Skip to content

How Use Stable Diffusion SDXL ControlNet LoRAs For FREE Without A GPU On Kaggle Like Google Colab

Jump to bottom
FurkanGozukara edited this page Oct 23, 2025 · 1 revision

How Use Stable Diffusion, SDXL, ControlNet, LoRAs For FREE Without A GPU On Kaggle Like Google Colab

Full tutorial link > https://www.youtube.com/watch?v=dpM02YMj8FY

How Use Stable Diffusion, SDXL, ControlNet, LoRAs For FREE Without A GPU On Kaggle Like Google Colab

image Hits Patreon BuyMeACoffee Furkan Gözükara Medium Codio Furkan Gözükara Medium

YouTube Channel Furkan Gözükara LinkedIn Udemy Twitter Follow Furkan Gözükara

You want to use Stable Diffusion, use image generative #AI models for free, but you can't pay online services or you don't have a strong computer. Then this is the tutorial you were looking for. By watching this tutorial, you will learn how to use Kaggle free cloud service with famous Stable Diffusion #Automatic1111 SD Web UI as easy as it is running on your local computer. I have prepared an amazing Kaggle notebook that even supports SDXL and ControlNet of SDXL and LoRAs and custom models of #SDXL. Of course it supports all of the Stable Diffusion SD 1.5 based models too.

Kaggle Automatic1111 Notebook File ⤵️

https://www.patreon.com/posts/run-on-free-like-88714330

Tutorial GitHub Readme File ⤵️

https://github.com/FurkanGozukara/Stable-Diffusion/blob/main/Tutorials/How-To-Use-Automatic1111-Web-UI-On-A-Free-Kaggle-Notebook-Like-Google-Colab.md

00:00:00 Introduction to how to use Stable Diffusion for free without a computer or a GPU

00:02:44 How to register a Free Kaggle Account and activate it

00:03:28 How to create a new Notebook on Kaggle

00:03:42 The two very important steps that you need to make before start using Kaggle Notebook

00:03:52 How to set accelerator (select GPUs) and enable Internet

00:04:28 What is persistence and should you use it

00:04:45 How to write code on a Kaggle notebook and use the code shared in the GitHub readme file

00:04:55 How to manually setup your Stable Diffusion Automatic1111 Web UI notebook if you are not my Patreon supporter

00:05:08 How to download and import the Automatic1111 SD Web UI notebook used in this tutorial

00:06:20 Suggested resources before following this tutorial

00:07:28 How to start your Free Kaggle Dual GPU session

00:07:39 How to see how many resources you are using in your current session in a Kaggle notebook

00:08:04 How to install Automatic1111 on a Kaggle notebook

00:08:43 Explanation of the right section of the Kaggle notebook, working directory

00:09:00 How to clear the outputs of a Kaggle notebook session

00:09:10 How to download Stable Diffusion SD 1.5, LoRAs and SDXL models into the correct Kaggle directory

00:09:39 How to download models manually if you are not my Patreon supporter

00:10:14 An example of how to download a LoRA model from CivitAI

00:11:11 An example of how to download a full model checkpoint from CivitAI

00:11:48 How to start downloading all the model files

00:13:41 How to install ControlNet extension and download ControlNet models

00:15:50 How to use custom LoRAs from CivitAI or Hugging Face

00:16:29 How to start Automatic1111 Web UI instance with correct parameters and settings

00:18:42 How to understand installation of Automatic1111 Web UI has been completed and ready to use

00:19:09 First time model loading may be very slow

00:19:26 How to enable quick VAE selection drop down list

00:19:50 How to set correct folder scan path for ControlNet

00:20:29 How to reload UI for extensions and setting changes to be effective

00:20:51 Automatic1111 Web UI is ready to use on a Free Kaggle notebook

00:21:22 How to use embedded VAE of the models

00:21:48 Which image generation sampler is the best one

00:21:58 Why and how much first image generation is slower

00:22:57 How to install extensions (e.g. After Detailer) of Automatic1111 on a Free Kaggle notebook

00:23:35 You need to reload Web UI to see new extensions

00:23:55 How to use your downloaded LoRA models in Automatic1111 Web UI

00:25:17 First image generation with SDXL model

00:26:16 First time image generation speed vs consequent images generation on SDXL

00:26:54 First image generation with the custom LoRA model from CivitAI

00:27:20 Image generation speed of SDXL when using LoRA

00:27:41 How to use your own trained LoRAs, models or LoRAs, models from your computer

00:29:03 How to import your files, datasets in to your current session on Kaggle

00:29:21 How to use files imported as data set such as LoRAs or models that you have uploaded as dataset

