Greetings everyone. In this video, I will show you how to install
and use DreamBooth extension of Automatic1111 web UI from scratch. How to generate amazing quality images like
this. On the left, it is my real image. On the right, we are seeing the raw output
of my trained DreamBooth model. Here again, on the left, we have my real image. On the right, we have the raw output of DreamBooth
model. And here on the left, my real image and on
the right, this is inpainted output of the DreamBooth model. I will show all of them. By following this tutorial, you will be able
to generate studio quality images that you can use on your LinkedIn profile. You can use on your Twitter profile. You can use on your GitHub profile. You can possibly use on your official documents
or wherever you need studio shot quality images. This will save tons of time of yours. This will save you huge amount of money. So this tutorial will teach you how to obtain
amazing realism. So I have prepared an amazing GitHub readme
file for this tutorial. Every information, links and commands will
be in this file. This file will be the primary source. As it be necessary, I will update this readme
file. The link for this readme file will be in the
description of the video and also in the comment section of the video. When you click here, you will get to my Stable
Diffusion repository. I am posting a lot of useful stuff here. So please star it, fork it and watch it and
check out the links and descriptions that I put here. So we will begin with installing the DreamBooth
extension of Automatic1111 web UI. If you don't know what is DreamBooth training
and if you wonder and if you want to learn more details, watch this tutorial that I have
put the link here. To be able to follow this tutorial, you need
to have Python and Git installed. If you don't know how to install Python and
Git, look out for this tutorial video that I linked here. You can download the Python 3.10.9 version
from here and Git from here. If you don't know how to install Automatic1111
Stable Diffusion web UI, I made a recent video and in the first 15 minutes, I have explained
how to install it very detailed. So watch this tutorial for that. We will use realistic vision version 2 for
this tutorial as a base training model. The link is here to download it. So I will show both automatic installation
and manual installation. I have prepared an auto-installer script. When you open this link, you will get to this:
Patreon post that I have downloaded part 1 and part 2.bat files. Don't worry, I will also show manual installation
as well step by step. So let's make a new folder in my c drive. DreamBooth tutorial. Let's make two folders, auto, manual. For auto installation, copy paste the downloaded
bat files here. First execute part 1. It will ask you to run it. Click more info, run anyway. You can right click the bat file, edit and
see what is inside it like this. So here I have my manual installation of Stable
Diffusion web UI. Let's say you want to use the exact version
that I have used for this tutorial. Then after cloning your Automatic1111 web
UI, execute this git checkout command. Open a cmd window inside here, copy paste
it, and you will get to this checkpoint. Then do a fresh installation. So download the realistic vision model from
this link. When you click it, you will see it will start
downloading like this. Put the downloaded realistic vision model
into your models folder. Don't forget to join our Discord channel for
any kind of help and also discussions with other expert people we have. Click this link. From here join our server. Automatic installation part 1 has been completed. I closed the cmd window. Then start part 2. Click more info, run anyway. When you see the url like this, it means that
the first part of the automatic installation has been completed. And now automatic installation will automatically
clone the DreamBooth extension and install everything automatically for you. Manual installation is also ready. So let's say you already have an existing
Automatic1111 web UI instance. You don't have to do a fresh installation. All you need to do is start a cmd window,
do git pull which will update the Automatic1111 web UI to the latest version. If you get a message like this, then do git
checkout master and then do git pull again. And it will update everything to the latest. Let's say you have previously installed DreamBooth
extension. Then go to the extensions folder and delete
it. This is important. You have to be inside this folder. Open a new cmd window. Copy the command written on the readme file
here. Paste it and hit enter and it will clone it. Let's say you want to use the version I used
in this tutorial. Then what you need to do is executing this
command, copy it, move into sd DreamBooth extension folder like this, then execute here. And then you will use the same version that
I used for this tutorial. Why this is important? Because sometimes the extensions may get broken
temporarily and when you are watching this tutorial, if it doesn't work, you can use
