kohya_ss Trainer Colab

[deep geyser]

That looks correct

[lime meteor]

wtf is colab's download speed

its so fast

its working

what

scam

naw

[deep geyser]

I'm glad to know I'm not clinically insane

[coarse willow]

yet

[celest crater]

shouldnt the folder containing the other folder with repeats be the project name?

oh nvm blind

[deep geyser]

Has anyone paid for 100 Google Colab compute units? How long does it last training loras via kohya?

[coarse willow]

if I remember correctly, A100 uses around 13 compute units per hours

it's not that great

regular T4 uses 1 or 2 at max

sadly the A100 is not 10x faster than a T4 so yeah

there are much better alternatives than Colab

for example, at paperspace if you have "Growth" subscription (~50 usd per month), you can use A100-80GB GPUs

one session lasts 6 hours at max, tho you have persistent storage, and also if a session ends, you can restart one immediately without any limitations

you can buy 100 compute units at Google for 12,5 USD, and it is enough for 7.5 hours on an A100 (only 40GB tho), that means that you should use paperspace for at least 30 hours in a month to make it worth more

I'm sure there are other gpu clouds also. Lamda GPU is also a pretty good value, however it is almost always at full capacity

[deep geyser]

thanks leggeka, I bought the thing and will stick to the T4 so I can use colab more comfortably for now

[dusky knoll]

hi guys im new to lora training and picked the easiest collab i found for now and wanted to ask is there something like best settings or it deepends what data set u have for an example 20-50-100 images wat settings is good to use here?

[coarse willow]

are you using the Colab in the pinned messages?

or is it Lynn's

anyways, the settings seems good for a baseline

i'm asking because there are some other important parameters like learning rate, or training mode (unet+text encoder, or unet only) which are not in the image

[dusky knoll]

https://colab.research.google.com/github/Linaqruf/kohya-trainer/blob/main/fast-kohya-trainer.ipynb#scrollTo=TN0aDXTsiwW9

im using this one

other ones look a little bit not for me :lulw:

tried to train about 10 loras but im not getting that good results

[coarse willow]

ah, well. Lynn's has all the parameters which can be edited during training, it's a more advanced Colab than these. The default values are fine but probably you should change them if you didn't get good results with the simpler Colabs.

[deep geyser]

Try the pinned one, it has instructions and less things to worry about

[dusky knoll]

i get just weird bodies like its made of dirt

[coarse willow]

ah that's because you missed your activation word from your prompt

i think mine and hollow's Colab are not training the Textual Encoder part, so it doesn't need activation words

if you apply the lora it just works

the text encoder training part was a bit confusing for me too in Lynn's colab

[shy mural]

I tought it wasnt was necessary to

[deep geyser]

I had text encoder at 5e-5 on mine, today I made it so you can change it if you want. Most of the settings say "leave as is if you're not sure"

[coarse willow]

Hollow's and my colab is using an older version of kohya trainer

[dusky knoll]

is it ok if i use small data set like 20-30 images

[coarse willow]

it's fine

[dusky knoll]

some with full body pics other portraits

[deep geyser]

yes

[dusky knoll]

so i should try both of ur colabs 😅 ok

[deep geyser]

Mine is just a cleaner version of legekka's, props to him

[coarse willow]

yeah you should use hollow's from the pinned message

[deep geyser]

Also I fixed the xformers with the help of some guy

I think it's faster now

Just had to change one line cause the dependencies changed at some point

[coarse willow]

you don't want to know how many builds did I make...

it's definitely faster with xformers

also it will be even faster with PyTorch 2.0 and model compile
5 days left until release \o/

[celest crater]

does anyone know why it defaulted to dreambooth method even if i had nothing on the reg folder?

[coarse willow]

btw

it's deprecated

https://colab.research.google.com/github/Linaqruf/kohya-trainer/blob/main/kohya-LoRA-finetuner.ipynb here's the good one

[dusky knoll]

does the images have to be at 512

because on the last one i trained with any size

[coarse willow]

no, if you have buckets enabled (which is enabled by default if I remember correctly), it will resize all of them

[dusky knoll]

alright im gonna start now training my lora

[celest crater]

this didnt happen before

weird

[dusky knoll]

should i go with default settings for first run or change something

my data set is from 20 images

so does it mean its ok to let it go with 10 or should i higher the number

[celest crater]

depends on what youre training

style or character

also how many repeats have you set on the folder

[dusky knoll]

character

and i didnt set any repeats

thats the thing i dont see here any data set repeats

on the last colab i used there was a setting how many times i should repeat my data set

[coarse willow]

you set the repeats by renaming the folder of the images

[dusky knoll]

waitwat

xd

[dusky knoll]

i see i see my bad for not reading

[coarse willow]

np it's retarded

but that's how mr. kohya_ss made the repo so /shrug

[dusky knoll]

so i create a folder with project name and create another folder in it and i rename it to xx_repetiotions

[coarse willow]

5_anythingnamecanbehereitdoesntmatter

should be "concepts"

[dusky knoll]

alright so how many repeats should i put on my data set with the 20 images for an example any idea what i should try first

[deep geyser]

you can name it anything you want just needs the number and underscore at the beginning

[dusky knoll]

yea i got it but i dont know how many times i should repeat my data

xd

[deep geyser]

Ok for 20 images I'd recommend 10 repeats and 10 epochs
Every other setting stays the same
Try that, should be like 15 minutes, if it doesn't work try more repeats or unet 1e-3

[dusky knoll]

watt i used to wait 1 hour or so kekw

[coarse willow]

lora training meant to be "fast"

[dusky knoll]

alright brb when training is done

[deep geyser]

20x10x10/2 is 100 steps per epoch, 1000 steps total.
1000 to 2000 works well for characters
I just did 6600 steps at 1e-4 unet for a style

[celest crater]

for styles you kinda just need to eyeball it

depends a lot on your dataset, lr and netdim

just make sure its not overbaked/overfit

[coarse willow]

at least you can play around with the settings and you don't have to wait 4 hours in between

[celest crater]

that looks really nice which artist was it trained on?

[deep geyser]

A videogame

600 images

https://civitai.com/models/18006/seeds-of-chaos-style-lora

[celest crater]

what dim/alpha and lr did you go for?

ill give it a try later

i might retrain my mimonel lora again so i kinda want to do it with better lr as that one was trained on like week one of the lora training days

[deep geyser]

For this one I went with 1e-4 unet, 0 text encoder, 32/16 dim/alpha, 600 images, 1 repeat, 30 epochs

[celest crater]

yeah i see 128/128 was abandoned long ago huh?

[deep geyser]

128 is cringe

[celest crater]

do you feel like bigger dim like 64 that i heard helps at all?

[coarse willow]

you can train a universe into 128

[deep geyser]

Probably not, I just went with 32 here because it was my first style, for characters I do 16
I saw a dim 1 style and it worked fine

[celest crater]

ive been training multiple outfits/characters lora and 128 dim got the best details but i guess it might only be decent for there and might also be overkill

[celest crater]

any reason for text encoder at 0 as well?

[deep geyser]

ao started doing it for styles and I trust him

[celest crater]

i guess i might as well just try with 0 and other setting since they train that fast

only way to know really lol

[dusky knoll]

yo guys can try the anything v4.5 version for training ?

