ULTIMATE FREE LORA Training In Stable Diffusion! Less Than 7GB VRAM!

Summary

This video tutorial introduces 'Laura,' an efficient method for training AI models using personal images, optimized for devices with limited GPU resources. It combines elements of Dreambooth and textual inversion, enabling the creation of lightweight embeddings for various models. The guide walks viewers through the process using Koya SS GUI, from setting up the software to training a model with detailed steps and configurations. The result is a personalized AI model that can be mixed and matched for unique image generation, all achievable with minimal hardware requirements.

Takeaways

• 😀 Laura is a method for training AI models using personal images, optimized for small graphics cards with only 6-7 GB of VRAM.
• 🤖 Laura combines elements of Dreambooth and Textual Inversion, creating small files that can be used with any model, unlike larger Dream Booth models.
• 🚀 The video demonstrates using Koya SS GUI, a user-friendly software for training models like Dreambooth, Textual Inversion embeddings, and fine-tuning models.
• 🔧 The Koya SS GUI installation process requires Python, git, and specific Visual Studio redistributables, with detailed steps provided in the video.
• 📁 A specific folder structure is necessary for Laura training, including image, model, and log folders, with the image folder containing subfolders named after training step counts.
• 🖼️ Image preparation for Laura involves high-quality, varied images, resizing to 512x512 resolution, and using tools like blip for image captioning.
• ✍️ Captioning images for Laura requires adding the character's name to the beginning of the caption for more precise training results.
• 🔢 The number of training steps is determined by the formula of at least 100 steps per image, with a total minimum of 1500 steps, influencing the subfolder naming.
• 📝 Two configuration files are provided for training: one basic and one for systems with less than 8 GB of VRAM, simplifying the setup process.
• 🔧 Training parameters like batch size, learning rate, and resolution can be adjusted, with suggestions given for weak GPUs to use memory-efficient options.
• 🎨 The final Laura model can be used in Stable Diffusion with an extension, allowing the weight of the model to be adjusted in prompts for varied image results.

Q & A

What is Laura in the context of this video?

Laura is a method for training a subject using your own images, optimized for small graphics cards, requiring only 6 to 7 gigabytes of VRAM, making it accessible for users with limited GPU resources.

How does Laura compare to Dreambooth and Textual Inversion in terms of resource requirements?

Laura is more resource-friendly compared to Dreambooth and Textual Inversion, as it allows for training a subject with significantly less VRAM, making it suitable for users with less powerful GPUs.

What is the file size of the embeddings created by Laura?

The embeddings created by Laura are small, typically ranging between 300 megabytes and 100 megabytes, which is considerably less than the size of a Dream Booth model.

Can Laura embeddings be used with any model?

Yes, Laura embeddings can be applied to any model in the same way as Textual Inversion embeddings, making them versatile for various applications.

Why is the Dreambooth extension not used for training Laura in this video?

The Dreambooth extension is not used because it often does not work well with the automatic 11 due to new updates that can break the training process, which the presenter finds annoying and prefers to avoid.

What is the Koya SS GUI and how is it used in the video?

The Koya SS GUI is a software that simplifies the training process for a Dreambooth model, a checkpoint, and a texture inversion embedding, or even fine-tuning your own model. It is used in the video to train Laura efficiently and easily.

What are the system requirements for installing the Koya SS GUI?

To install the Koya SS GUI, one needs to have Python and git installed, as well as Visual Studio 2015, 2017, 2019, or 2022 with the distributable libraries. If not already installed, an Excel file is provided to assist with the installation.

How does one set up the Koya SS GUI for training?

The setup involves installing the necessary software, creating a new folder for the Koya installation, copying and pasting command lines into Windows Powershell, and following the prompts to complete the installation process.

What is the recommended image resolution for training with Laura?

The recommended image resolution for training with Laura is 512 by 512 pixels. However, for higher quality results with a powerful GPU, training at 768 by 768 pixels is suggested.

What is the minimum number of images recommended for training with Laura?

It is recommended to have at least 10 high-quality images for training with Laura. The number of training steps should be at least 100 steps per image, with a total of at least 1500 steps.

How can one use the trained Laura model in Stable Diffusion?

After training, the Laura model can be used in Stable Diffusion by copying the saved tensor file and pasting it into the Stable Diffusion UI folder's models/Laura directory. An extension called 'Koya SS Additional Networks' needs to be installed to utilize the model within the UI.