00:30:23 How to restart Automatic1111 Web UI on a Kaggle notebook

00:31:06 How to prompt your own trained LoRA

00:31:32 How to enable After Detailer (adetailer) extension to improve faces of Stable Diffusion generated images

00:33:01 Where are the SD generated images are saved on a Kaggle notebook

00:33:45 SDXL with LoRA image generation speed

00:34:20 How to use Stable Diffusion XL (SDXL) ControlNet models in Automatic1111 Web UI on a free Kaggle

00:35:05 Where to download SDXL ControlNet models if you are not my Patreon supporter

00:36:13 Notebook crashes due to insufficient RAM when first time using SDXL ControlNet and how I fix it

00:37:39 First image generation results of SDXL ControlNet

00:39:21 SDXL ControlNet development discussion topic

00:40:22 Possible bug with ControlNet

00:41:10 How to use Stable Diffusion 1.5 version based ControlNet instead of SDXL ControlNet

00:43:29 What happens if you exceed your assigned disk space on a free Kaggle notebook

00:44:35 First time using SD 1.5 ControlNet model on a SD 1.5 based Realistic Vision 5.1 model

00:44:55 How to fix runtime error when changing SDXL to SD 1.5 based model

00:46:42 How to use PNG info to get the prompt

00:47:57 How to download all generated images

Video Transcription

  • 00:00:00 Have you ever wanted to harness the full power of image-generation AI, Stable Diffusion,

  • 00:00:05 but felt held back by the limitations of your computer or budget?

  • 00:00:10 Today I have just the solution for you.

  • 00:00:13 In this comprehensive tutorial, I will guide you through a Kaggle notebook detailing how

  • 00:00:19 to utilize custom models from CivitAI and Hugging Face such as the renowned Realistic

  • 00:00:25 Vision.

  • 00:00:26 Explore the capabilities of Stable Diffusion X-Large commonly referred to as SDXL.

  • 00:00:33 Implement custom LoRAs from CivitAI like the Pixel Art XL.

  • 00:00:38 Incorporate your very own trained LoRAs compatible even with SDXL.

  • 00:00:43 Dive into the advanced extensions of Stable Diffusion ControlNet, tailored for both Stable

  • 00:00:50 Diffusion 1.5 based models and the SDXL models.

  • 00:00:56 The SDXL support for Automatic1111 Web UI just arrived today.

  • 00:01:02 I will show you how to use ControlNet models for SDXL in this tutorial.

  • 00:01:07 Hopefully, a much more comprehensive tutorial for ControlNet for SDXL is coming.

  • 00:01:14 Moreover, install any extension you want such as After Detailer.

  • 00:01:19 And finally, download all of the generated images with just one click.

  • 00:01:25 The best part, we will achieve all of these by using Kaggle, a leading machine learning

  • 00:01:32 platform.

  • 00:01:33 Not only does Kaggle offer free GPU capabilities akin to the Google Colab, but it stands out

  • 00:01:40 in its generosity.

  • 00:01:41 With Kaggle, you gain access to powerful T4 GPU for 30 hours weekly, and you are always

  • 00:01:48 in the know about your usage and refresh times.

  • 00:01:52 You can follow this tutorial even from your mobile phone.

  • 00:01:55 You don't even need a computer.

  • 00:01:58 So if you are ready to elevate your AI capabilities without spending any money, without owning

  • 00:02:03 a strong GPU, without owning a strong computer, this tutorial is tailor-made for you.

  • 00:02:09 And we will use the best and most easy to use UI for Stable Diffusion models, Automatic1111

  • 00:02:16 Web UI.

  • 00:02:17 It will be exactly as using in your computer like you have a very strong computer.

  • 00:02:22 I have prepared a very detailed GitHub readme file for this tutorial.

  • 00:02:27 Everything that you are going to need will be in this file, the links, the codes, and

  • 00:02:32 everything else.

  • 00:02:34 I will update this file if it be necessary in future.

  • 00:02:38 Also, the link of this file will be in the description of the video and also in the pinned

  • 00:02:43 comments of the video.

  • 00:02:44 So we will begin with registering and logging in to our Kaggle account.

  • 00:02:51 If you don't have a Kaggle account, open this link.

  • 00:02:54 Register with either Google or with email.

  • 00:02:57 I prefer to register with Google.

  • 00:02:59 I have already registered my account.

  • 00:03:02 After you registered your account and logged in, go to settings.

  • 00:03:05 You can click this link to go to settings.

  • 00:03:08 And in here, you have to verify your phone number.

  • 00:03:11 As you are seeing right now, my phone is verified.

  • 00:03:14 This is necessary to be able to use the free GPUs that the Kaggle provides us.

  • 00:03:20 If your phone number verification fails for any reason, you can contact the support of

  • 00:03:26 the Kaggle from this link.