[celest crater]

i train with the model we cant mention since everything has that merged

[deep geyser]

It's all a lie, you can get the same details with dim16 and dim128 if you just raise the weight when using the lora

[celest crater]

dunno if anything v4.5 is better

[deep geyser]

left is 128, right is resized to 16

[deep geyser]

Same

[deep geyser]

Bad idea imo, it'll bake in the shading and make it look flatter in the end

[dusky knoll]

so i should use the link it already was in there

was that anything v3 wait xd

[deep geyser]

The link that was already there is the one we can't mention

Works best

I just can't mention it

[dusky knoll]

oh yea mb

alright

im gonna try it

[celest crater]

btw does it say its using dreambooth method now even if it isnt?

thats kinda bugging me a bit

since db method fucks with style lora's apparently

[dusky knoll]

o well run in to issue 💀

expected for first time :x

like are u putting here the name of the project or the data set path

[deep geyser]

there you put the name of the outer folder

The inner folder has the number at the start

And you put the outer folder in the datasets folder

It sounds confusing but just look at the graph again

https://media.discordapp.net/attachments/1062362883058581544/1083840140901896192/image.png

[dusky knoll]

oh i think i skipped the project_name folder which i gotta put the 10_xx folder in it

sucks to cant read :lulw:

[deep geyser]

I could write a small python script so people don't have to go through folder hell
But it is tradition at this point

[celest crater]

you could just set a base folder and only make people add the repeats folder

but i guess people just need to understand folder structure anyway

[dusky knoll]

ok we ran it to other error code

[deep geyser]

The project name is just the name, nothing else

[dusky knoll]

is it supposed to be so ? sorry for bothering

[deep geyser]

erase the projectname, it should just be Ushio_V1

[dusky knoll]

u mean the datasets folder or

[deep geyser]

In the 🔴setup

[dusky knoll]

oh ok

[deep geyser]

When it fails you just need to run C and D again

[narrow patrol]

I was just going to ask with the new locon technique https://github.com/KohakuBlueleaf/LyCORIS/tree/locon-archive but it seems that they already integrated it

[coarse willow]

locon?

lol what a tryhard name

[fleet dagger]

LoCon, LoRA, LoHa, and LyCORIS

L Time

[deep geyser]

L

[coarse willow]

L

[fleet dagger]

L is Real

[ashen swan]

What's the difference ?

[round stag]

LoCon needs https://github.com/KohakuBlueleaf/Lycoris I believe

while Kohya LoCon is built in from what i recall?

Someone please correct me

[deep geyser]

I made a new version of my Colab. As it is a breaking change it will stay separate from the old one. Find it here:

https://colab.research.google.com/drive/1fs7oHytA4Va0IzDK-F8q_q9RIJRIh5Dv?usp=sharing

-Updated to the latest kohya version
-One click to install and run everything (will pick up from where it left if interrupted)
-More options but also more instructions
-No messing with folder names, choose your repeats directly from the colab
-Helpful messages for common errors

In the future I might add locon support

[celest crater]

[deep geyser]

Also saves the configs you used in a folder

[hollow yoke]

[deep geyser]

Added LoCon and LoHa support

[celest crater]

Havent tried any of those before heard they can get better results at lower dims but dunno what params they use

[olive beacon]

For the "official" example of loha: linear dim 8, conv dim 4, alpha 1, constant ADAMW, 2e-4 for both tte and unet
Compared against locon with linear dim 16 and conv dim 8 (the rest are the same)
Both are of size around 30MB, 5 characters + multiple styles trained in with the capacity of combining them.
This is actually documented in my civitai model page https://civitai.com/models/17336/roukin8-character-lohaloconfullckpt-8

The problem is that I have not tested on other types of tasks and a few changes have happened since then so you may still need to play a bit around learn rates for the specific tasks. Would be great if you could get back to us or just share the results on civitai with parameters documented if you are able to make a good one.

[deep geyser]

Thank you very much but WHY clip skip 1

[olive beacon]

Why not?

[deep geyser]

Because clip skip 2 is standard for every anime model and lora

[olive beacon]

They are better in the fine-tuning experiments

Clip skip 1 is MyneFactory standard

[deep geyser]

Do yo mean ADAMW8bit?

[olive beacon]

Clip skip 2 copies pastes the training images more easily in the experiments

I did not use 8bit. There is not so big difference for lora-type training.

But I think everything depends on the configuration. With both clip skip 1 and 2 you can get good results. Cilp skip 2 supposedly trains slower by the way.

[deep geyser]

One of my users made this locon with my default parameters

I say "my users" because it's some guy who speaks spanish in a small group we have

I have a spanish version of the trainer :b

[olive beacon]

Honestly I still don't understand the benefit of having tte learning rate smaller than unet learning rate, but I guess this is beneficial in some task.

[deep geyser]

Do you know if having text encoder learning rate at 0 is the same as having no text files?

[olive beacon]

No it is not. It just fixes the tte during training but the information is still forwarded to unet.

I am actually more for pivotal tuning then training tte, but need to see when this gets implemented in kohya.

[deep geyser]

Thank you - and related question, kohya lets me set "only train unet" - is this the same as TE learning rate at 0?
These might be annoying questions but I'm trying to optimize my trainer

[olive beacon]

I think so.

[deep geyser]

Thank you. I can dig into the code if I wanna be sure

[deep geyser]

I don't even know if loras can do this by themselves

but nice locon

[model_arguments]
pretrained_model_name_or_path = "/content/model.safetensors"

[additional_network_arguments]
unet_lr = 0.0005
text_encoder_lr = 5e-5
network_module = "lycoris.kohya"
network_dim = 16
network_alpha = 8
network_args = [ "conv_dim=8", "conv_alpha=1", "algo=lora",]
network_train_unet_only = false

[optimizer_arguments]
optimizer_type = "AdamW8bit"
learning_rate = 0.0005
lr_scheduler = "cosine_with_restarts"
lr_warmup_steps = 50
lr_scheduler_num_cycles = 3

[dataset_arguments]
cache_latents = true
debug_dataset = false

[training_arguments]
output_dir = "/content/drive/MyDrive/lora_training/output/toyfaces_lc"
output_name = "toyfaces_lc"
save_precision = "fp16"
save_every_n_epochs = 1
train_batch_size = 2
max_token_length = 225
mem_eff_attn = false
xformers = true
max_train_epochs = 10
max_data_loader_n_workers = 8
persistent_data_loader_workers = true
seed = 42
mixed_precision = "fp16"
clip_skip = 2
logging_dir = "/content/drive/MyDrive/lora_training/log"
log_prefix = "toyfaces_lc"
lowram = true

[dreambooth_arguments]
prior_loss_weight = 1.0

[saving_arguments]
save_model_as = "safetensors"

[latent crane]

@deep geyser is the default model in your trainer animefull-final?

[deep geyser]

yes

[celest crater]

@deep geyser i reckon you cant do multiple folders in the new version?

the feature was useful for multiple concepts

[deep geyser]

you can't, new version of kohya changed how it works, currently on my colab you would need to edit the code to add another dataset

[celest crater]

zamn

[deep geyser]

I'll see what I can do

In the meantime the old one is still there

[celest crater]

yeah ill keep using the old one for this purpose ty for your work anyway

[celest crater]

i just noticed on your page it says you trained for 6000 steps but that would be 18000 steps on epoch 30 did you use batch size 3 and this is like an absurd ammount of steps or am i being dumb somewhere?

[deep geyser]

yes batch size 3
The locon version I added today was done by much earlier

[celest crater]

zamn and its not overfit?

i usually stop way earlier

[deep geyser]

the eyes on the v1 are fucked but otherwise looks good

[celest crater]

but i guess having that many images does help

[deep geyser]

The locon was epoch 16 I think

[celest crater]

oh yeah i kinda wanna try to do an style on the new collab any settings i need to be wary for the locon?

[deep geyser]

Default settings is about right

Here's what I used

But again it was 600 images on batch size 3 so

[celest crater]

ill give it a look ty very much

[deep geyser]

My colab is cool now and spits out the config file if you wanna share it
And possibly load it back into the colab in the future

[celest crater]

oh thats really neat

[deep geyser]

I made a colab to gather a dataset for Lora training
-Choose desired booru tags and download hundreds of images in seconds
-Run the WD Tagger AI to quickly assign tags to those images
Useful if you're lazy like me
https://colab.research.google.com/drive/1-D58Nx782_aj_BYtJS0uPU79WAhS1ldL?usp=sharing

[lime meteor]

yay

[celest crater]

@deep geyser wont you happen to know some default seetings for baking a pose/concept lora? i heard somewhere of using 5e-5 for unet lr and lr but idk if it will work on 32/16

[deep geyser]

depends on the number of images, probably 2e-4 for dim32

the rest of the settings well I recommend what's default on my trainer because that's worked for me

[celest crater]

oh yeah i meant 5e-4 mb

i think that was the default for 128/128

im kinda lost with the new defaults

also would that still apply if i try to train a loha on your new collab?

