDU to clean your drivers and reinstall latests

[wispy ether]

what is ddu

[sterile kiln]

https://www.google.com/search?q=ddu&rlz=1C1GCEA_enFR911FR911&oq=ddu&aqs=chrome..69i57j0i131i433i512j0i512l2j46i512j0i512j0i10i512j46i175i199i512j0i10i512j0i512.576j0j9&sourceid=chrome&ie=UTF-8

[wispy ether]

no

why

https://developer.nvidia.com/rdp/cudnn-archive

cant i just install this

[sterile kiln]

Cuda is already included in your drivers, you don't have to install them normally. You risk making the problem worse.

[wispy ether]

can i just click reinstall driver in nvidia experience

done

what now

[sterile kiln]

https://github.com/bmaltais/kohya_ss/search?q=cudart64_110.dll+not+found&type=issues

[wispy ether]

🗿

so what do i do

there are no solutions

only to reinstall the repo but i did that already

[random pendant]

why bother with kohya

just use a1111

[wispy ether]

i did

[sterile kiln]

Auto1111 have LoRA training now?

[wispy ether]

its kinda weird

[random pendant]

it always has

[wispy ether]

with dreambooth

[random pendant]

it just got broken for a while

[sterile kiln]

Peoples with less than 8 GB can't use Dreambooth

[random pendant]

but then again I'm allergic to lora's i'm strictly dreambooth master race

ah i see

[wispy ether]

i have more

[sterile kiln]

That's why Kohya SS is usefull

[random pendant]

nep what are you trying to do

train a style or a character or what

[wispy ether]

train a lora

[random pendant]

of what

[wispy ether]

palkia

[random pendant]

so if its a character why not try dreambooth on a1111

just saying its better than lora

and u can run it if u have more than 8gb

[wispy ether]

you didnt read ig

i just said

i did

[random pendant]

so then what happened

[wispy ether]

wait

wdy,

m

.

[random pendant]

what gpu u got

[wispy ether]

3060

[smoky oak]

@random pendantI suggest not wasting time on this, truthfully

[random pendant]

ur gucci

[wispy ether]

what

[random pendant]

it should work. u getting a dll error u said

[smoky oak]

I have been trying to help him for close to 5 hours, and he just doesn't listent, to either me or @sterile kiln

[random pendant]

have u tried --reinstall torch

oic

o wait its --reinstall-xformers but I think theres one for torch too

[wispy ether]

uh

[sterile kiln]

Answer is probably in Github link i sent for his error, but he have to read them...

[wispy ether]

i did

[random pendant]

trolol

[sterile kiln]

5 posts?

[wispy ether]

3 of them werent that related

[sterile kiln]

https://github.com/bmaltais/kohya_ss/discussions/418

[wispy ether]

hi

bro doesnt tell reinstalled what

[smoky oak]

everything

delete the whole install

[wispy ether]

u mean the repo?

[smoky oak]

all of kohya

[wispy ether]

thats the only thing of kohya i have

[smoky oak]

WHY am I falling back into this

[wispy ether]

and i already reinstalled it

[random pendant]

didnt he say he wasnt using kohya

[sterile kiln]

Personnally, i use 3.10.7, auto1111 and Kohya works. Maybe this

[random pendant]

I suggest starting from absolute scratch

watch aitrepreneur's tut

u cant go wrong

[wispy ether]

and by 3.10.7 you mean python?

[random pendant]

yeh

[random pendant]

dont use any other version tbh

[wispy ether]

[random pendant]

well theres yer problem

[sterile kiln]

No, go on dev website

[wispy ether]

and then

[wispy ether]

?

its 3.10

wdym

[random pendant]

oh nvm

[sterile kiln]

https://www.python.org/ftp/python/3.10.7/python-3.10.7-amd64.exe

[random pendant]

yeh u got it good lolol

[wispy ether]

[sterile kiln]

uninstall pyth

[wispy ether]

no

[sterile kiln]

and install the one i linked

[wispy ether]

ok

[sterile kiln]

don't forget to check "use in envirronment variable"

[wispy ether]

tf did u do

[smoky oak]

WE haven't done anything but help you

[wispy ether]

then why is xformers gone now

[smoky oak]

no idea, but apparently we did that and not you :/

[wispy ether]

uh

shouldnt 10.9 work tho

[sterile kiln]

No, some peoples tell 3.10.9 cause problems

[wispy ether]

ok

[sterile kiln]

but you didn't fully read the page i sent

[wispy ether]

so

i installed 10.7

and what now

[sterile kiln]

in fact you only read half of the things that are sent to you, so that inevitably poses a problem

[wispy ether]

so

i tried it again

same error

[smoky oak]

https://tenor.com/view/nah-nope-gif-24148633

[wispy ether]

ehm

wait

[sterile kiln]

uninstall pyth
and install the one i linked
don't forget to check "use in envirronment variable" (old setup)

[wispy ether]

i did that

what is pyth

[sterile kiln]

PYTHON?????