  • 00:03:28 After registered and logged into your account, we will begin with creating a new notebook.

  • 00:03:34 Click this create button here and click new notebook.

  • 00:03:38 It will start a new fresh notebook like you are seeing right now.

  • 00:03:42 There are two important things before you start.

  • 00:03:45 First of all, you need to select accelerator from here.

  • 00:03:49 Select GPU T4 x 2.

  • 00:03:52 This will provide us dual T4 GPU.

  • 00:03:56 You see, you can use it 30 hours every week.

  • 00:03:59 This is amazing.

  • 00:04:00 This is for free and each week this resets.

  • 00:04:04 So each week you can use 30 hours this dual GPU.

  • 00:04:08 These GPUs will be only assigned to you while you are using.

  • 00:04:12 This is not like Hugging Face spaces if you have ever used them.

  • 00:04:16 This will be totally reserved to you so you will have full control of these dual GPUs.

  • 00:04:22 And one another thing is Internet on.

  • 00:04:24 You need to set Internet on otherwise you will get error.

  • 00:04:28 You can also select persistence.

  • 00:04:30 However, this is not working very well in many cases.

  • 00:04:34 Also, it will slow you down when you start your session, restart your session to reload

  • 00:04:40 the data.

  • 00:04:41 But you can also enable persistence if you wish.

  • 00:04:44 I don't prefer it.

  • 00:04:46 So what you need to do is you need to delete this code and use the codes I shared here.

  • 00:04:52 You see, for example, copy this one, paste it and run the notebook.

  • 00:04:56 Then copy this one, paste it and run the code and continue.

  • 00:05:01 You can also add new code, copy paste the new code to there like this, and execute them.

  • 00:05:08 However, if you are my Patreon supporter, which I hope you will be, you can download

  • 00:05:13 the entire notebook from this link.

  • 00:05:16 Let me download it.

  • 00:05:17 When you open this link, you will get to this Patreon page and you can download the notebook

  • 00:05:23 from this attachment or you can also click here to download.

  • 00:05:26 The notebook is downloaded.

  • 00:05:28 Then we are going to click file and import notebook.

  • 00:05:33 Then browse files, go to download, select the notebook, click import and the entire

  • 00:05:40 notebook is imported.

  • 00:05:42 I am going to show entire code in this video.

  • 00:05:46 So if you are not my Patreon supporter, you can still follow this video perfectly fine

  • 00:05:52 and you will be able to do everything that we are doing with this notebook because the

  • 00:05:57 code that is necessary and the instructions are all shared in this GitHub readme file.

  • 00:06:04 So read this readme file very carefully.

  • 00:06:06 Moreover, I will update this readme file if it be necessary in future.

  • 00:06:11 So if it gets broken, just reply to this video or join our Discord channel and ask me questions

  • 00:06:18 or tell me what is broken.

  • 00:06:20 Moreover, before starting, I suggest you to check out the suggested resources.

  • 00:06:25 For example, this link will open the ControlNet extension development for Stable Diffusion

  • 00:06:33 XL with Automatic1111 Web UI.

  • 00:06:36 Also, I suggest you to watch this tutorial video where I have shown how to train your

  • 00:06:43 own Stable Diffusion XL LoRA on a free Kaggle account.

  • 00:06:47 You can also do Stable Diffusion 1.5 based versions training as well.

  • 00:06:53 So it is not limited to SDXL.

  • 00:06:55 Whatever I show is not limited to Stable Diffusion XL.

  • 00:07:00 SDXL.

  • 00:07:01 You can do everything with Stable Diffusion 1.5 based versions as well.

  • 00:07:06 And I also suggest you to watch this ControlNet tutorial.

  • 00:07:10 This is a huge tutorial where I have shown how to install Automatic1111 Web UI, how to

  • 00:07:15 install ControlNet, how to download its models, which ControlNet model and option do what.

  • 00:07:21 So this is a huge tutorial that you should watch.

  • 00:07:25 After that, you are ready to start following this tutorial.

  • 00:07:28 First of all, we need to start our session.

  • 00:07:31 So we click this icon.

  • 00:07:33 It will start our session.

  • 00:07:35 Once the session has been started, you will get a green circle here.

  • 00:07:39 When you click this tab, it will open the session details.

  • 00:07:44 You see, I am using 4GB of my maximum temporary disk, which is 73GB, which is a huge disk

  • 00:07:52 space.

  • 00:07:53 We are using 500MB RAM memory at the moment.

  • 00:07:56 We are not using any of the GPUs that we have.

  • 00:08:00 You see, I have 2 GPUs.

  • 00:08:02 You should also have dual GPU here.