[deep geyser]

i love technology

[celest crater]

lol

[deep geyser]

i trained a locon with uhh lemme check

[celest crater]

you sent me your locon setting the other day

if thats what you meant for an style

[deep geyser]

# num_images = 664
# num_repeats = 1
# resolution = 512
# keep_tokens = 0
# shuffle_caption = True
# flip_aug = True

[additional_network_arguments]
unet_lr = 0.0001
text_encoder_lr = 1e-5
network_dim = 16
network_alpha = 8
network_module = "lycoris.kohya"
network_args = [ "conv_dim=8", "conv_alpha=1", "algo=lora",]

[optimizer_arguments]
learning_rate = 0.0001
lr_scheduler = "cosine_with_restarts"
lr_scheduler_num_cycles = 3
lr_warmup_steps = 300
optimizer_type = "AdamW8bit"

[training_arguments]
max_train_epochs = 30
save_every_n_epochs = 1
train_batch_size = 3```

yea

[celest crater]

welp ill give it a try only way to know

ty very much

[narrow patrol]

Nothing goes in the tag, any keyword? style etc?

[deep geyser]

I don't understand the question
For this locon as it is a style there's no activation tag, it just works

[tacit zinc]

if i want to train a single lora with finetune on multiple related concepts/trigger words (like tamamo no mae wearing default costume + casual + swimsuit) how should i structure my dataset folder
im not even sure how to do that with the dreambooth colab, could it even be done in the newest versions?

[tacit zinc]

should i just dump all the 3 datasets in the same folder and only distinguish them by tagging them correctly by concept?

[hollow yoke]

hollow, i need help

[deep geyser]

I removed the vae from the anime model, I might've done it wrong, for now use model_with_vae.safetensors

[hollow yoke]

Ah, I don't really understand. I will try again later. Thanks for the answer.

[celest crater]

change the link for the model for this one https://huggingface.co/LarryAIDraw/animefull-final-pruned/resolve/main/animefull-final-pruned.ckpt

dunno why it happens but this one worked for me

btw hollow first concept locon trained and the results are amazing

[hollow yoke]

[digital girder]

Hi all, kinda noob question. Could someone recommend some models to train LoRAs on? I'd like to train a mangaka's style but not sure which model I should use

[celest crater]

Usually its ||novelai|| for anime models and stable diffusion for realistic models.

[digital girder]

thanks a lot, I just found Hollowstrawberry's colab and the guide linked there also provides good pointers so that's what I'll do.

[warped cloud]

anyone have thoughts on specific details of a character? like should I add cropped images to the dataset of that specific detail (specifically katarina's belly tattoo)

[deep geyser]

One or two can't hurt, just make sure to tag it correctly

[celest crater]

For more detailed lora's I've seen people say 768 resolution is better, for example to get nahida's cross shaped pupils was a nightmare for everyone until someone trained at 768 and only tagger her pupils when you could make them out by the thumbnail alone

[warped cloud]

I added like half tattoo shots and half normal pics and removed tattoo tags and it seemed to work well and consistently gave her tattoos so need to do more testing with her eye scar and then to see if removing the tags was the reason why or what be

[pulsar gorge]

haven't checked. Is this updated for LyCORIS?

which, coincidentally, sounds a lot like Lykon

[fathom summit]

After curating dataset images, ran 4️⃣ Tag your images cell, this error occured

[deep geyser]

oops <= instead of <

[fathom summit]

will try

[deep geyser]

off by one errors, my favorite

[fathom summit]

I also saw you're trying to make fancaps scrapper

maybe there's something you can take from here: https://github.com/Fannovel16/fancaps-scraper

[fathom summit]

works btw

[deep geyser]

also oops i was working on the live version earlier

[fathom summit]

lolol

[deep geyser]

nice, fixed on the original too

[deep geyser]

thamks

[fathom summit]

Also, can I request a few feature here.. Idk if this will help in training tho

Character tags in wd tagger
and a way to get tags from gelbooru then merge it with tags from wd tagger (you have remove tag duplicates already ig)

[unkempt mantle]

hi

[deep geyser]

I can add that yes

The reason we don't just grab the tags from the website is there's a bunch of tags we don't care about like artists and franchises and rating and stuff

But it's easy for me to save them if you need them

It's harder to mix them with the other ones though

Cause wd tagger just replaces the whole file

[fathom summit]

I see

https://huggingface.co/hollowstrawberry/animemodel/resolve/main/model.safetensors
is this nai + nai vae or wd vae? 🤔

nvm it's wd vae

[deep geyser]

it's the original animefull-final-pruned but fp16 and safetensors

it came with its vae baked in

[deep geyser]

no

[fathom summit]

You sure?

The outputs from that model really vibrant, but after I used nai vae from my repo it's less vibrant

[fathom summit]

@deep geyser
can I request one more feature?
fiftyone for removing duplicates,
since even after deselecting include_posts_with_parent dup still exists sometimes, and I have 400+ images so manual curating is a bit painful

colab in ss: https://colab.research.google.com/drive/1oBSntB40BKzNmKceXUlkXzujzdQw-Ci7

[fathom summit]

hollowstrawberry's model.safetensors test
Left, VAE: None
Right, VAE: Nai Vae

[fathom summit]

emptycanvas = nai w/ nai vae baked

[deep geyser]

Surprising findings, let me test
But I can assure you it's not wd vae

[deep geyser]

Yes I would love to add this, thanks for the reference

[deep geyser]

What the fuck is happening

[deep geyser]

now anime.vae.pt gives me an error?

and this is with my model which has baked vae

[deep geyser]

I have added this feature

duplicates are marked for deletion and you can mark more if you want

much better than sifting through google drive

thank you

[hybrid juniper]

at the fourth step it just does this for me

well i just fixed it somehow by instead of running the steps in order i just restarted the run time and ran step 1 and then 4

[narrow patrol]

it's perfect LoCon xd

[narrow patrol]

#1062362883058581544 message

I use this

[deep geyser]

thank you, hopefully I fixed it now

[hollow dew]

I am struggling to make it work on jupyter for several days now, does anyone have a hint on what I should do please ?

I am not used to jupyter and colab, is there a guide somewhere ? I just want to run the training locally instead of on google's servers

[coarse willow]

it is possible, but you have to use a venv or conda at least for it. There are a lot of guides how to setup local runtime for Colab

https://robotwealth.com/how-to-connect-google-colab-to-a-local-jupyter-runtime/

[hollow dew]

Thank you I will try

[deep geyser]

[deep geyser]

advanced features

[celest crater]

Does it work for lycoris?

[deep geyser]

ye of course

[deep geyser]

it seems theres some issue after xformers updated yesterday

all my loras come out fried from epoch 1

[celest crater]

aww bummer wanted to bake an style lora before going to bed

[fathom summit]

input ":question: Enter the word 'yes' if you want to proceed with the download: " didn't shows up
@deep geyser

I used this as temporary fix

[deep geyser]

oh that happened to me once, idk whats up
i just ran the cell again and it worked

also nice color scheme

[fathom summit]

I used dark theme + this colorization

[deep geyser]

how do i get colorizations

[fathom summit]

tools > settings > site > change theme to dark > editor tab > colorization

then double click a cell to refresh the theme

[deep geyser]

oh i had to switch from adaptive to dark

[fathom summit]

yep

[fathom summit]

also why no aria2c in here
I saw it was installed

[deep geyser]

aria2c works with single files? whack

[fathom summit]

da speed dude

https://cdn.discordapp.com/emojis/940306480555495444.gif?size=64&quality=lossless

[deep geyser]

eh needs some tweaking so I won't do it right away

That was civitai crashing, it's usually like 100 MB/s
aria2c is still at least twice as fast

[shy topaz]

tf

lmfao

keeps happenin

rip

[deep geyser]

Hello what are your configs

[deep geyser]

done

[shy topaz]

just the default ones

well, mostly

[deep geyser]

did you happen to select locon/loha?