How does one mix different Laura models in a prompt for Stable Diffusion?

Multiple Laura models can be used together in a prompt by selecting them and adjusting their respective weights to control the emphasis on each model's influence on the final image, ensuring the total weight does not exceed one.

What is the impact of using a lower batch size during training?

Using a lower batch size, such as one, can improve the quality of training, especially when the number of images is limited. However, it may increase the training time and use less VRAM compared to higher batch sizes.

What are the benefits of training at a higher resolution like 768x768?

Training at a higher resolution such as 768x768 can result in better-looking models, especially when high-quality images are available. However, it requires a more powerful GPU due to increased VRAM usage.

How can one ensure the proper naming of folders for training steps in Laura?

The folder name should include the number of training steps per image followed by an underscore and the name of the character, ensuring the total training steps are at least 1500 and each image has at least 100 steps.

What are the configuration files provided in the description for?

The configuration files provided are used to automatically set optimal training parameters for Laura, including one for general use and another for systems with less than 8 gigabytes of VRAM.

What is the role of the 'enable buckets' option during training?

The 'enable buckets' option is used when training with images of different dimensions, ratios, and resolutions. It allows the training process to accommodate images that are not uniformly cropped or sized.

How can one modify the weight of a Laura model in a prompt?

The weight of a Laura model in a prompt can be modified by adjusting the number associated with the model's call-out text in the prompt. A value of 1 corresponds to 100% influence, 0.9 to 90%, and so on.

Outlines

00:00

🤖 Introduction to Laura AI Training Method

This paragraph introduces Laura, an AI training method that uses personal images and is optimized for small graphics cards, requiring only 6-7 gigabytes of VRAM. Laura is described as a hybrid between dreambooth and textual inversion, creating small files for use with any model. The video will not use the dreambooth extension due to compatibility issues with automatic updates, instead focusing on the Koya SS GUI, a user-friendly software for training models and fine-tuning. The paragraph concludes with acknowledgments to contributors who provided tips and created instructional videos for using the Koya SS GUI.

05:02

🛠️ Setting Up Koya SS GUI for AI Training

The paragraph outlines the process of installing the Koya SS GUI, starting with accessing the repository and ensuring prerequisites like Python and git are installed. It details the installation steps, including running commands in Windows Powershell as an administrator, and provides a link for downloading required libraries if not already installed. The process involves creating a new folder for the installation, copying and pasting code from GitHub, and following the installation prompts. An optional step is also mentioned for users with NVIDIA 30 or 40 series graphics cards to improve training speed.

10:04

🖼️ Preparing Images for Laura Training

This section explains the preparation of images for Laura training, emphasizing the importance of high-quality and varied images. It suggests using a website for resizing images to 512x512 resolution and manually modifying captions for each image to perfectly describe them. The process involves using a tool like blip for image captioning and then editing these captions to include the character's name for more precise training. The paragraph also describes creating a specific folder structure for training, including an images folder, a model folder, and a log folder, with the training steps determined by the number of images provided.

15:05

🔧 Configuring Training Parameters for Laura

The paragraph discusses configuring training parameters for Laura, starting with choosing a base stable diffusion model and selecting appropriate checkboxes based on the model's version. It introduces two configuration files provided by the video creator for optimal training settings, one for general use and another for systems with less than 8 gigabytes of VRAM. The video creator guides viewers on how to input folder URLs for images, outputs, and logs, and how to set the model output name. It also touches on various training parameters such as batch size, learning rate, and resolution, with advice on selecting settings based on the strength of the user's GPU.

20:06

🚀 Starting the Laura Training Process

This paragraph instructs viewers on initiating the Laura training process by clicking the 'train model' button after configuring all settings. It provides insights into training parameters like batch size, which affects training speed and VRAM usage, and the importance of selecting the right settings for users with limited VRAM. The paragraph also explains how to use the provided configuration files for easy setup and how to modify settings for low VRAM systems. The training time is highlighted, noting that 1500 steps with a batch size of 2 takes only 6 minutes, and the final result is a save tension file that can be used in stable diffusion with the help of an extension.

🎨 Using Trained Laura Models in Stable Diffusion

The final paragraph demonstrates how to use the trained Laura model in stable diffusion, detailing the process of copying the trained model file into the appropriate folder and installing the necessary extension. It explains how to integrate the Laura model into prompts by using additional networks and adjusting the model's weight to influence the image generation. The creator shows examples of image generation with different model weights and mixing two Laura models to create unique images, emphasizing the flexibility and power of Laura for quick AI training even with limited GPU resources.