  • 00:08:04 Then we will begin with installing everything written here.

  • 00:08:08 This is going to install this library to manage the RAM better in a Kaggle notebook.

  • 00:08:15 This is going to clone the Stable Diffusion Web UI into our working directory.

  • 00:08:20 These lines are going to generate these temporary folders because we are going to download ControlNet

  • 00:08:27 models and the Stable Diffusion models into the temporary disk.

  • 00:08:30 Because the temporary disk is 70GB.

  • 00:08:33 However, our working directory is only 20GB.

  • 00:08:37 Since we have cloned our repo into our working directory, it is now shown here.

  • 00:08:43 So this right section of the notebook is really important.

  • 00:08:46 This right section can be hidden and opened back from here as you are seeing right now.

  • 00:08:52 So the Kaggle working directory is our main directory.

  • 00:08:56 I will show everything, so don't worry.

  • 00:08:58 You see, everything were properly installed.

  • 00:09:01 Let's clear the output.

  • 00:09:02 So for clearing the output, right click anywhere and clear all outputs and it will clear all

  • 00:09:07 the output.

  • 00:09:08 Now the second step.

  • 00:09:10 In here, I am going to download the SDXL base model, SDXL refiner model, SDXL best VAE model

  • 00:09:19 into the respected folders.

  • 00:09:21 I am going to download the main models into the Kaggle temporary models folder because

  • 00:09:26 as I said, we have much more space in the temporary folder.

  • 00:09:29 However, I am going to download the best VAE into the working directory.

  • 00:09:35 Don't worry, these are all explained here.

  • 00:09:38 The logic is so simple.

  • 00:09:40 First, you need to find the repository.

  • 00:09:42 For example Stable Diffusion XL base repository on Hugging Face.

  • 00:09:47 Then go to the files and versions tab and in here you need to copy the link of the file.

  • 00:09:53 To copy the download link, you see there is a download arrow here.

  • 00:09:57 Right click and copy link address.

  • 00:09:59 Then all you need to do is come here, change this link with that one.

  • 00:10:05 For example, I deleted it and copy pasted it.

  • 00:10:08 Then change the model file name that you want it to be saved as.

  • 00:10:13 You can also change this.

  • 00:10:14 So with this strategy, you can use any custom model on Hugging Face or even on CivitAI.

  • 00:10:20 The difference with CivitAI is that it is like this.

  • 00:10:25 Let me show an example from CivitAI.

  • 00:10:28 For example, this model.

  • 00:10:31 This is a pixel art SDXL LoRA.

  • 00:10:34 Let me open its page from CivitAI.

  • 00:10:37 So this is the CivitAI page of this LoRA file.

  • 00:10:41 Either it can be LoRA or it can be model.

  • 00:10:43 It doesn't matter.

  • 00:10:44 The logic is same so I am teaching you the logic.

  • 00:10:47 So to download this, right click here and copy the link address from here.

  • 00:10:51 Then go back to your notebook, delete this link, paste it.

  • 00:10:56 You see the link is here.

  • 00:10:58 And in here, just define its path and the download file name.

  • 00:11:03 The extension of the file name is really important.

  • 00:11:05 Since this is a SafeTensors file, I need to make it a SafeTensors extension.

  • 00:11:12 What if it was a checkpoint like you are seeing Juggernaut XL?

  • 00:11:17 Still same.

  • 00:11:18 I just right click, copy the link address.

  • 00:11:20 And in the Kaggle, then this time I need to download it into the models folder here.

  • 00:11:26 So I just need to change the download link from here, paste it and download it into the

  • 00:11:32 models folder with its name.

  • 00:11:35 If you also reply to the Patreon post and if you need any additional models that you

  • 00:11:39 need here, I will change and modify the notebook file, update the Patreon post.

  • 00:11:46 So I will make it easier for you.

  • 00:11:48 So when I click this link, it will start downloading the models one by one.

  • 00:11:52 You see the download speed is huge on Kaggle.

  • 00:11:56 You are seeing 300 megabytes, 200 megabytes at the moment.

  • 00:11:59 200 megabytes per second means that actually it is around 2000 megabits per second.

  • 00:12:06 So this is a huge speed.

  • 00:12:07 You see the entire SDXL model is being downloaded in just under one minute as you are seeing

  • 00:12:14 right now.

  • 00:12:15 So Kaggle is super fast in downloads as well.

  • 00:12:19 Kaggle is much better than Google Colab if you ask my opinion.

  • 00:12:24 So the entire download is completed in just 32 seconds for the SDXL main model as you

  • 00:12:31 are seeing right now.

  • 00:12:32 32 seconds.

  • 00:12:33 So it will download all of the models written here.