I haven't seen this issue before
Please try again as some stuff broke yesterday

[shy topaz]

lora

i think its fixed

i didnt change anything

except resolution

from 768 to default

[manic goblet]

WARNING The following values were not passed to

any idea on how to fix for lora training

[deep geyser]

We'll need more context

[manic goblet]

i tried to run the kohya easy script filled out everything and it would load the folder when it was in the dataset folder

got that error

[hard wing]

help
Found 259 images with 10 repeats, equaling 2590 steps. 📉 Divide 2590 steps by 2 batch size to get 1295.0 steps per epoch. 🔮 There will be 10 epochs, for around 12950 total training steps. 💥 Error: Your total steps are too high. You probably made a mistake. Aborting...

[supple urchin]

Lower your repeats and epochs

[deep geyser]

Your repeats are too high, it would've taken like 6 hours to train

And possibly burn it

use 1 or 2 repeat

[hard wing]

i use 2 and its working thank you 😄

[supple agate]

how do i check the first post?

[deep geyser]

You can't but you can check the pinned post

[supple agate]

[coarse willow]

it's not a bug, it's a feature

[manic goblet]

shut it tod

[crude zephyr]

i got the error how i can fix it ,

?

[deep geyser]

Try to disconnect from the session (top right corner) and reload the page to try again

I used the colab half an hour ago so it should work

[celest crater]

Share your training config and dataset if you can and it ends up looking good cause i might want to retrain a multi outfit lora i did long ago with the new tools we have now

[deep geyser]

Uploaded the dataset and config here
https://civitai.com/models/28686

[celest crater]

Also saw your guide on pruning and dunno if its the best way to go cause other people do go for the overboard tagging

[deep geyser]

Worked out after 3 tries

Had to un-prune many tags

[celest crater]

And individual tags for stuff that repats

Instead of just pruning

[deep geyser]

What did I say about pruning I forgor

[celest crater]

Prune stuff you want to be always there or is common to the character

Which is true if you just want to prompt one outfit hairstyle etc

I think you could probably rephrase it or something in case people wanted more flexibility

[deep geyser]

I got told I'm doing it wrong by not pruning everything and not having a single tag
My feelings got hurt lmao

[celest crater]

I think that way of doing things stayed from embeddings and idk if the ai doesn't benefit from the help of the tags it already understand

[deep geyser]

There's many things to say about pruning and I'm not sure about any of them at this point

[celest crater]

No no its not wrong since the tag you add absorbs basically what your prune and tagger usually does a lot of false positives

But we certainly would need more info and data on the matter

Yeah

In my experiments

90% of the time

Dataset was the defining factor

That and not baking stuff to hell and back

Haven't tried regularization images either which might or might not help

[deep geyser]

This lora is probably my best one cause it actually balances multiple outfits and gives each an activation tag

even if I had to un-prune a bunch to prevent them from bleeding into each other

[celest crater]

It does look very good but you basically did what i told you more or less

There are people that legit prune all outfit tags etc

And just add an inicialization token

And call it off

[deep geyser]

derrian says they're not working correctly in kohya
idk

[celest crater]

There was one dude experimenting with that in one of the multiple lora rentries but idk if he ended up saying anything definitive about it

[deep geyser]

Did you tell me previously or you mean right now?

[celest crater]

Now

[deep geyser]

gg

What do you think I should say about pruning in the guide?

I want to present flexible options while also guiding people on the right track

[celest crater]

I didnt see your other tries if you posted em but if you can reliably get the 4 outfits at 0.7-1 and can gen more than those outfits and its not super overfit i think thats a sucess for multiconcept,multicharacter,multioutfit

At least from what ive seen so far

Like other either that or finetuning a model and thats painful as fuck

[deep geyser]

yeah it works fine at 1
Only the hair bow is baked in for some reason

[celest crater]

Honestly i dont really know cause almost everything you said and have provided already streamlines the process well enough for everyone

And people that just want to train their first lora wont do schizo tagging

[deep geyser]

This was my first try I think

[celest crater]

But you could probably mske the distinction for more flexibility

Actually let me search for a tagging guide i saw a few weeks ago

From an anon that does blue archive lora's

It goes really deep into it

[deep geyser]

the human form got fucked

the eyes were not accurate

[celest crater]

Yeah eyes you can give up

Nahida was the test subject gor getting decent eyes on loras

And getting them required 1mp training

And autism tagging

• only tagging the eyes when you coulf clearly see them by the thumbnail alone

Usually its just better to do an inpaint lora on top

@deep geyser https://huggingface.co/khanon/lora-training/blob/main/kazusa/tagging methodology.md

There is also the issue of duplicate tags on multiple outfits

And idk if just the trigger tag can soak the differences alone

So you probably also need to change the dupe tags for one of the outfits which is a pain in the ads

And another thing ive seen people focus on recently is multiple angles with proper tagging

Which helps a lot on getting details right

Which is also more work

At the end of the day if the dataset is good you can just autotag and make a quick lora and the results will be decent enough

So this is probably only for people that want to go the extramile

Additionally idk if its the best way to go about it but ive seen people train on 128/128 and then resize the lora with the resize script thats on kohya repo

I think it retains more details that directly training at lower dim/alpha but it might be biased

[deep geyser]

ye i did which is how I got these 2 slightly different dresses to work

Still not perfect but who's gonna notice

[deep geyser]

I've resized several from 128 to 16 and it works nicely

It's impossible to know if training at 16 directly would've been better for those as they're old and not mine

[deep geyser]

ZOOMIN

[narrow patrol]

in any case, use lora locon, it works for style art, but try using it to make characters, and it works well I think

[celest crater]

I found locon easy enough for styles for characters i couldn't get the same ammount of detail

Also found more fucked up hands for some reason dunno why

[deep geyser]

Locon isn't good for characters

[chrome crest]

For the dataset maker, is this needed if I want to do a style without any activation words?

Same with this setting in the Lora Trainer

[deep geyser]

nope

[deep geyser]

You can still tweak the tags but it's not too important
I heard text encoder is important for locon styles btw, so do make sure the tags are well represented

[chrome crest]

Ok, so I leave the activation tag to "0" in the lora trainer?

[deep geyser]

yes

[chrome crest]

Alright, thanks

[celest crater]

@deep geyser so how are you doing your dataset for these holo multioutfits lora's? as in number of images per outfit and lr

[deep geyser]

It should say so in the version description

As for how, you mean with my colabs?

For example for Ina
In the dataset maker I set the project name to ina/priestess, and do everything for that outfit. Then do the same for ina/kimono, etc.

In the lora trainer I set the project name to ina, then scroll down to the custom dataset box and write the folders I'm using and their repeats

[deep geyser]

As to the number of images, ideally you want at least 20 per outfit, but less still works.

What's important is the images×repetitions. You need to balance those if you want the outfits to work equally well. For example, imagine you have 3 folders:
Folder A with 100 images
Folder B with 50 images
Folder C with 20 images
You'd set 1 repeat for A, 2 repeats for B, and 5 repeats for C, then all of them have 100 images×repeats for a total of 300.
A healthy amount of images×repeats is either 400 in total or around 200 per outfit, whichever you can get.
It's not always so simple but yeah

[celest crater]

i just have my nene dataset from super long ago laying around and it was honestly such a messy job so i was wondering if maybe doing it in your collab with more careful tagging would do the work but like idk the learning rates i should use in this case given the huge ammount of images

[deep geyser]

nene huh? I forgot nene existed

Is there a lora on civitai? There's one in a mega link at least

[deep geyser]

2000 steps with min_snr_gamma and 5e-4 unet works really well for me

3000 if I'm feeling fancy

[celest crater]

min_snr_gamma is that a new thing?

oh smoother training that seems nice

did you use flip aug

?