the exact versions that I have used. Then close this window, move into your main
installation folder, go to virtual environment folder venv, go to inside scripts folder,
open a new cmd window here, copy this, activate and type it and it will be activated. You see now the virtual environment is activated. I have explained a lot of details about virtual
environments in this tutorial video. Then execute these commands one by one. Copy this, paste, hit enter, paste, hit enter,
then copy this, paste, hit enter, and then copy this, and paste, hit enter. And it will install all of the requirements
of the DreamBooth extension. This is the most proper way of installing
DreamBooth extension. Do not install it from extensions tab of the
Automatic1111 web UI. Close your Automatic1111 web UI and follow
the steps that I have shown here. So what is the benefit of using my automatic
installer? If there be an error. Just leave a comment here and I will update
the scripts and you will get the most up-to-date fixed script file and you won't have any issues. Alternatively, leave a comment to this video
and hopefully I will update this readme file and fix the issue for you. So now we can use both automatic and manual
installation. I will continue with automatic Stable Diffusion
and DreamBooth installation. I will start my web UI instance. The web UI has started. You see the commit hash and the DreamBooth
hash revisions here. When you open your interface, you will see
the DreamBooth tab. So let's begin with creating our model. Click the create tab, give a name, any name. This doesn't matter. Me Best realism: select your source checkpoint. Click refresh icon here If you don't see the
model. Select realistic vision. This is for realism. Check unfreeze model. You may wonder why check this? because this
is slightly improving the quality. I have done over 30 different experimentation
to find the best parameters. Hit create model: follow what is happening
in your cmd window. This is important. Also, in the very bottom of my interface you
see it displays version of the Automatic1111 web UI, python, torch, and xFormers. Currently my xFormers version is the latest
version. Why? Because the automatic script automatically
upgrades the xFormers to the latest version. If you don't know how to upgrade it to latest
version, watch this tutorial. If you don't see any progress in your cmd
window, hit enter because it means that it was frozen and then it will continue to display
messages and continue operation. Okay checkpoint successfully extracted. You can see where the extracted model files
are stored. It is inside in this folder. Let me open it. So here the extracted model files. This is where Stable Diffusion training model
files are saved in this folder. You can see with the name of your training
model and other variables are her. In this folder you will find the native diffuser
files. So based on these diffuser files it will continue
do training, continue doing fine tuning. You see all of the files are here. Okay, then select your model. When you first time create your model, it
will automatically select the freshly composed model like this. Later time if you want to continue then you
need to select your model from here. It shows the loaded model, model revision,
model epoch, and the source checkpoint used to compose this model. I will show the best settings for the best
possible realism. This is usually best settings for fine tuning
and other tasks as well. So what matters for best realism is the used
base model, used resolution and used images quality. I will explain all of them. First click performance wizard then click
train wizard person. I train up to 200 epochs. Save model frequency is 10. So with each 10 epochs we will save our model. I am not saving any previews anymore, it is
just slowing us down. I use x/y/z plot to find the best checkpoint. I will show you how to do that. For training the best possible realism I suggest
you to use batch size 1 and gradient accumulation steps 1. Class batch size doesn't matter because you
should never use DreamBooth extension to generate your classification images. Either you should use pre-made classification
images or use text to image to generate your classification images. I don't check gradient checkpointing because
I have 24 gigabyte VRAM having gpu, but I will also show the best settings for 12 gigabyte
having gpus as well. So the learning rate will be 1e-7 this is
important. Make the learning rate 1e-7 both for UNET
and text encoder learning 1e-7. Image processing. Now this is really important. As higher as you go with training max resolution
you will get better results, but max resolution also depends on your base trained model. Realistic vision has a very good resolution,
but probably the max we can go is 1024 so make this 1024 like this. Do not select apply horizontal flip or dynamic
image normalization. I tested this and they are decreasing the
quality. Use EMA. Select Lion. I tested all of this and Lion is the best