Keywords

💡Laura

Laura is a method for training AI models using personal images, optimized for devices with limited GPU resources. It allows users to train a subject with as little as 6-7 gigabytes of VRAM, making it accessible for those without high-end graphics cards. In the video, Laura is described as a hybrid of Dreambooth and textual inversion techniques, creating small files for use in AI models, which is central to the video's theme of efficient AI training.

💡VRAM

VRAM, or Video Random Access Memory, refers to the dedicated memory used by a graphics processing unit (GPU) for storing image data. In the context of the video, VRAM is a critical resource for AI model training, with Laura requiring significantly less VRAM compared to other methods like Dream Wolf or textual inversion, thus democratizing the training process for users with smaller GPUs.

💡Dreambooth

Dreambooth is a technique for training AI models on custom datasets, allowing the creation of models that can generate images based on specific subjects or styles. The video script mentions Dreambooth as a comparison point for Laura, indicating that Laura is a more efficient method in terms of VRAM usage while still achieving similar results.

💡Textual Inversion

Textual inversion is a process in AI training where text embeddings are used to influence the generation of images. The video describes Laura as being similar to textual inversion in that it creates small files that can be used as embeddings in AI models, but with the added benefit of lower VRAM requirements.

💡Koya SS GUI

Koya SS GUI is a software tool mentioned in the video that simplifies the process of training AI models, including Dreambooth models, checkpoints, and embeddings like Laura. It is highlighted for its user-friendly interface and the ability to fine-tune models, making it a key component in the video's tutorial on training with Laura.

💡Training Parameters

Training parameters are the settings used to configure the AI model training process, such as batch size, learning rate, and resolution. The video script provides detailed instructions on how to set these parameters for optimal training with Laura, emphasizing the importance of these settings in achieving high-quality results.

💡Batch Size

Batch size in AI training refers to the number of samples processed before the model's internal parameters are updated. The video suggests choosing a batch size of one for a small number of high-quality images to improve training quality, or two by default for a balance between speed and VRAM usage.

💡Embeddings

In the context of AI, embeddings are vector representations of words, phrases, or images that capture their semantic meaning. The video explains that Laura creates small embeddings that can be applied to any model, similar to textual inversion, but with a smaller file size, which is crucial for efficient model usage.

💡Stable Diffusion

Stable Diffusion is a type of AI model used for generating images from text descriptions. The video mentions Stable Diffusion as the base model on which Laura embeddings can be applied, showcasing its versatility and the integration of Laura with existing AI technologies.

💡Extensions

In the video, extensions refer to additional features or functionalities that can be added to AI models or training software. The script discusses the use of extensions like 'safe tensors' with Stable Diffusion to enhance model performance, emphasizing the importance of selecting the right extensions for training stability.

💡Regularization

Regularization in AI training is a technique used to prevent overfitting by introducing additional information or constraints to the model. The video mentions that regularization images are not necessary for Laura training when training a single subject, indicating a simplification in the training process compared to other methods.

Highlights

Introduction of Laura, a method for training AI models using personal images with minimal VRAM requirements.

Laura is optimized for small graphics cards, requiring only 6-7 gigabytes of VRAM.

Comparison of Laura with Dreambooth and Textual Inversion, noting its smaller file size and flexibility.

Explanation of how Laura creates small 100-megabyte embeddings applicable to any model.

Mention of Koya SS GUI, a user-friendly software for training models and fine-tuning.

Details on the installation process of Koya SS GUI and prerequisites like Python and git.

Instructions for setting up the command line and executing the installation script.

Optional steps to improve training speed for NVIDIA 30 or 40 series graphics cards.

Description of the folder structure necessary for training with Laura.

Importance of image quality and variation for effective training.

Process of resizing images to 512x512 resolution for training.

Use of blip for image captioning and the manual modification of captions for precise training.

How to name training folders to determine training steps based on the number of images.

Preparation of configuration files for training parameters to simplify the process.

Selection of base stable diffusion models and the impact on training settings.

Recommendation to use 'safe tensors' for stable diffusion models.

Discussion on training parameters like batch size, learning rate, and resolution.

Techniques for mixing multiple Laura models to create unique images.

Demonstration of the final results and the capability to adjust model weights for different outcomes.