  • 00:12:36 If you don't need any model, just put a hashtag like this in front of it and it will get commented

  • 00:12:44 out and that line will not be executed.

  • 00:12:47 So wait until you see this circle has been stopped.

  • 00:12:51 That means the cell execution has been completed.

  • 00:12:55 You can only execute one cell at a time.

  • 00:12:58 This is a rule commonly shared in Google Colab or Kaggle or other similar platforms that

  • 00:13:05 you may have used or you may use.

  • 00:13:08 So it is downloading all of the models.

  • 00:13:10 And you see while that cell is being executed, I have cancel run option here.

  • 00:13:15 So when you see this option is gone, that means that the cell execution has been completed.

  • 00:13:22 You can also watch the parameters here.

  • 00:13:24 You see our disk usage is getting increased.

  • 00:13:26 Okay, all files have been downloaded.

  • 00:13:29 You see 7 gigabytes file downloaded in just 51 seconds.

  • 00:13:33 This is super fast.

  • 00:13:35 And that cancel run option is not here anymore.

  • 00:13:39 So I right click and clear all outputs.

  • 00:13:41 Now the next thing is whether you want to use ControlNet or not.

  • 00:13:45 The ControlNet is the best extension and the most advanced extension of the Stable Diffusion

  • 00:13:51 and Automatic1111 Web UI.

  • 00:13:54 So if you don't know what is ControlNet, as I said, watch this ControlNet tutorial.

  • 00:14:00 This is a huge tutorial, about 100 minutes fully chaptered.

  • 00:14:05 So by watching this tutorial, you will learn a lot of things, almost everything about ControlNet.

  • 00:14:12 Let's begin with downloading the ControlNet for Stable Diffusion X-Large, SDXL.

  • 00:14:18 This just arrived like yesterday.

  • 00:14:20 This wasn't supported by Automatic 1111 Web UI.

  • 00:14:24 However, now Automatic 1111 Web UI also supporting ControlNet models for SDXL.

  • 00:14:30 Hopefully, I will make a new whole tutorial about ControlNet in SDXL.

  • 00:14:36 But it is not ready yet.

  • 00:14:37 So in this tutorial, I will just show an example of how to use it.

  • 00:14:42 So wait my next tutorial about ControlNet in Stable Diffusion XL when using Automatic1111

  • 00:14:47 Web UI.

  • 00:14:48 Previously, you had to use Comfy UI to use SDXL and ControlNet.

  • 00:14:55 However, now we can both use SDXL Refiner and ControlNet with Automatic 1111 Web UI.

  • 00:15:02 All thanks to the guy that develops this extension, which is this guy you see.

  • 00:15:10 The name is Lyumin Zhang.

  • 00:15:12 I am not sure how am I pronouncing it.

  • 00:15:15 Lyumin Zhang.

  • 00:15:16 Yes, he written his pronouns.

  • 00:15:18 So this guy is an amazing developer.

  • 00:15:20 You can also read this thread to learn a lot of information regarding Stable Diffusion

  • 00:15:26 XL ControlNet models.

  • 00:15:28 Okay, all ControlNet models are downloaded while I was talking in just few minutes because

  • 00:15:35 of download speed of Kaggle.

  • 00:15:37 Then I am going to skip this cell for now because this is going to download ControlNet

  • 00:15:42 models for SD 1.5 based models.

  • 00:15:44 You see, we are already using 50 GB of our disk space.

  • 00:15:50 Now, whether you want to use LoRA or not, this is the cell you need.

  • 00:15:55 We are also going to download the LoRA's into the temporary folder.

  • 00:16:00 So I will just download the LoRA of this pixel art as an example.

  • 00:16:05 But you can download and use any LoRA's.

  • 00:16:07 I will also show example of your trained LoRA's how to use them.

  • 00:16:12 But let's just begin with a custom LoRA hosted on either CivitAI or HuggingFace.

  • 00:16:19 Currently I am going to use from CivitAI.

  • 00:16:22 So the LoRA file is also downloaded.

  • 00:16:23 It is already very small file only 160 MB.

  • 00:16:28 And now we are ready.

  • 00:16:30 These are our Automatic1111 Web UI command line arguments.

  • 00:16:34 You see xFormers, share and I am giving a custom folder name for the model files.

  • 00:16:41 Why?

  • 00:16:42 Because I downloaded models not inside models Stable Diffusion folder, but inside Kaggle

  • 00:16:47 temporary models folder.

  • 00:16:49 This is really really important because we are utilizing the temporary folder that Kaggle

  • 00:16:53 gives us.

  • 00:16:55 If you use my Patreon Kaggle notebook you don't need to make any changes.

  • 00:16:59 This is also shared in the GitHub readme file in the bottom as you are seeing right now.