[deep geyser]

Make TRIPLE SURE that you won't regret using flip aug...
Any detail that goes to the left or right will start going to either side, and sometimes to both sides (though this also happens without it)
I went from having it off by default, to on by default, to off by default

You can also flip aug specific folders only

[celest crater]

i would care for this except for the fact that the outfits that have less images are the ones that have either the hairclip or the flower ornament on one side

and honestly i dont think its worth the loss of quality overall

over those being on the wrong side of the hair

[deep geyser]

ye whenever loss goes back up from a training step it tries to diminish it
min_snr_gamma halves the amount of epochs I need to train for

[celest crater]

did you notice if using ina_x

made outfits bleed into ina?

cause when i did mine for nene i used outfitx

and i suppose it was overbaking but a bit bleed onto the others

but i didnt do as much careful tagging

in theory it should be a different token for each

but idk anymore

[deep geyser]

i didn't get outfit bleed with ina, but I did with ollie. Took 3 tries, I had to un-prune tags

[celest crater]

yeah...

[deep geyser]

That's when I turned off flip aug

[celest crater]

well you would get weird swords i reckon

with flip aug

[deep geyser]

Worse

Her FACE is asymmetrical

I didn't think about it either

I'm an idiot

Instead of a face like (\) I got (X), and both eyes the same color

[celest crater]

[deep geyser]

[celest crater]

I kneel

🧎‍♂️

[proud belfry]

@deep geyser Thanks for sharing lot of information!
What repo do you use? With Kohya I don't have min_snr_gamma.
For unet 5e-4 but which value do you use for Text encoder?
And as LR is related to batch size, which do you use?

[deep geyser]

I use my colab which is pinned here and #1093035978320523264
For local training I use #1073837013758902355
I use text encoder 1e-4 and batch size 2

[proud belfry]

Thanks a lot. I'll make more test.
For now I can't reproduce my results in a quick way.
The settings I used take 12 hours to get a correct lora 😢

[proud belfry]

I use exactly the settings you recommend and it was very bad for me.

[deep geyser]

Sorry to hear. What are you training?

[proud belfry]

a style on a small dataset

[deep geyser]

ah yes that'll probably need different settings
I'm most knowledgeable on character loras

[proud belfry]

ok.

[celest crater]

how small of a dataset are we talking about?

[proud belfry]

24 images.

[celest crater]

Yeah thats really small of a dataset for artstyles

Ideally you have 100-200

What you can do is make closeup shots of faces for example if its an anime style

With proper tagging of closeups to artificially create 2x images

Then you could use flip_aug and play with the learning rates

Cause it will probably overbake before you get what you want

Also make sure for artstyles you need to train for longer characters usually are done between 1-2k steps with normal settings, styles is not a certain science and depends a lot on your dataset like everything honestly but it should bake for longer

Or until you basically can notice it with it losing flexibility

[supple urchin]

When I was researching for my style lora (example attached) I saw that usually training between 4k-6k steps will get a style down pretty well. I trained on 48 images. My understanding is that quality of data is better than quantity of data (once you get above a minimum dataset size) so as long as your tags are good, you should be able to get a nice and flexible lora

[celest crater]

@deep geyser did you use easy_scripts to resize your loras?

[deep geyser]

yes

[celest crater]

i saw he added new stuff to the bat file if you dont mind sharing the settings you used i would appreciate it

[deep geyser]

Well I use it to resize other people's loras

[deep geyser]

There should be a text file with the recommended settings

[celest crater]

these ones?

[deep geyser]

[celest crater]

i reckon you used dinamic

[deep geyser]

I used fixed actually, but dynamic is better

Dynamic didn't exist yet

[celest crater]

thats what i heard

kinda curious to actually finish tagging my nene lora again

and see if resize doesnt butcher the details

my last try got outfits right but because i pruned the hairstyles like an idiot they got all mangeld up

nene has a million variants for each outfit which doesnt help at all

and not that decent art

like the china girl outfit and the main one are okay

even the blazer one

but the new years, swimsuit and a few random ones here and there are wew

[deep geyser]

Yeah I learned that in my 3 tries for Ollie

[deep geyser]

Mind if I try?

[deep geyser]

Are you tagging manually?

[celest crater]

yeah

but the dataset is really big

so sometimes i miss some stuff

im gonna go to sleep in a bit

ill hit u up tomorrow

this is the old one i did when pruning everything and added an activation tag was a thing, its also gigabaked so i used it at like 0.7 and had to use outfit0,outfit1,outfit2,outfit3,outfit4 + some common tags and there was a lot of stuff appearing on the others if i went above 0.7

[deep geyser]

[celest crater]

it was also done when we barely knew anything about loras

i reckon with your collab and the new tools it will go easier

actually let me just link you the mega link that also had the dataset

so you can see what a mess it was

it was basically all pruned lol

like 1 tag cant hold everything

[deep geyser]

I didn't realize nene was missing from civitai and she's my wife so now I want to try
But I don't want to step on your toes if this is important to you

[celest crater]

https://mega.nz/folder/Zt9BzJza#T9235jamelU1ULNJJ-0ABQ

[celest crater]

no no

go for it

feel free to reuse this old dataset

if you want

[deep geyser]

I have the means cause I did 9 outfits for rushia, the idol outfit had only 5 images and I went to get screenshots and it came out nice

[celest crater]

if you end up making something better than i do

i benefit as well

lol

[deep geyser]

That's a nice mindset, thanks
Will look into it
Will probably do it the real way and sort all images into folders one at a time

for rushia i just sorted by booru tags but it's more inaccurate than one would hope

[celest crater]

i dont care for the likes, etc on civitai

specially since we didnt really do any images ourselves

and got them without artists permission

but i like ai image generation and want to learn

so its the best mentality to have

[deep geyser]

True, that's why I have the «do whatever the fuck you want to» license

[celest crater]

i always put dont sell

but the rest is fine

[deep geyser]

if people wanna sell im like
ok
you can find it for free so go ahead

[celest crater]

i find it a bit scummy to sell but people that do ai stuff on pixiv set up their accounts with fanbox

like at frame 0

so its whatever

at the end of the day i still want to learn to draw on my own

but not learning this stuff for the foreseeable future would be shooting myself in the foot lol

hmmm

the example txt file is not working

thats weird

[deep geyser]

[celest crater]

nvm im a dumb fuck and was on the wrong option

[celest crater]

mine are ordered btw its just the old dataset cause i havent uploaded the new one yet

so at least you wont have to order the og outfit, default, blazer, swimsuit and new years one

just remove the txt files

and tagging hell

also i felt like there were like mayhaps

like a gigazillion pics

[deep geyser]

but if i don't slave myself like a pig what's the fun of making a lora?

[celest crater]

for default outfit, og outfit and random pics

and then blazer, new years outfit and swimsuit took a hit

[deep geyser]

anyway thanks will take a look

[celest crater]

because of repeats

i saw you try to keep em on the hundreds

which might be the right call

[deep geyser]

200 steps per outfit is the way to go

20 images minimum

[celest crater]

new years didnt get that shit

even if it had low images

but swimsuit took a hit and i think its cause there were random pics

with swimsuit tagged as well

that and well

the prunning of all fucking tags

lol

the ai had to do extra work that day

[deep geyser]

see i have 18 images of new years it's fine
but 11 images of the fucking bikini
it's barely an outfit I can't find a single source for it

I have more images of the unofficial bikini

[celest crater]

yeah....

[deep geyser]

I'm dropping the bikini it's not happening

[deep geyser]

right is easy but left is not enough images

[deep geyser]

@celest crater

will civitai tomorrow

[proud belfry]

Yes, I know dataset must be bigger but... it can't be 🙂
I got some good results with this dataset but the training takes between 5 and 12 hours.
So I'm pretty sure I can optimize those settings.
Batch 8, 10 repeats, 100 epochs, (3000 steps), constant, lr 5e-5, text 2e-5, 128/64 dim/alpha.

[celest crater]

🧎‍♂️🧎‍♂️

Only thing i can notice at a glance is the skirts took a hit

Otherwise it looks really good

Lets see if ill be able to do any better

[celest crater]

Only god know what you have to be doing to get those times qith your small dataset

For starters dont do batch size 8

Do 1-2

[supple urchin]

My research told me to have repeats and epochs equal, also your dim/alpha is pretty large, I'd go smaller

[proud belfry]

thanks.