[wispy ether]

ok

i did that

[sterile kiln]

Oh yeah you did that, that's exactly why you have 3.10.9 installed?

[wispy ether]

wdym use in envirronment variable

where does it say that

[sterile kiln]

It was in older python setups, now its "Add Python 3.10 to PATH"

[wispy ether]

Fatal error in launcher: Unable to create process using '"C:\Users\Lukag\AppData\Local\Programs\Python\Python39\python.exe" "C:\Users\Lukag\AppData\Local\Programs\Python\Python39\Scripts\pip.exe" install gradio': Das System kann die angegebene Datei nicht finden.

I DID ADD 😭

i always do

its pip

[sterile kiln]

you have to remove 3.10.9 before

[wispy ether]

is that the command

[sterile kiln]

i could set a 4000x12000 pixel picture if you want

non

[wispy ether]

if thats the command then cmd says wrong typed

[sterile kiln]

jesus christ

help me

[wispy ether]

i did remove it

[smoky oak]

I think we are all getting pretty pyth'd off at this point

[wispy ether]

or do i have to uninstall conda aswell

[smoky oak]

@sterile kilnI think we have all tried enough lmao

We got our good karma from this

[wispy ether]

.

YES OR NO

ok

fym

i did uninstall 10.9

[surreal thistle]

Nobody:
Stable Diffusion 2.1:

[wispy ether]

bro i checked add to path

C:\Windows\system32>pip
Fatal error in launcher: Unable to create process using '"C:\Users\Lukag\AppData\Local\Programs\Python\Python39\python.exe" "C:\Users\Lukag\AppData\Local\Programs\Python\Python39\Scripts\pip.exe" ': Das System kann die angegebene Datei nicht finden.

😭

pip is broken

[smoky oak]

I am actually going to explode

python-39 is 3.9

[wispy ether]

how tf do i uninstall 10.9 if i already uninstalled it

[smoky oak]

find 3.9 and uninstall it

you have some version of 3.9.x, and its trying to default to that for god knows why

[sterile kiln]

https://www.youtube.com/watch?v=1gXj72rM7zw

[wispy ether]

[smoky oak]

lmfaooo

[wispy ether]

waiting

[sterile kiln]

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

[wispy ether]

should i delete that

so i deleted everything except that with the name python 3.9

i removed whole python

reinstalled 10.7 WITH add path

no pip

wtf

[dry crow]

Install 3.10.9

[wispy ether]

.

[wispy ether]

,

[dry crow]

He meant 3.9 i guess

[sterile kiln]

8 lines, god

[wispy ether]

bro i had 3.10.9 all the time

and you told me to uninstall it

and install 3.10.7 instead

[sterile kiln]

https://tenor.com/view/idiot-stupid-gif-19958237

[wispy ether]

dev

ok python is fucked

i uninstalled 10.7

installed 10.9

with add to path

still no pip

[smoky oak]

at this point, I think everything is messed up

[wispy ether]

yea

[dry crow]

Delete the venv

[wispy ether]

how

i didnt make one

[dry crow]

Oh okay

Do you install sd or kohya ?

[wispy ether]

kohya ss

[dry crow]

Hmm okay

Did you uninstalled Python 3.9 ?

[wispy ether]

doing rn

[dry crow]

You then may need a restart after that

[wispy ether]

ok

pip not recognized

[smoky oak]

I genuinely think you messed up on our instructions so many times that something is actually really messed up

or you did something wrong somewhere else at a different time

I know I was running SD without cuda toolkit for a month lol

that was a bit of a shock lol

[wispy ether]

i just ran the get pip file

still no pip

its fixed

[wispy ether]

soi need cuda toolkit?

[smoky oak]

I am not getting back into this-

that had nothing to do with it, anyways

[wispy ether]

cuda dll was missing tho

[smoky oak]

best of luck

[wispy ether]

wdym

even after all this

returned non-zero exit status 1.

i giv up

your method ´sucks

[calm canopy]

Nice red name 👑

Dunno if it's just a title and/or mod powers but you and guizmus definitely keep this discord running with all the questions you answer. Stay cool friend 😎

[wispy ether]

any idea why it makes random images instead of the one i use for training

i have no idea whats going on

[calm canopy]

Means it needs more training. Mabey more images idk. If you get some really baked in images, it means you trained too much.