Conclusion emphasizing the power of Laura for quick training on weak GPUs.

Transcripts

and in this video I will show you how

you can train your own subject using

Laura now what is Laura well Laura is a

method of training your subject using

your own images that is optimized for

small graphics card meaning that

compared to dream wolf or textual

inversion you can train a subject with

only 6 to 7 gigabytes of vram which is a

super good news for all of you who don't

have a good GPU now basically Laura is

kinda like a love child between

dreambooth and textual inversion in the

sense that it creates these small 100

megabytes file that you can use in the

exact same way as textual new version

embeddings meaning that they can be

applied to any model and the size of

these embeddings usually range between

300 megabytes in 100 megabytes which is

definitely way less than a dream Booth

model and of course just like dreamwoof

and textual inversion you can train a

style or a character absolutely anything

you want now in this video I'm not going

to be using the dreambooth extension to

train Laura now I've already explained

why in my previous Dreamboat video then

the dreambooth extension often does not

work very well with the automatic 11

because of new updates so sometimes a

new update can completely break the

training and it's very annoying and I

personally don't want to deal with this

anymore but don't worry because we're

going to be using something even better

and that is the Koya SS GUI it is a

super cool piece of software that is

super easy to use where you can train a

dreamboof model and lower a checkpoint

and texture inversion embedding or even

fine-tune your own model and thanks to

the configuration files that I'm gonna

provide you in the description down

below it is super easy and fast to set

up so let's go but before we begin let

me thank two people for making this

video possible the first one is spy BG

who gave me a lot of tips and tricks on

how you can use Laura efficiently and

the second one is Bernard malte that

made two amazing videos explaining how

to create a lower weight using the Koya

SS GUI and that is also responsible for

making the Koya SS GUI that we're gonna

be using today so again the link for

their channels will be in the

description down below alright then now

let's begin alright so to install the

Koya SS GUI you're gonna click the link

to description down below you're gonna

arrive on this page on the Korea SS GUI

repository you're gonna scroll down

you're gonna make sure that you have

Python and git installed if you already

have stable division installed it should

be already there if you follow my

tutorial video but also make sure that

you have the visual studio 2015 2017 19

2022 with distributable already

installed if you don't have it installed

already you can click on this link right

here and it will download an Excel file

and then just run the Excel file to

install the required libraries it's very

easy very simple very fast nothing to

worry about so then once you have

everything installed you're gonna come

here and select this command line and

Ctrl C to copy it then you're gonna go

into your Windows startup menu and look

for Windows Powershell and then click on

run as administrator and then just paste

the command line that we copied

previously and then press enter it's

going to ask you if you want to change

the execution policy and here you're

gonna type

Aya that basically corresponds to yes

for all and then press enter and there

we go so now that this is done you can

close the window and we're going to

create a brand new folder on our

computer so right click new folder I'm

gonna call mine Koya which is where

we're gonna put the koyan installation

then go inside that folder click on the

folder URL Ctrl C to copy it then go

back to the Windows startup again look

for Powershell but this time no need to

run it as administrator you can just run

it normally and then what we need to do

is that we need to go and design the

folder that we created right here and to

do this you're gonna type CD space and

then paste your folder URL right here

now if you created a folder that is not

on the C drive instead of Simply putting

CD you're gonna have to add slash D and

then the path of your folder let's say

your folder is on your drive e this is

what this would look like but in my case

since this is on my C drive I don't need

the slash D so I'm simply going to be

using CD and then the URL of my folder

and then press enter and as you can see

now we are inside the folder that we

just created right here so now we go

back to GitHub and we're going to click

on this button right here to copy this

entire code and then we're gonna paste

the entire code in the center Windows

Powershell window and if it gives you a

warning you can just click on test

anyway and as you can see right now it

has started running all the lines of

code that we copy and pasted before now

this might take some time so be patient

it shouldn't take too long this is not a

stable diffusion installation after all

but for some of you it might still take

some time now at the end of the

installation it will stop at the

accelerate config line code and to

finish the installation you can just

press enter so then it's going to ask

you a few questions so for the first

question you're going to choose this

machine so here just press enter no

distributed training so press enter

again here you're gonna type no press

enter torch Dynamo you're gonna type no

again then press enter they want to use

deep speed you're going to press no then

press enter again What GPU should be

used you're gonna type all then press

enter Then you wish to use fp16 or bf16

for this one you're going to be choosing

fp16 so just press the down arrow and

then press enter and now finally as you

can see the installation is done now