[supple urchin]

Maybe try something like 22 repeats, 22 epochs, batch size 2. For 24 images that should get you to 5,808 steps. You can probably get something passable with the tiny dataset as long as the tags are near perfect

[proud belfry]

Thanks a lot, I'll try that.

[supple urchin]

Good luck!!

[proud belfry]

thanks

[upbeat matrix]

what model do yall train loras on? I want to train a lora for AOM3A3 but since the style is 3dish theres not many images online for that kinda training. So if I train with anythingv3 and then just use that lora with AOM3A3 will it work?

[celest crater]

aom3a3 was mixed with nai so train on nai

[deep geyser]

nai is almost always the best

[normal stirrup]

Train in nai
Test in nai

[deep geyser]

I test in the model I actually wanna use

[upbeat matrix]

nai as in novel ai?

[deep geyser]

yes

animefull-final-pruned

[upbeat matrix]

anything v3 is basically nai but better right

[deep geyser]

it's worse

(For training)

[upbeat matrix]

I see

[upbeat matrix]

where can i get this

[deep geyser]

https://huggingface.co/hollowstrawberry/stable-diffusion-guide/blob/main/models/animefull-final-pruned-fp16.safetensors

[upbeat matrix]

fp16 fine to train with?

[deep geyser]

all loras are fp16

Unless you're weird and change it

[upbeat matrix]

seems trash for generation tho lol

unless my prompts are trash

[deep geyser]

Absolutely

I don't know the logic behind it but it makes great loras

[upbeat matrix]

AOM3A3 vs Anyv3 vs NAI

same seed and all

[proud belfry]

it's still not perfectly close but closer than before.

[deep geyser]

nice job

[knotty rain]

when i'm tryna use the lora colab it makes a duplicate dataset folder instead of recognising my dataset folder

[deep geyser]

[coarse willow]

dataset <> datasets also everything is case sensitive

[knotty rain]

i did dataset

it's not detecting because of the text files containing the prompt

do i place them separately in a folder?

[deep geyser]

Is it a duplicate folder with the same name? This is a rare bug with Google Drive...
Try moving the images to the new folder and deleting the old one
Alternatively, rename both folders, wait 5 minutes, rename your dataset folder again, and wait 5 more minutes, then restart the colab

[knotty rain]

Yep that's it

I'll try tomorrow and update you

[deep geyser]

[knotty rain]

I did the moving images thing too, it created yet another folder

[deep geyser]

[chrome crest]

ok

[void dragon]

Hello, I want to install Kohya to train custom loras. I found a couple of youtube tutorials that are 2months old and apparently outdated. What is the most recent, beginner friendly guide I can follow?

[proud idol]

it hasn't moved in 50 minutes

[deep geyser]

Oh wow that's a lot of images

But I don't know why it would fail

[proud idol]

xd

[proud idol]

So I shouldn't wait?

[proud idol]

I can give you a link to this dataset, if you want to find bug or something

[obtuse oak]

what were you trying to train?

[proud idol]

arts by tag vocaloid in danbooru

[proud idol]

(Locon Lycoris)

[unkempt mantle]

whats the best amount of dataset for a usual style lora?

[deep geyser]

100 images is the sweet spot

If you can get 200, cool
If you can get 50, fine

[unkempt mantle]

so 100 is like the amount that doesnt fry it

[deep geyser]

Within some assumptions, the answer to that question is yes

[unkempt mantle]

i am a lazy potato so whats the best settings for 62

dataset

images

or whatever

[deep geyser]

With 62 images, do 5 repetitions and 10 epochs, 2 batch size, 5e-4 unet, 1e-4 text encoder, min_snr_gamma at 5, cosine with 3 restarts, 150 warmup steps
That's how I'd do it but there are many methods and none of them is perfect

[unkempt mantle]

thank you i will now forever save that in my notepad

[deep geyser]

You could also do 2 repetitions to match the 2 batch size and just have more epochs like 20

But that's usually what you'd do with 100 or 200 images

[unkempt mantle]

so 2 rep 2 batch for 20 epochs

so what if its 30

is it going to be 1 rep and 1 batch

[deep geyser]

30 images is not quite enough to stabilize it, but it'll still be usable
I always do 2 batch size though

So maybe 4 repeats with 30 images

[unkempt mantle]

wait i mean 30 epochs-

[deep geyser]

And enough epochs to reach 1500 to 3000 steps

[unkempt mantle]

i forgot to add the epochs

im now going to die in a hole

<3

[deep geyser]

Dont die in a hole thats bad for your health

Anyway 30 epochs would be 1 repeat yeah
But like I said I always do 2 batch size

[unkempt mantle]

but if i die i wont need to worry for my health

[deep geyser]

I tried a 7 image lora and it seemed to do better with 2 batch size than 1 batch size
I tried a 1000 image lora and it seemed to do better with 2 batch size than 3 batch size

[unkempt mantle]

1k image?

girl how

[deep geyser]

well if you don't give a fuck and do popular characters it's easy

but if you do multiple outfits like me then yeah you do reach those numbers too

[unkempt mantle]

im a style lora person

[deep geyser]

I'm a character lora person

Also supposedly locons are better for styles so probably do that

[unkempt mantle]

are locons smaller

[high talon]

Or be me an do whatever and pay the consequences later XD

[deep geyser]

they're, bigger, but you make them smaller, so they're the same size

[unkempt mantle]

im sticking with loras

i dont want to learn an entire math equation again

i took 2 months to learn how loras work

i dont need another 2 months of my life gone

[high talon]

Oo; same math

[deep geyser]

Nah locons are just

[high talon]

if you use holo's notebooks just same math XD

[unkempt mantle]

well imma go retrain that one lora that failed

successfully

[deep geyser]

A locon is just a lora but you have a second layer of neurons
Normally you'd have a lora with dim32 alpha16
For locon you halve it, then halve it again for the conv layer, so dim16 alpha8 conv_dim8 conv_alpha4

It's that extra layer that learns styles better

[deep geyser]

I forgot to say dim32 alpha16 is epic

mb

[unkempt mantle]

wait why did i come here again

i have memory loss

[unkempt mantle]

what colab are you using

[deep geyser]

mine, pinned here

[unkempt mantle]

.

how do you add sample steps

[deep geyser]

There's no samples in my colab

[proud idol]

If you can get 6059?

fsfsfsfsfsfsff

[deep geyser]

Trim it down to 400 and do 1 repeat

Pick the best ones, no way you have 6000 good images

[proud idol]

I just initially wanted to tune natively, but then I realized how long and inefficient it is

[unkempt mantle]

is the trainer only made for character training?

because when i trained my style lora it didnt apply any style

[deep geyser]

the default values are what I use for characters but it's not "made" for it
You might've had a learning rate too low. Also, did you try locon?

[unkempt mantle]

ah nope

though whats good for locon 62 dataset?

[wispy urchin]

HOW DO I USE IT

AAAAAAAAAAAAAAAAAAAAAAAAAA

where do i put my images

[unkempt mantle]

uhh

you go throguh 12308998939023 steps

[wispy urchin]

HELP

[deep geyser]

the defaults should be fine - don't go over 5 repeats though

[unkempt mantle]

ok thanks

[unkempt mantle]

but what do you mean by "over 5 repeats?" also can you add auto tagger?

[deep geyser]

there's a tagger in the dataset maker

[unkempt mantle]

oh but can you do the same in the regular colab

[deep geyser]

no

[unkempt mantle]

.

[deep geyser]

They both use the same drive folder just go to the other colab

[unkempt mantle]

is the dataset maker able to gather artstyles

[deep geyser]

from gelbooru? you can use an artist tag yes

[unkempt mantle]

ok thanks

[unkempt mantle]

what did you mean by 150 warmup steps here

[deep geyser]

5%

[unkempt mantle]

is it for other colabs?

[deep geyser]

my colab has warmup ratio

[unkempt mantle]

oh

can you add emojis to the name of the lora for no reason

ok so default settings is fine for 62 dataset?

locon

[unkempt mantle]

?