one that's for sure. Select mixed precision bf16. If your graphic card do not support this. You can also select fp16. I will continue with bf16. Do not use xFormers. This is also decreasing the training quality. Cache latents. Train UNET. Step ratio of text encoder is 75%. Offset noise is used to have more contrast
of the output images. I don't find this useful for training yourself,
but you can try this with like this. Freeze clip normalizations layers. This is not also improving so do not check
it. Clip skip 1 for training yourself. Now this is important. Weight decay. This will be 10%. You see I did set a lower learning rate and
higher weight decay. Why? Because I am using Lion optimizer. So these learning rates and these weight decays
are set for Lion. So apply everything like I am doing. Do not touch anything else here. Keep them as they are. Now concept tabs. This is really important. In the dataset directory we will set our training
images folder path. Here I have preprocessed different dimension
images of myself. How did I prepare them? I used my human cropping script which I have
shown in this video and prepared automatically different aspect ratios. For most realism we need the highest resolution
with 768 by 1024. These are the raw images. They are bigger than 768 and 1024. Then I used the train tab. In here, preprocess images. Set the source directory. For example let's set this like this and here
let's set another folder as a destination. So I will set the destination here. Let's make it as new train. Set the width and height of the images. Select auto focal point crop, focal point
face weight 1, and preprocess. I am not using captions for training a person. It is not improving the realism. So here now we have our new training images. They are processed. They were already properly cropped for the
correct aspect ratio. So when I click view details and sort by width
and height, we can see the width and height of the images like this. They are ready to be used. So the most important part of DreamBooth training
for realism is the training images data set. This data set is at best medium quality. Why? Because I have repeating backgrounds and repeating
clothing. So as you have more diverse backgrounds and
clothing, it is better. Whatever is repeating in the training images,
the model will memorize them. Therefore you will have lesser flexibility. However, even this data set works very well
for realism. So if you use a better data set then it will
be better. In your images the sharpness of the image
should be very good. It must have very high quality. It must have good lightning. So capture your training images during daytime
when you have sufficient amount of lightning. This is really important. Your images should be sharp. Your images should be high quality. Sharp focus, clear, not blurry, not anything
else. So the most important part, as I said, the
training images quality. Prepare different backgrounds and clothing
having images. This is at best a medium quality data set. Okay, let's return back to our DreamBooth
tab. Let's select data set directory. So I copy here, paste it here and training
images data set is ready. Then classification images. This is a commonly asked question to me. What is classification images for. The classification images will prevent model
from overtraining and also it is used to keep the model previous knowledge. Therefore, normally you need to generate your
classification images by using the model itself. Let's say we are going to use class prompt
as a man and you may wonder, what is this man, what is this classification images and
everything? I explain all these concepts in this amazing
tutorial video. So watch this video to learn more about classification
images and rare tokens and other concepts. Since our target is realism, I will fine tune
the model with real images. So with the classification of real images,
the model will become even more realistic and I will get better results. So here my preprocessed images. I will use 768 by 1024 pixel images. So here my data set. These are collected from unsplash.com. These are available to use even for commercial
purposes. You see these are real images. They are super high quality. Let me open some of them for you. Here an example image. Here another example image. By using these real images, I will improve
the realism of my model and I will prevent overtraining. So with these images, I will obtain the maximum
realism from this training. This classification images data set is shared
on this post. You can alternatively also watch this tutorial
video and prepare your own classification images data set as well. Let's say you want to use the model itself
classification images. Then you can generate classification images
with photo of man, set the resolution the half of the target resolution so it will be
384 by 512. Then apply high resolution fix why? Because this way you will get higher resolution,
better quality classification images data set. For example this is an example image but it
has some noise as you are seeing with high resolution data set. So right click here and generate forever as