[wispy ether]

no

i trained today already

i made a new model

it generates random images

even if i try this

[calm canopy]

Right but perhaps you used too few images or too little training strength

[wispy ether]

i did the same thing on my first model try

there it worked

that was a result

[calm canopy]

Those are all the same images so it's not really going to learn much

[wispy ether]

why 0

[wispy ether]

look at the image

i did the same as i do now

but now it doesnt learn anything

i think i did something wrong

[calm canopy]

Are these not all the same images?

[wispy ether]

yea

[wispy ether]

but thats what i did when i trained a lora with that result

[calm canopy]

Correct me if I'm wrong, I'm only seeing a tile view of it

Training under 1 image isn't really going to give you a good result

[wispy ether]

that was the result

why does it give random images

[calm canopy]

I'd be surprised if you actually did get a good lora based off of it

It's only really learning one image

[wispy ether]

and why doesnt this give results

[calm canopy]

Try feeding it at least 10 images

[wispy ether]

good?

[calm canopy]

Right, the graphs are ok but that's based of 1 image. Remember that SD is a bunch of algorithms.

[wispy ether]

why didnt it give any results tho

i just clicked train again, and it does the same

all it gives me it the parameter

graph*

[calm canopy]

Best thing to do is look for a lot more images.

[wispy ether]

nvm

cmd said restart

[calm canopy]

Sure you can train a lora on 1 image but it's not gonna be great compared to one that is fed a bunch of others (the correct way)

[wispy ether]

any tips

[calm canopy]

Best tip is to just get more images for the lora to train on.

[wispy ether]

Exception training model: 'No inf checks were recorded for this optimizer.'.

[wispy ether]

raise RuntimeError("No executable batch size found, reached zero.")
RuntimeError: No executable batch size found, reached zero.
Restored system models.

ok i give up

[smoky oak]

I'm trying to generate reference art for a character I want to make who has three heads, similar to the idea of Cerberus, and I'm trying to find the right combination of words and negatives in order to try and maximize the results that have three heads on one body, does anybody have any suggestions?

I'm wondering if I'll have to try and train a three-headed LoRA, which I have no idea how I would go about

[wispy ether]

why does it look so weird

[wispy ether]

@dry crow u got time?

😭

why is it like that

[ebon loom]

Has anyone tried to ever switch the video cards? because my NVIDIA has a better Dedicated Ram compared to the current one.
I need more Vram, as the Textual inversion training method i'm following, after seeing Ai-entrepreneur's video. But it's been an hour already, and it still says "Preparing dataset..." from the location i have pointed it out. Will it take longer?? Coz i need to quickly create something for someone's Bday, using their face, and time's running out!
Total images to be trained are 15.
Embedding Learning rate is : 0.05:10, 0.02:20, 0.01:60, 0.005:200, 0.002:500, 0.001:3000, 0.0005
Batch size : 15
Gradient accumulation steps : 1
& Max steps : 3000
.
✅ - Move VAE and CLIP to RAM when training if possible. Saves VRAM
.
✅ - Use cross attention optimizations while training
.
Please help, Or any suggestions/alternatives

[wispy ether]

[dense tapir]

@marsh lily Any new revelations?

I read it was released a couple of days ago

no windows support and they are mad

no idea I just was emailed via a github thread

Ahhh, yeah a PR

a merge

Well, I am watching a lot on AMD and they are doing a lot of stupid shit and I am beginning to think, as a lot are, that AMD is just Nvidia lite these days (which is very bad) for their shit they are pulling with releases and this.

AMD only people are pissed. Not one video from anyone being cheerful.

Their newest card to release next month makes the 4070 released recently look stellar by comparison.

pricing, and lack of vram is another

I think it is 8gb for 500 or it was 600

7600xt

599

yep 8GB of GDDR6

Well, I am with the others as this is a bridge they are crossing too far. price to perfomance to ram offered is shite. As one AMD reviewer said this is a 249 USD card at most 269 MSRP

As another reviewer said all of this hate would not be here if these damn companies offered more vram. 2023 not 2017

Well, as to that this stuff really should have been waiting or so close to being finished to coincide with their releases. We shall see so keep me tuned in for the training with it.

Well, their enterprise line up for ML is jaw dropping

A 6 year old MI25 is 16GB and as fast in SD as a lot of modern cards are. Same age as mine but for its day was 4 years, maybe six, advanced