  • 00:17:04 So don't worry about that either.

  • 00:17:06 And we are going to start with no-half-vae command because this command is for SDXL.

  • 00:17:13 I think they also added a new command.

  • 00:17:16 Let's also find it with you.

  • 00:17:18 So this is the space where all of the command line arguments are shared.

  • 00:17:23 Okay it is not written here yet.

  • 00:17:26 So let's search the repository.

  • 00:17:28 Okay I couldn't find it.

  • 00:17:30 Currently they added --medvram-sdxl option which enables the medium VRAM only for SDXL.

  • 00:17:38 I believe I have seen they added the similar option for VAE as well.

  • 00:17:44 So let me look one more time.

  • 00:17:46 Okay I couldn't find it, but we are going to start with no half VAE.

  • 00:17:51 This is going to make the SD 1.5 based models image generation slower.

  • 00:17:57 So if you are not going to use SDXL, you should also comment this VAE precision like this.

  • 00:18:05 But if you are going to use SDXL make it like this.

  • 00:18:09 Okay let's begin.

  • 00:18:10 So I will click this cell and start Web UI.

  • 00:18:13 The first time when you start this cell.

  • 00:18:16 When you execute this cell, it will install the necessary requirements of the Stable Diffusion

  • 00:18:22 Automatic1111 Web UI.

  • 00:18:23 You see it is downloading Torch version 2.0.1.

  • 00:18:27 So this is only first time installation.

  • 00:18:32 After that if you just stop this cell and restart it, it won't install the libraries

  • 00:18:37 again until you restart your session or terminate your session and start again.

  • 00:18:42 The installation has been completed.

  • 00:18:44 How will you know?

  • 00:18:45 You will get this Gradio live link.

  • 00:18:48 This is really important.

  • 00:18:49 Let me zoom in.

  • 00:18:51 This link.

  • 00:18:52 Open this link.

  • 00:18:53 After opening this link, you can right click and clear all outputs like this.

  • 00:18:58 And this cell is still being executed because this is the cell which is running our Automatic1111

  • 00:19:05 Web UI instance.

  • 00:19:07 And our instance is getting loaded.

  • 00:19:09 When you first time load your Web UI the load of the model may take a lot of time because

  • 00:19:16 we have a very limited RAM as you are seeing here.

  • 00:19:19 So it has loaded with Realistic Vision version 5.1 because we have also downloaded that model.

  • 00:19:26 First let's go to the settings.

  • 00:19:28 Go to the interface in the settings section.

  • 00:19:32 You will find it here.

  • 00:19:33 User interface.

  • 00:19:35 Go to the info quick settings list.

  • 00:19:38 Type VAE and select SD VAE.

  • 00:19:40 This is really important because with this we will be able to select which VAE to use.

  • 00:19:47 Apply settings.

  • 00:19:48 Then reload UI.

  • 00:19:50 But before doing that since we are also going to use ControlNet.

  • 00:19:54 Let's also select its folder.

  • 00:19:57 So go to the ControlNet tab in here.

  • 00:20:00 And you see it has extra path to scan for ControlNet models.

  • 00:20:04 This is really important.

  • 00:20:05 This is the place where we will set the model folder.

  • 00:20:09 So the model folder is going to be.

  • 00:20:12 Let's find it in here.

  • 00:20:14 Kaggle temp CN models.

  • 00:20:16 This is the folder we have given.

  • 00:20:18 You can also define different folder pathing.

  • 00:20:21 It is totally up to you.

  • 00:20:23 The Kaggle temp part is the important part.

  • 00:20:26 Apply settings.

  • 00:20:27 Now ControlNet will scan this folder.

  • 00:20:30 Then reload UI.

  • 00:20:31 When you reload UI you will get a new Gradio public link.

  • 00:20:35 This won't work anymore.

  • 00:20:37 So close this.

  • 00:20:39 Wait link to appear here.

  • 00:20:41 You see now we got a new public URL for the Gradio live.

  • 00:20:46 Open it.

  • 00:20:47 And Automatic1111 Web UI has been started successfully.

  • 00:20:51 This is now going to work as in your computer as it is running locally.

  • 00:20:57 But this is going to run on the remote server.

  • 00:21:00 So nothing will be actually executed in your computer.

  • 00:21:04 Everything will be executed on the cloud.

  • 00:21:06 However you will be able to use it as easy as in your computer.

  • 00:21:11 So let's begin with generating some images with Realistic Vision.

  • 00:21:14 Therefore I am going to select the best SD 1.5 based models VAE from here.

  • 00:21:22 If you want to use the embedded VAE of the model you need to select none option from

  • 00:21:29 this selection.

  • 00:21:30 You see none.