many as classification images you want, then move them into another folder and give its
path. However for realism, real images performing
better. So I will copy the path of this classification
images data set here. I don't use file words instance token or class
token. I use training prompt so it will be ohwx man,
this will be my instance prompt. Ohwx is our rare token. Man is our class token. So we teach ourselves into ohwx rare token
and we leverage the knowledge of man token of the model. The class prompt will be also man. Because my classification images are all man. Let's say you also wanted to use captions
for your images. Then all you need to do is adding here [filewords]. So it will read the captions and append them
to here and this will be your final instance prompt when during training. You can prepare captions from train, preprocess
and here you can use BLIP for caption or deepbooru for caption. For example, let me show you use BLIP for
caption. There are also other captionings, but I find
the captioning is necessary for fine tuning, not teaching a style, or for not teaching
a subject. In this case, we are teaching ourselves as
subject. Therefore, I don't find using captions is
better. So new train2, preprocess. When you first time use captioning, it will
download the used model. In this case it is BLIP and here our second
training data set with captions. When you open caption file you will see a
description, a man with glasses and a blue shirt is standing in front of a building and
a street light and street light. So you see the class token is already exists
in my caption. Therefore, if class token exists in all of
the captions, remove it from here. So when this file is read, what will be the
instance prompt. The instance prompt will be equal to this
one. It will just append the file words to here. However, as I said, I don't find it better
to use because it will diverse your training and it will take your model away from realism. In realism training, we want to model to learn
as much as the details of our face. And the class prompt is man and class images
per instance. Now this is important. I use 100. Therefore, since I have 13 training images,
I need 1300 classification images. How many I have in my folder? I have 2700 classification images. Therefore, calculate your number of classification
images that you need. Generate them from here. Open the folder, look how many you have, move
them into another folder and use them as a classification images. However, you won't get as good as results
as using the real images. Because when we use real images like these
are. Actually we are further fine-tuning the base
model with these new images for this token. So it becomes even more realistic. Better quality for man token. I don't use any sample prompts because I will
check the each checkpoint with x/y/z plot I will show. So these are the settings of concept tab. Data set directory for training images. Classification dataset directory for classification,
regularization images. Filewords these are empty. Training prompts ohwx man. Class prompt man. If you are training woman, then this is woman. If you are training girl, then this is girl. If you are training boy, then this is boy. Moreover, you should also clear the bad images
from your classification images. In the saving tab do not save EMA weights
for generated model. This is not improving the quality. Generate a ckpt file when saving during training. By the way, not anymore it is generating ckpt. It is generating safetensors file. It is same but this is the safe version. So these are the saving options. Always save settings, then click load settings. After loading settings verify that all of
them are correct or not. Training steps per epochs, same model frequency:
10. Batch size 1. Gradient 1. Set gradients to none when zeroing. Gradient checkpointing not selected. Learning rate, text encoder rate, max resolution
this is important. Make sure that it is 1024. Also, make sure that your classification images
and training images are also at this resolution. If you don't increase this resolution, then
the DreamBooth extension will downscale them first. Then start training. So you won't get this much high resolution
quality. Use EMA. This is important. Lion bf16, memory attention default, cache
latents, train UNET. Offset noise 0. Clip skip 1. Weight decay 10%. Text encoder weight decay 10%. I also tested this text encoder gradient clip
normalization. This didn't improve my results. Concept tabs looking correct.Training images,
classification images, dataset training prompts correct. Don't use other words here. Because rare token logic is very important. And as I said, if you don't know rare token
logic, please watch this video. This will help you tremendously. Okay, everything is looking good then hit
train. Currently my VRAM usage is like this as you
are seeing right now. First it will start with preprocessing and
looking for the classification images whether they are correct dimension or not and every
one of them is in the correct dimension so it should work very well. First, it will look for the classification