technically we could stop right here but

there is an additional optional stand

that you can take that will improve the

training speed if you have a 30 or 40

series graphics card for NVIDIA but

don't worry it's very simple all you

have to do is just click on this link

right here and it will download a zip

file and if for some reason this link

does not work or is too slow I will also

give you a mega link that you can use

instead so once you've downloaded the

zip archive on your video you just right

click and then extract it then you're

gonna go inside that folder select the

cud and N Windows Ctrl C to copy it then

go inside the Koya SS folder and paste

the folder right here then we go back to

GitHub click on this button right here

to copy these two lines of code go back

to the windows Powershell window and

then paste the lines of codes right here

and again if it gives you a warning you

can just click pay list anyway and then

press enter and as you can see now

everything is done we did the most

difficult part of this video

congratulations and now if we go inside

our Konya SS folder you will see a bunch

of files but don't worry because

actually there is only two files that

interest us that is the goo either but

in the upgrade.ps one for example next

time that you want to update Koya SS all

you have to do is just come here on the

upgrade.ps1 right click and then run

with Powershell and as you can see it

will automatically update the Korea SS

folder and then finally if we want to

run the Korea SS GUI all you have to do

is just double click on the gui.bat file

and to load the CSS and just like stable

diffusion it will give you a local URL

that you can easily open in your browser

but just holding down control and then

click on the local URL and there you go

now we are inside the Koya SS GUI and we

can finally begin the dreambooth Laura

training now obviously before we can

begin the training there is a few steps

that we need to do before and the first

step is of course to prepare our images

now will not be going into too much

details about that in this video because

I've already talked about it in my

previous text uni version video so

definitely watch this video before

watching this one or at least watch the

beginning of the video where I explain

how to perfectly caption an image

because the process is pretty much the

exact same but if I had to summarize

first make sure that all the images are

of high quality and that the images have

a lot of variation so you have different

lights in different angles make sure to

have at least 10 images as long as they

are of high quality it's better to have

a small amount of images of high quality

than a lot of images with bad quality

then you can use a website like berm.net

to resize your images to 512 by 512

resolution although for lower training

you don't necessarily need to resize any

images you can use images of any

resolution but if you're not confident

and you want to make sure that

everything works well just process the

images manually it will work absolutely

fine then you're gonna capture each

image using a process like blip that

will do like 80 of the world work for

you and you can do this process either

inside stable diffusion or inside the

koyagui by going into utilities and then

believe captioning it will basically do

the exact same thing it will analyze

every image and then create a text file

with the same name as your image and

then for each image you're gonna have to

go and manually modify the captioning so

that it perfectly describes the picture

now the small difference compared to

textual inversion captioning is that in

the beginning of the captioning you're

going to input the name of your

character now you don't necessarily need

to do it you could simply leave it like

that so that the next time that you use

the Laura embedding every women in your

prompt will look like your character but

if you want more Precision for your

prompt you can simply input the name of

your character in the beginning of the

caption this way when you use your name

in The Prompt the name of your character

in the lower embedding will be linked

together but that's not a big deal if

you don't want to do it it's really up

to you so now that our images are

prepared we're gonna have to create a

specific folder structure so this is

also something that is different because

compared to the textual inversion

training but again don't worry it is

fairly easy all you have to do is just

right click new folder I'm not going to

call mine Wednesday Adams Laura I'm

going to select all the files and images

Ctrl C to copy it then going to send the

Wednesday Adam's Laura folder and here

we're going to create three folder an

image folder a model folder and finally

a log folder now we're gonna go and send

the image folder and again here we're

going to create a new folder but this

time this will be a little special

because the way we name our folder will

determine the amount of training steps

that the koyagui will do and this will

depend on the amount of images that you

have so first you need to make sure that

you do at least 100 steps per image with

a final training steps of at least 1500

so for example if you only have 10

images and you need at least 1500 steps

of training you're gonna take 1500

divided by the number of images that you

have and it's going to give you 150 and

this number is the amount of training

Style types per image that we need to

input into folder name so for this

you're going to right click new folder

and you're gonna type 150 underscore and

then the name of your character so in my

case it's Wednesday Adams then you're

going inside that folder and you paste

your images right here so as I said

previously this number right here is

what determines the amount of training

steps per image that the Lora training

will do so in this example I only give