  • 00:21:31 When you select none it will use the embedded VAE.

  • 00:21:34 If you don't know what is VAE just search Google.

  • 00:21:36 You will get a lot of information.

  • 00:21:38 So let's generate an example image.

  • 00:21:40 Photo of an amazing sports car.

  • 00:21:44 Let's make the resolution 768.

  • 00:21:46 768.

  • 00:21:48 I find that the best sampler is DPM++ 2M SDE Karras.

  • 00:21:55 So I am going to select it.

  • 00:21:57 And let's hit generate.

  • 00:21:58 The first generation may be slower than expected.

  • 00:22:01 Let's see the speed.

  • 00:22:02 So you see for the speed I am checking here.

  • 00:22:05 It is reaching about 2.4 IT per second which is a very very good speed because we are generating

  • 00:22:11 with a higher resolution.

  • 00:22:13 When the first time it is generating image, the VAE decomposition decoding part is going

  • 00:22:20 to take more time as you are seeing right now.

  • 00:22:23 So it has generated all the steps and now generating the final image.

  • 00:22:27 So the entire generation took 24 seconds.

  • 00:22:31 Let's look at the image.

  • 00:22:32 So this is the image we got.

  • 00:22:34 We didn't use any negative prompts.

  • 00:22:36 Let's generate another one.

  • 00:22:38 Okay it is generating with 2.4 IT per second.

  • 00:22:42 You see this time VAE didn't take almost any time because it was cached in the RAM memory.

  • 00:22:48 And this time the total generation took 9 seconds.

  • 00:22:52 So the first time was 24 seconds.

  • 00:22:54 The second time is 9 seconds.

  • 00:22:56 And we got another image.

  • 00:22:57 When you are running on Kaggle you can also install extensions.

  • 00:23:02 All you need to do is go to extensions tab.

  • 00:23:04 Install from let's say available.

  • 00:23:07 Click load from.

  • 00:23:08 Let's install After Detailer which is one of the favorite extensions that I have.

  • 00:23:13 You see After Detailer.

  • 00:23:14 Click install.

  • 00:23:16 After you clicked install you should watch what is happening in here.

  • 00:23:21 This is the command line interface that we have when we are running on our local computer.

  • 00:23:25 You see it is installing all of the extension requirements.

  • 00:23:28 And it is successfully installed.

  • 00:23:31 Also this screen is now unfrozen.

  • 00:23:33 When we go to text to image tab.

  • 00:23:35 Why?

  • 00:23:36 Because we need to reload the UI.

  • 00:23:37 So I clicked to reload UI.

  • 00:23:40 It is going to reload the UI.

  • 00:23:42 You see it is downloading the necessary files for that extension to become available.

  • 00:23:47 And then I have got a new link.

  • 00:23:49 Click it.

  • 00:23:50 Now I have the After Detailer extension installed.

  • 00:23:53 So it is also supporting extension installation.

  • 00:23:56 So let's say you have downloaded some CivitAI LoRAs.

  • 00:23:59 You just need to click here.

  • 00:24:01 After that you need to click refresh.

  • 00:24:03 But currently I don't have any Stable Diffusion 1.5 version based LoRA.

  • 00:24:09 However I have a LoRA for SDXL.

  • 00:24:13 So to use SDXL LoRA I will first select the SDXL base safetensors from here.

  • 00:24:20 It is going to take some time to load it because we have a very limited RAM.

  • 00:24:25 Therefore the model changes are taking some time like one minute.

  • 00:24:30 Maybe sometimes over one minute.

  • 00:24:32 Just patiently wait here to fully executed.

  • 00:24:36 Okay it has been like 110 seconds.

  • 00:24:39 It is still loading.

  • 00:24:40 Okay almost done.

  • 00:24:42 Yes it is finally loaded.

  • 00:24:44 You should also watch the messages printed here.

  • 00:24:47 So model loading took a lot of time.

  • 00:24:50 It says actually 49 seconds but it was more here.

  • 00:24:53 Then you also need to change the VAE.

  • 00:24:56 This is really important.

  • 00:24:58 You need to change it to either none or SDXL VAE.

  • 00:25:01 Now it is going to load the SDXL VAE.

  • 00:25:05 You can also watch here and see what is happening.

  • 00:25:08 You see it is using more RAM because it is going to de-load the previous VAE.

  • 00:25:12 Okay VAE is loaded.

  • 00:25:16 Okay now we are ready.

  • 00:25:17 So let's generate our first image with SDXL base model.

  • 00:25:22 For SDXL there are several resolutions.

  • 00:25:25 The very best one is 1024x1024.

  • 00:25:29 So let's generate photo of an amazing sports car.