[deep geyser]

Yes this is what I told you last time and that's the default except for the repeats

If you tried it and it didn't work then oof
locon for styles is better though

On the last part the defaults should work
dim16 alpha8 conv_dim8 conv_alpha4
or conv_alpha1 idk

[unkempt mantle]

you were correct i successfully trained the locon a bit nsfw though

[deep geyser]

Looks cool

Congrats

[wispy urchin]

HELPPPP

MINE IS UGLY

[normal stirrup]

Maybe it s dataset issue sir

[pastel vine]

Helloo I would like to ask senpais for some tagging advice

I was training Lora for Kaguya Otsutsuki, but considering she's wearing a white robe that makes her look flat on most of the dataset images it's hard to derive from small breasts and no robe when rendering txt2img.

Soo, if it's recommended to remove characteristics tags to stabilize the character's traits should I tag her with "small breasts, white robe" to achieve the opposite - no robe and bigger breasts?

[deep geyser]

That sounds about right

There's also flat chest

[pastel vine]

Will try, thank you also did someone maybe put together a document of efficient training settings for different dataset sizes? I've been looking them up in channel histories so far 😆

[celest crater]

Most of what's been researched in general resumes to "it worked for me" so idk what to tell you

Usually having a good dataset and decent ammount of images in it does most of the work

[deep geyser]

Pretty much yeah

I'm sticking by my 400 images×repeats, 5e-4 unet, dim16, batch size 2 setup

I just trained a Locon with 66 images × 5 repeats plus 35 images × 2 repeats
Thanks to some tweaking that comes out to exactly 400

And it worked perfectly

[pastel vine]

appreciate it, thanks for the info!

[unkempt mantle]

@deep geyser could you add an auto background adder for zip datasets on the trainer?

[deep geyser]

Thank you for the suggestion but I believe it would be useless as transparent backgrounds are treated as black by the AI

[unkempt mantle]

i really need it

and its helpful if you have a transparent dataset

[gentle kettle]

Id like to tag my dataset in a way where it focuses only on the face. What ive tested is that the clothes and other characteristic seem to leak during genearion. Would like to ask some tagging tips. Thanks for your attention :)

[deep geyser]

Tag everything

Except facial features

[unkempt mantle]

will there be a feature to train lokr's?

[gentle kettle]

Ok? I wanna use WD 1.4 tags on this dataset

Also does the normal settings for the learning rate and scheduler work well for my objective?

[deep geyser]

I'm not really sure, but it probably is

[deep geyser]

I'll look into it if there's demand for it

[unkempt mantle]

its smaller and seems a bit better

https://civitai.com/models/35136/mika-pikazo-lokr
this one is 2mb

[gentle kettle]

Okay man. Btw can i DM you sometime? I have some further questions

[deep geyser]

Sure, just leave me a question and I'll try to answer

[little shuttle]

yeah

[deep geyser]

Sounds like demand to me

I'll dedicate a couple hours to it when I feel less tired
Just need to update kohya to the latest release if applicable, test it, look up the parameters for lokr, account for those, test it, profit

[little shuttle]

See https://civitai.com/models/48738/megumin-ed-style-lokr for another example of how Lokr is overpowered

[proud idol]

What type of Lora is better to use for some phenomenon/action?

[little shuttle]

#1094438173892943892

[unkempt mantle]

@deep geyser

[faint bronze]

Recently compared my newer Loras with old stuff I made with the previous version of the lora trainer and saw that old Loras were better. Probably using a bad setting but does anyone know what the default setting for old training was?

[high talon]

Dunno but my newer ones using holo's are better than my old ones XD

[faint bronze]

Mean as said I'm probably using wrong settings so trying to figure out what I'm doing wrong

[deep geyser]

My trainer spits out the config file so you can check

As long as you didn't delete the old folders

[deep geyser]

You're using the wrong model, possibly stable diffusion 2.0

[faint bronze]

Old one didn't at least for me

[deep geyser]

Oh so it's really old

A lot has changed since then I guess

[vital aspen]

The trainer isn't working for me anymore, gives ERROR: No matching distribution found for torch==2.0.0+cu118

[deep geyser]

[deep geyser]

Try now

[vital aspen]

yup its fixed, awesome, appreciate it. Was the change super easy? I'd love to know how to fix it if it happens again

[deep geyser]

when google updates colab itself the versions change
this time I could've tried to adapt to their new versions but I just forced it to load it from the old place

[vital aspen]

ah okay

[pulsar gorge]

I wonder how a 8dim resize of my mika pikazo lora would compare

at the moment it's 64dim

in theory lokr is just a different method for parametrization

I wonder if a normal lora training + resize approximation is a viable alternative

[deep geyser]

Dynamic resizing is pretty effective
I don't know how it works though

[high talon]

~_~ i shave your loras down a bit so i can use them for stuff COUGHS i should literally give you the 32dim ones for your stuff - AND any mooshed versions i've made XD if it's Vox/Western and it's entirely just your stuff i keep A LOT OF my lora squishes XD (I haven't tested vox+western outside of applying it to a model for that specific reason lol)

[pulsar gorge]

I might post 32dim version, slowly

[obtuse lark]

trying to load safetensor into the colab

[deep geyser]

probably the wrong link and it loaded a website instead

it's saying the file is invalid

Remember it has to be a direct download

[obtuse lark]

i see

[deep geyser]

drive or mega wont work

[obtuse lark]

i try linking it directly to the safetensor on my huggingface repo

if i dont it just gives me another error cuz it trys to save model as ckpt

[obtuse lark]

yea i use the url to the huggingface repo, it just doesnt work

[deep geyser]

huh

ping me about this later

[obtuse lark]

i put "https://huggingface.co/ModernNoob/Deyui_Model/blob/main/Deyui_Diffusion_Model-04.safetensors" as url

[deep geyser]

ah
I thought I was handling those

you need this link instead
https://huggingface.co/ModernNoob/Deyui_Model/resolve/main/Deyui_Diffusion_Model-04.safetensors

(right click the download button -> copy link)

[obtuse lark]

yea thats wut i put (without the quotes around url)

wait hold up give me a sec

[deep geyser]

notice it says "resolve" instead of "blob"

again, I thought I was replacing that automatically, guess not

[obtuse lark]

Oh i see

bro im blind as a fking bat

i should prob make a pr and just change blob to resolve for u :v

[obtuse lark]

same error

[deep geyser]

ill check

[obtuse lark]

aight

also, not sure if thats telling something

imma train on the default model on another dataset in the meantime.

[deep geyser]

I think the model is busted, idk

might also be for sd2.0

[obtuse lark]

yea, seem like it happens before too

did the custom url worked in the past?

[deep geyser]

it works generally

the key error means the model is invalid somehow

it's looking for something but it's not there

[obtuse lark]

can u send me an url that does work? I want to try to find a difference or something

[gentle kettle]

hello! i have dataset composed of images that is not 512x512. any tips? or guides for training lora with these images

i dont want to crop them too

[deep geyser]

@gentle kettle the training resolution doesn't matter, you can use any image sizes

[gentle kettle]

ohhh what resolution will I use during training for the LoRa?

[celest crater]

@deep geyser does your dadaptation implementation in the collab support "betas=0.9,0.99" as a param or is it not needed?

[deep geyser]

I think that's set as the default but not sure

[celest crater]

think the default only had decouple true and weight decay

[arctic marsh]

Hey, i get this pip error that says “pip’s dependency resolver does not take into ac… blah blah and 4 lines of respectively torchaudio, torchdata, torchtext and torchvision being incompatible(?)”

I am currently training even after this error (hopefully its actually training even after this error)

I can send the actual text if im asked.

Sorry I couldn’t search the thread

[deep geyser]

Hey, no worries, this is normal
The only downside is it takes longer to load

One day I'll fix that warning ...

[arctic marsh]

thats cool with me, i was worried from the “ERROR:” tbh

thanks for instant reply

[deep geyser]

Yeah it's spooky

[deep geyser]

I had notifications turned on it seems

[arctic marsh]

Actually, while at it, I would like to ask a few things to avoid multiple notifications. Of course, these are a lot of questions and will probably waste your time so please answer whenever you like or if you like to at all. Thanks for the notebook, too.