images and preprocess them if they are not in the correct dimension. It took like four to five seconds to preprocess
my classification images. It says that it has found 1300 regularization
images which was the necessary number. Why? Because we have 13 training images and we
are using 100 classification images per instance image. Therefore, we need 1300 classification regularization
images. Since it has found all of the necessary classification
images, it didn't generate any new ones. However, if we didn't have sufficient amount
of them, it would use these settings to generate those missing number of classification images
based on the class prompt here. However, as I said, do not depend on this. Use text to image tab to generate your classification
images. It is better. Always watch this cmd window to see what is
happening. This rule applies both on windows on linux,
on RunPod, wherever you are doing your training, always look at the cmd window that you started
your web UI. Now it is caching latents. It is caching both classification images and
training images. Therefore, 1313 images are being cached in
the gpu VRAM memory. So caching latents will increase your VRAM
memory usage. Caching latents has been completed and now
it started training. You see the training will going to take about
6 hours to be completed for 200 epochs. It is using 19 gigabytes VRAM. Why? Because we are not using gradient checkpointing,
we are not using xFormers and we are using 1024 pixels resolution and we are using EMA. All of these are increasing the VRAM usage
and decreasing our training speed. However, these are necessary for maximum quality
training, and when you consider that you don't have to go to a photo studio anymore, this
is much more convenient both economically and from both timing and the results perspective. By the way, currently I am recording a video. Therefore it is also using a lot of GPU and
VRAM of the GPU as well. Moreover, I am using Nvidia broadcast for
reducing the noise and this is also using a lot of GPU. What if if you don't have such powerful GPU
for such quality training, don't you worry. I have two excellent tutorials that shows
you how to use Stable Diffusion and DreamBooth on RunPod. First one is this one and second one is this
one. I even have automatic DreamBooth installer
on RunPod. In this post it is shared. Follow this video to install DreamBooth on
RunPod. For doing this training you don't even need
a powerful GPU from RunPod. Just rent RTX3090. It is 29 cents per hour and with like spending
two or three dollars you can have your amazing trained model, then download it and use it
on your computer or on Google Colab. Wherever you want after you have done the
training. Everything is same on RunPod. Just watch these two tutorials and you will
be able to do everything I have shown in this video also on RunPod. Meanwhile our training is continuing. I had done exact same training previously
me best. So I have a lot of checkpoints and I have
no idea which one is the best. Therefore now I will show you checkpoint comparison. Select like the middle like 100 epoch like
this. Decide your prompt. I will use this prompt. This is from the readme file. I will use this negative prompt. Now since we have done our training in 1024
you should select with 768 and height 1024. This is really important and that's all you
need. Now generate some images and we will pick
good seeds to do x/y/z comparison. So I am making batch count 100. We got the 100 output. Now we will select few good seeds and we will
do x/y/z checkpoint comparison. So let's find some good seeds for comparing. Okay, this is a decent seed so I will look
for the seed and save it. This is the seed. I copy paste it into a notepad file. Let's continue. Okay this is another decent seed. Let's also copy it. Paste it here. Continue. You can find up to 9 seeds or 4 seeds. It's up to you. Wow this is very interesting one. The face is almost like me but it has added
more hair. I will test this. Here the seed. This usually don't happen when you aim for
realism. The face is very good and it has also extra
hair. Okay this is another decent seed. So I have got four seeds for experimentation. Then make the batch size 1. Go to scripts: x/y/z plot, copy, paste the
seeds like this. Select the checkpoint names. From here you should select the checkpoints
with the order that will be easier for you. So I will start from epoch 10. Why epoch 10 is 260? Because I have 13 training images. Therefore for one epoch it has to do 13 steps
because in each step since batch size is 1, it will train one image. But I also have classification images. Thus, it is double of the required step. So 260 is 10 epoch, then 20 epoch, 30 epoch
40, 50, 60, 70. I am selecting all of the epochs with the
order they are like this. Up to 200 epochs. So all epochs are selected. Make the grid margins like 100. It will be easier to see and nothing else. Hit generate. Now it will generate x/y/z plot comparison