you a number if you only had 10 images

but let's say you add 20 images for

example well in that case there would be

1500 divided by 20 and gives you 75 but

as I said this number should be at least

100 so if you already have 20 images you

don't need to do this math anymore and

you can just input instead of 150 100

same if you have 25 images 30 images etc

etc this number should be at least 100

but if you have less than 15 images you

need to do this little math and for

example simple in my case I only have 11

images 11 is less than 15 so I'm gonna

have to do some math so 1500 divided by

11 gives me 136 well this is the number

that I'm going to be using for the

folder name so again instead of 100 I'm

gonna put

136 and now finally after we've prepared

your images after we prepare the folder

structures we can finally start the last

step which is training and for this I

will make your life very simple because

I prepared for you two configuration

files that you can use to automatically

determine the training parameters so for

this you're gonna click the link in

description down below and you're gonna

have two Json files the first is the

Lora basic settings which is basically

the basic lower settings that work with

pretty much every single training and

the second one is a special settings

file that you can use if you have less

than 8 gigabytes of vram so just

download these two files put them

somewhere safe then you're gonna come

here click on the configuration file

button click on open and then choose one

of these two files so here I'm just

gonna choose the Laura basic settings

file and then click open and if you

click on training parameters you can see

that some of these options were already

done for you these are the most optimal

settings that work for the Laura

training and that you can use as is so

if you don't want to waste time this is

a super quick way to start a training

but I'm also gonna explain a few

settings that you might want to modify

the padding on your training now first

before we begin let's actually choose

our base stable diffusion model

basically what stable diffusion model

are you going to be using to train a

Lora model and here in the model quick

pick you have already a bunch of

pre-selected models the 1.4 the 1.5 the

2.0 2.1 the 2.0 512 by 512 and the 2.1

512 by 512 and if you select one of

these models it will be automatically

downloaded from GitHub and also

depending on the model that you select

some options will be selected so if you

select for example the 1.5 you will see

that this checkbox will not be selected

because this is not a V2 model however

if you select the 2.0 base you will see

that this checkbox is now checked and if

you choose the 2.0 768 version you will

see that now this V parameterization

checkbox is checked also now you need to

keep these checkboxes in mind because if

you come here and click on custom you

can actually choose your own model and

if you choose a custom model you need to

know what model this was based on so for

example if I come here and I choose the

protogen V 2.2 model I know that this

model is based on the 1.5 meaning that I

don't need to check any of these

checkboxes however if I select the model

that is based on the 2.0 model I need to

check this checkbox right here and if

it's based on the 2.0 768 version I need

to check this additional checkbox right

here so this is something to keep in

mind but I personally think that most of

you will use the 1.5 model as a base for

training so you will probably not need

to use any of these check boxes and here

with the safe train model as we have a

bunch of extensions that you can choose

from but of course I highly recommend

that you keep safe tensors because it is

now the safest extension for stable

diffusion models so whenever you have

the choice just choose saved answers so

now you're gonna click on the folders

Tab and here for each section you're

going to input the folder URL so for

example for my image folder you're going

to click here and as you remember we

created a brand new folder for that and

mine is inside Wednesday Adam's Laura

folder and this is the folder that we

want to use do not go inside that folder

you need to select this image folder

right here so just click select folder

for the output folder mine is in

Wednesday Addams Laura and then select

folder and same thing with the login

folder once they add UPS Laura and then

sync folder now you can also choose a

regularization folder where you have

your regularization images but for lower

training you don't really need

regularization images since you're going

to be training one subject anyway so

here you're gonna input the model output

name so in my case I will put Wednesday

Addams then click on the training

parameters and here you're gonna see a

bunch of options but as I said since we

have loaded the configuration file you

don't actually need to touch anything if

you want to start the training right now

all you have to do is just click on the

train model button and you are done but

let's go through some of these training

parameters anyway so the train batch

size is basically the amount of images

that you're gonna train at one time so

the higher the number the faster the

training will be because you basically

divided the amount of training steps by

this number but here's a little advice

if you don't have a lot of images if you

only have like 10 or 15 images I highly

suggest that you choose a batch size of

one it actually improve a little bit the

quality of the training sure it might

take longer but the final result will be

worth it and of course a higher batch

size will also increase the amount of

vram used for the training so if you

have a weak graphics card I suggest that

you put a batch size of one otherwise

you just leave it at 2 by default so

when it comes to everything right here