1 - If interrupted, can i just go start the same main cell and continue from where I left off without really changing anything?

2 - If I wanted to train with a lot more images, would lora no longer be a good choice? Lets say, 3-4k or more images.

3 - Since I downloaded the AnyLoRa, is it best if I use this lora by first loading AnyLoRa itself into webui or can I load any other model just fine?

4 - Is setting resolution to maximum possible good idea (dataset may have less than that resolution images) or will it exponentially slow down?

5 - Why is there a max total step limit? Funny but I literally reached there and changed how_many as a result. Is it an overcooking issue?

6 - When is more dim holding more information worse? Another overcooking or big file size lora that doesnt do any better issue?

Thanks in advance no matter how many you want to answer.

Wow thats a text wall, really really sorry

[deep geyser]

1. Training will start from scratch but it will skip the 3-4 minutes of installation.

2. For that you could look into LoHa but most likely you want a checkpoint. Otherwise 400 images is my personal limit. At some point it can't learn any more, so quality over quantity.

3. Loras work with any checkpoint, personally I like training on animefull then testing on anylora.

4. Multiplying the training resolution by 2 will multiply the training time by 4 or more. Stable Diffusion models don't really understand pixels, it's all math to them, so the training resolution is less important than you think. You only need 768, 1024 or more if you need to capture small details like accessories or pupils.

5. It was a common mistake to set one of the values wrong and end up with huge training times, I know the limit has saved me and others a few times. But the limit itself is unnecessary. Maybe you could train a loha for more than 10000 steps, but I try to encourage best practices and I don't think there's a good reason to go over in most cases. Same reason I don't have a VAE option in the trainer (they make things worse, last time we checked)

6. Large dim may introduce small visual artifacts, possibly because it makes stuff up to to fill all the data. Those are rarely perceptible, though. Regardless, dim32 for lora or dim16 for locon should be good for 99% of cases. dim64 may be useful for something. dim128 is the old way of doing things and I would advise against it. I know I've filled 10 different sets of clothes of an anime character into a single dim32 lora.

[deep geyser]

It's okay

[arctic marsh]

guess loras dont need multiple sessions training

[deep geyser]

I love helping and answering questions in this subject, I also have a ko-fi which keeps me motivated to do so thanks to generous peeps

[arctic marsh]

thats okay, thanks for your help

you deserve donations tbh

[deep geyser]

There are 2 ways to continue lora training, I implemented them both but they're so bad they're now hidden in the colab code

[arctic marsh]

i see, np

i will wait for any captchas lol

[deep geyser]

One way is to store the weights which is several GB, I ruled it out due to the limited available space and the weird latency on google drive
The other is to load the previous safetensors and continue from there, but it seemed to not work
Those were beyond my control, I don't know if it's been improved since then

[ashen swan]

Can somebody help, I am trying to train a Lora 321 Images, 1 Repeat, 6 Epoches, Total Steps are 1926 , My problem is I only get really messy stuff, However I've been using an Collab for a while and do around 10 Epoches most of the time 5000 Steps and It isn't some overcooked garbage

At around 1200 Steps

Learning Rate 0,00001, Text Encoder LR: 0,0001, UNET LR: 0,00001

DIM 16 Alpha 8

[normal stirrup]

Bad dataset i suppose

[ashen swan]

That's the weird thing, I've used the same datasets before and zero problems

[deep geyser]

that unet is too low

But I don't know why it would generate trash like that

[ashen swan]

What would you recommend ?

[deep geyser]

What model are you using to train and generate

[ashen swan]

NAI FULL Pruned

Is what I'm using to Train, AOM3 is what I use to generate

[deep geyser]

Something else must be wrong
Send the config file that was generated by the trainer in your Google Drive

There's 2

[ashen swan]

The Log File or the Settings I've used ?

[deep geyser]

settings

[ashen swan]

Trough Google Drive or here ?

[deep geyser]

send the 2 toml files here

[ashen swan]

I don't have any toml files

I'm gonna send this regardless

[deep geyser]

Oh I thought you were using my trainer

[ashen swan]

Oh

[ashen swan]

I have been using your trainer however I copied the values from your trainer into mine and I only get garbage

Even at 10 Epoches your trainer does not output anything bad, mine does and I don't know why

[normal stirrup]

style or character¿?

[ashen swan]

Character

[deep geyser]

I don't know what "epoch": 6 does
I think you want "max_train_epochs": 6

Also your lr and unet lr should be 1e-4 probably

[ashen swan]

Let me try that

What's the difference between Epoch and Max Epoch ?

[normal stirrup]

try adam8w
1e-4 lr/unet rate
5e-5 txt
32/32 dim
1rep
12 epoc

the rest looks ok

[deep geyser]

I don't know, I've never seen "epoch" as just 1 word in the settings file

[deep geyser]

These are good settings
I would personally do 32/16 and 10 epochs

[ashen swan]

Entering 0 for Epoch will give me none, so I may just add the same number as Max Epoch?

[deep geyser]

Also adamw8bit i suppose

[deep geyser]

Set epoch as none maybe

[ashen swan]

will only give me 321 Steps

[deep geyser]

Hmmmmmmmmmmmmmmmmmm

Ok big question which trainer are you using

[ashen swan]

https://github.com/bmaltais/kohya_ss

[deep geyser]

I would recommend this trainer
https://github.com/derrian-distro/LoRA_Easy_Training_Scripts

I haven't checked bmaltais recently so I can't comment on it

[little shuttle]

Very good trainer

Based holo

[ashen swan]

Looks nice, gonna give it a try

[ashen swan]

That's the one your trainer is using ?

[little shuttle]

Nope

[deep geyser]

We all use the same base trainer behind the scenes, but this one is by a friend
My theory is that something went wrong with your trainer somehow so it could work to try a different one

[normal stirrup]

try collab meanwhile

[ashen swan]

Huh I'm gonna give it a try eitherway

Yeah

Which version of Torch should I Install?, 1.12.1, 2.0.0, 2.0.1?

[deep geyser]

2.0.1 imo

It's faster

[ashen swan]

@deep geyser Works like a Charm, Thank You

[deep geyser]

[sand cypress]

is there any comparison showing the difference in training LORA from one model to another?

[deep geyser]

You could see the difference easily but I don't know of anyone showing side to side

[sand cypress]

animefull or anylora? which one do i use to anime lora?

dont know the difference ._.

[deep geyser]

animefull is the base model, I recommend it
anylora works too

I train on animefull then test in anylora

[sand cypress]

ok

[deep geyser]

My lora trainer now takes 1 minute to load again instead of 3, cause I fixed the dependency error
Enjoy

[frigid cloud]

This is the 1984th message

https://tenor.com/view/1984-gif-19260546

[torpid shard]

1987

[pallid musk]

This one is good.
Recently I'm using this and I did my sdxl also with this. I'm now liking this trainer more than normal khoya ss gui

[celest crater]

@deep geyser does the collab have dadapt v3?

[deep geyser]

Nope it uses v1.5

[celest crater]

does only the kohya trainer have it so far? i kinda wanna try it

[faint bronze]

welp tried using the new one but i cannot make it see my GPU, it keeps trying to train on my integrated GPU and fail because not enough memory

back to the old ways ig

[deep geyser]

2 months late but I did add it as well as prodigy

[simple haven]

What's the username/password to the generated gradio link? I didn't see a way to set my own credentials when running this colab

[celest crater]

1 month late but ive tried them and honestly saves so much effort ty for your work, wanted to ask as well how to bypass the total step limit you put in place since sometimes some styles will be underbaked or should i just pump up the lr?

[deep geyser]

You'd have to open the cell's code to remove that check

[celest crater]

[celest crater]

@deep geyser how long should this take?
found directory /content/drive/MyDrive/lora_training/datasets/Scott-Pilgrim contains 4664 image files
i know its a lot of images but i think it might actually be stuck

[deep geyser]

huh