with the selected seeds so we will be able to see which checkpoint is performing best. You see in the first checkpoint. It is not even like me. As the training continues and more epochs
are trained, the images will become more like me. When doing x/y/z comparison, make sure that
the VAE file and the model files are correctly loaded. There aren't any errors. Also, if you have noticed I am using VAE ft
840000 ema pruned ckpt as VAE file. This is the best VAE file. It will significantly improve your image generation
output. This is not used during training, it is used
during inference. If you don't know how to install and use that
VAE file. If you don't know how to set these quick settings,
then watch the first 15 minutes of this tutorial and you will learn all of them. The x/y/z plot generation is completed. When we click the image, it will show us the
entire image. However, looking it from the folder is better. Click this folder icon. You see it has opened text to image images. But we need to check out text to image grids. Here. The grid file will be saved here. So in the very bottom we see our latest generated
grid file. Let's open it inpaint.net and let's zoom it
to original resolution. So it starts from the epoch 10. Let's look epoch 10 nothing like me. This is epoch 20. This is epoch 30. In epoch 30 a little bit like me. Epoch 40 you are seeing right now. Epoch 50. Okay, epoch 60. The face is very similar to me. 70 yeah, not bad. 80. This is epoch 80. Epoch 90 as you are seeing right now: epoch
100. This is the epoch 100. Epoch 110. This is the epoch we used to find these seeds
as you are seeing right now they are the same, okay 120 epoch. So by looking all of this you need to find
your best checkpoint. Then we will use that checkpoint to generate
images. You see as the epochs goes more it is changing
and it is more becoming like me as you are seeing right now. This is epoch 150. Actually, this was the epoch I found best. Then 160 as you are seeing right now. So the quality begins degrading I think. You see the face is better here than here
I think. Also better here than here. So you need to find your own best epoch. This totally depends on your training images
data set, therefore I can't say it will best with you. However, up to 200 epochs is usually sufficient
to find a sweet spot. Actually, after 110 I think all epochs can
be used, but you see now we are starting to get some noise at the epoch 200 so it started
to overtrain at the 200 epoch. That is for sure. That hair is also gone, the face is degraded,
so that is why I prefer to go to up to 200 epochs and then find the best one with this
way. I think this is the best epoch 150. The face is there. It is very high quality. After you decide your epoch, select it from
here. Turn off the x/y/z plot, then hit generate
forever. You can also add different prompts here. Try different prompts but why generate forever? I will explain you the logic of that. So in this folder I have over 2000 generated
images but they have a different numbering and there are some values you are seeing there. These images are sorted based on the similarity
to my real images. This makes my job much easier to find higher
quality images. Much easier. So by sorting images based on the similarity
and by generating hundreds or thousands of images, you can get very high quality images
very easily that you can use in your Linkedin, Twitter, or wherever you want. Very like you with very high quality. Like they are taken in a studio shot like
they are a professional photograph by wearing high quality suit. So how did I sort them. In this tutorial video I explained how to
sort all of the images. Everything is explained here. I already have a Patreon post that includes
script but you don't have to be my Patreon supporter to obtain the scripts. Everything is explained in this tutorial very
clearly. So can these images be further improved? Yes. You have inpainting option. There are two ways to improve generated images. However, when you use inpainting I think it
is not reaching the level of the raw output. Therefore, hopefully I am planning a fine-tuned
model with better resolution than the realistic vision and do higher resolution training on
that new fine-tuned model. Hopefully I will make a tutorial video for
that and release the model as well. So with that model, hopefully we will be able
to do training with even higher resolution and obtain better quality. But for now with realistic vision this is
the resolution I think we can go up to. Click send to inpaint. There are two options of inpainting. The first one is using the native inpainting
of the Automatic1111 web UI. Select the face select only masked and very
important thing here is only masked padding pixels. This will determine your output. This will affect your output significantly. So let's try with default values first. Okay, this is the default values first try
as you are seeing. Let's open it in a new tab. Then let's make the padding pixels 64. As you increase padding pixels, you will likely
to get better image. However, there is a sweet spot so this is