you're gonna leave everything by default

the learning rate is really good you

don't really need to change anything

this basic settings basically work for

every single training so the max

resolution this is the resolution that

you're going to be training your images

at so if you only have 512 by 5 12

images You're Gonna Leave It 512 by 512

by default but if you can this is my

advice if you have a good GPU instead of

512 or 512 try training this at 768 by

768 yes you will use way more vram but

if you have high quality images your

final model will look way better

otherwise if you have a weak GPU just

leave it at 512x512 by default also

right here the enable buckets if you

often crop your images yourself and you

have images of different dimensions and

ratio and resolutions you should enable

this option right here this will make it

so it can train images with different

resolutions otherwise if you just

cropped it yourself you can just disable

this option right here so again here you

can leave everything else by default but

if you have a weak GPU I suggest that

you enable memory efficient attention

and and gradient checkpointing this

might increase the training time but

will also use less vram and if you're

lazy like me and you don't want to

enable all of these options each time

that you want to train if you have a

weak GPU you can just come here in the

configuration file click on open and

then select the lower end of your arm

settings file and then click open and as

you can see all the low vram options

will be selected automatically and then

guess what now we are completely done

and if we want to begin the training

just scroll down and click on train

model and as you can see it will finally

start the training and for 1500 steps

with a batch size of 2 this training

takes only 6 minutes which is really

super super cool and then after the

training is complete if you go inside

the model folder you're gonna see your

final save tension file right here and

to use it inside stable diffusion you're

gonna select it Ctrl C to copy it go

inside your stable diffusion with UI

folder models Laura and then paste your

file right here and then you can launch

stable diffusion and not to be able to

use the lower weight that we train

inside koyagui we need to install a

special extension so for this you can

click on extensions click on available

load from then you're gonna scroll down

and look for Konya SS additional

networks and then click install then

you're gonna click on installed and then

click apply and restart UI and once you

see the additional networks tab that

means that the extension was installed

correctly and then finally to be able to

use Laura inside your prompt you're

going to select your model write your

prompt and then you're going to click on

this button right here to show the extra

networks and then choose Laura as you

can see right here you have all the

lower weight that we created previously

and if you want to use one of them

inside your prompt you can just click on

it and as you can see this line will

then appear inside your prompt then you

can then select and put it in the

beginning of your prompt and the way it

works is that this is the text that's

going to use to call out the lower

weight called Wednesday Addams and the

number that you see right here is the

weight of the model so exactly as if you

would select a prompt tag and put

brackets or or a higher number right

here this will use a higher percentage

of the model the one corresponds to a

hundred percent a 0.9 at 90 percent 0.8

at eighty percent

etc etc and you can create some really

cool images just by modifying the weight

right here and also what's super cool is

that you can use multiple lauraway

together for example if I click on

another one and I put it right together

we are now using two different lauraway

together but we need to make sure that

the number right here when put together

is not over one so if I use for example

0.5 here I'm gonna have to use 0.5 right

here if I use 0.2 right here I would

have to use 0 for the hand right here

and the bigger the number the more

emphasis it's gonna put on this slower

weight but for this example let's

actually do a normal one just keep it at

one then click on close and then finally

click on generate and this is the final

result looks pretty good and as you saw

with only 6 minutes of training with

less than 8 gigabytes of error M you can

have a results like this and that's

really super powerful and as I said if

we keep for example the same seed and if

you want to decrease for example the

weight of our model if I put something

like 0.8 and then click on generate

you're gonna have now a slightly

different image maybe better maybe worse

it kind of depends on what you're

looking for and as I showed you

previously you can easily mix it with

another Laura model so if I take this

one for example we just trained on top

of a Shelby if I put 60 percent of

Wednesday atoms and 40 of Thomas Shelby

using the same exact seed and I click on

generate it gives me this image and as

you can see we are slowly starting to

have a little bit more manly look and

with brighter eyes and that is because

Thomas Shelby has blue eyes so obviously

you're gonna have this look influence

into the image and if we go even higher

you'll see that our character is

starting to look more and more like

Thomas Shelby and less like Wednesday

Adams but it's still super super cool

because you can definitely mix and match

a lot of different lower weight to

create very interesting looking images

and there we have it folks now you can

train any subject you want in only a few

minutes with a very weak GPU all of that

thanks to Laura and there you go thank

you guys so much for watching don't

forget to subscribe and smash the like

button for the YouTube algorithm and

I'll see you guys next time bye bye