64 pixels. You see from this to this. This is looking better. Then let's try 96. So you should try different padding pixels
until you get the best result. Okay here it is. From this to this to this this is 96. This is 64. 64 is so far best. Let's say 128. By the way these are not cherry-picked this
is the first result that we are getting. So with doing multiple times generation you
can get even better ones. This. So this is 128. This is even looking better. Okay. But you see you don't get that natural look
when you do inpainting. I think most natural look comes from raw output
but this is still decent. So this is one way to improve. What is the other way of improving. By the way, this is a real improvement. The second way of improvement is using ControlNet. I have shown how to do this in another tutorial
but I will also show here. First click enable, click pixel perfect, select
tile from here. This doesn't produce always better results
but you can also try this and see if it will produce better for you. Select tile color fix, select my prompt is
more important. This is the first ControlNet then select ControlNet
unit one. Enable pixel perfect. Select inpaint from here and don't change
anything. This is important. Select whole picture and select inpaint not
masked okay and hit generate. This will inpaint the entire image and it
will likely to improve your face and also clothing. However, the best output is always from raw
images. This doesn't work. Let's also increase the denoising strength
to 1 and try again. Okay, this is a little bit better, but still
not good as the original inpainting. Sometimes it produces better results, sometimes
it is not producing better results. So I have an example here. The left one is the original, the right one
is the raw output and here we are seeing the ControlNet applied option. I think it is subjective which one is looking
better. So it is up to you. I have done a lot of testing and sometimes
it is producing better, sometimes not so it is up to you to test. I think with this way it is adding more details
like you are seeing here from here. But still both of them are good. So it is up to you to also apply this ControlNet
trick. So what is important here is as much as possibly
getting a good base image. Good raw output then inpainting face to fix. This is also very high quality inpainting
result. Let me show you the settings one more time. Inpaint masked, masked content original, only
masked, only masked padding pixels 96, resize is same, didn't change the resolution. Denoising strength is 75 percent. This is all the settings. So this is the inpainted result. Let's compare them. This was the base image. So from this base image to this inpainted
result which is a very significant improvement. It is also looking pretty natural. Let's also compare it with the real image. So here on the left we have the real image. On the right we have the inpainted face. However, we are able to get this quality because
we are doing training on a realistic vision model the most realistic model available right
now. We are using real class images and we are
using very high resolution. All these are affecting our results significantly. So I will put the summary and everything explained
in this video to this Patreon post. You will be able to find everything on there. Alternatively, you can watch the entire video
again. I am still going to work on even a better
workflow for more realism. I am going to work on styling workflow as
well because with realism and styling you can't get both of them at the same time. You can get either good styling or you can
get either good realism. Moreover, I will work on a new, better, fine-tuned
base model to train our images onto them to obtain even better quality images. Better output. If you get any question, just reply to this
video. This is all for today. I hope you have enjoyed. Please subscribe, support me by joining on
Youtube. I would appreciate that very much. Please consider supporting me on my Patreon
page. This is super important for me. We have over 250 supporters. I appreciate them very much. Supporting me on Patreon will also help you
tremendously. It will make your life easier. You will be able to access the scripts. I explain everything in my video so it is
not mandatory to support me on Patreon, but if you support me on Patreon I would appreciate
that very much. So as I said, this readme file will get updated
as it be necessary. Everything you need is written on this readme
file. Please read this readme file very carefully. Follow the steps, join the Discord channel
and ask any questions that you might get. Hopefully see you in another amazing video. By the way, I have written the used commit
hashes in the very bottom of the readme file. Moreover, I am sharing pip freeze command. With this command you are seeing all of the
libraries and their versions. If something gets broken, you can install
specific version of specific library and fix the error. However, this is probably won't be necessary. Just reply to the Youtube or the Patreon post
and hopefully I will update this readme file or the Patreon post.