  • 00:25:33 I am going to select DPM++ 2M SDE Karras.

  • 00:25:38 Let's generate.

  • 00:25:39 The first generation will be slower just like the previous time I presume.

  • 00:25:43 However if you don't have a strong GPU this is working very well.

  • 00:25:47 This speed is very decent.

  • 00:25:49 With RTX 3090 and a very powerful GPU on my computer I am only getting like 3.5 IT per

  • 00:25:57 second and this is like 1 IT per second.

  • 00:26:00 Which is only 3 times slower, but this is totally free.

  • 00:26:04 You don't spend like $1000 even more for a strong GPU.

  • 00:26:08 You are able to use this 30 hours every week.

  • 00:26:11 You see the VAE part is taking a lot of time because it is going to cache the VAE.

  • 00:26:16 Okay it took total 39 seconds.

  • 00:26:19 And this is the image we got.

  • 00:26:21 Let's generate another one and see how much time it is going to take this time.

  • 00:26:25 So since everything were cached, it is going to take lesser time.

  • 00:26:29 Let's wait.

  • 00:26:30 The speed is now 1.15 IT per second.

  • 00:26:34 Okay 15.

  • 00:26:35 Okay now this time it took only 19 seconds.

  • 00:26:39 So it was first time 49 seconds.

  • 00:26:42 In the second time it is only 19 seconds which is a very, very good timing.

  • 00:26:47 You can't get this timing in Google Colab.

  • 00:26:50 You can only get this like timing in a very strong computer.

  • 00:26:53 Now we can use the first LoRA which we downloaded from the CivitAI.

  • 00:26:59 So I click the LoRA tab.

  • 00:27:01 I click refresh.

  • 00:27:02 Now that LoRA should appear.

  • 00:27:04 Okay you see it has appeared.

  • 00:27:05 Let's append it and let's try to generate this image.

  • 00:27:09 So I click info.

  • 00:27:10 I copy the prompt.

  • 00:27:12 Let's copy paste it.

  • 00:27:14 Let's add the LoRA name and let's copy paste the negative prompt and paste it.

  • 00:27:19 So let's hit generate.

  • 00:27:20 Okay let's see the speed.

  • 00:27:22 Since we have added LoRA.

  • 00:27:24 It will be probably slower than the previous time.

  • 00:27:27 But I will generate two images and see the final speed when we are using LoRA.

  • 00:27:33 It is looking still very decent.

  • 00:27:35 All right it took 20 seconds so we didn't lose much speed and we got a picture.

  • 00:27:39 Okay it is working very well.

  • 00:27:41 How about if you want to use the your own trained LoRAs.

  • 00:27:47 Please watch this tutorial to learn how you can train your own LoRAs on a free Kaggle

  • 00:27:52 account.

  • 00:27:53 Currently I have my LoRAs uploaded into my data sets.

  • 00:27:58 These are private so to upload your own models, ownb LoRAs you just need to click this icon.

  • 00:28:05 Enter a title.

  • 00:28:07 Don't enter non-English characters.

  • 00:28:09 So let's say test 111 LoRA.

  • 00:28:13 Then select browse files.

  • 00:28:15 I have my last training LoRAs inside here, inside models, inside LoRA.

  • 00:28:20 Okay, I have test Kaggle LoRA.

  • 00:28:22 I will just going to upload it.

  • 00:28:24 Open it.

  • 00:28:25 It is going to get uploaded.

  • 00:28:27 This totally depends on your upload speed.

  • 00:28:29 Let's look at the upload speed from performance.

  • 00:28:33 Okay it is uploading with my full speed.

  • 00:28:35 It is 20 megabits per second.

  • 00:28:37 This is my upload speed.

  • 00:28:39 Okay the upload has been completed.

  • 00:28:40 Now I will click create.

  • 00:28:42 This will be private so no one else will have access to it.

  • 00:28:46 It is test 111 LoRA.

  • 00:28:48 Okay it was success.

  • 00:28:50 However we can't use it directly.

  • 00:28:52 You see it is getting added into my current session.

  • 00:28:56 You can also use your uploaded data sets, your files in every other session that you

  • 00:29:01 may have in future.

  • 00:29:03 All you need to do is click this add data.

  • 00:29:06 Select your data sets and click the plus icon here and then they will get added into this

  • 00:29:12 session.

  • 00:29:13 For example: let's also add class images from my previous video.

  • 00:29:16 You see it is going to get preparing to download and they will get added into my session.

  • 00:29:21 However we can't directly use this new LoRA.

  • 00:29:24 Why?

  • 00:29:25 Because our LoRA folder is not this one.

  • 00:29:29 It is using the default LoRA folder at the moment.

  • 00:29:31 So how am I going to use my new LoRA?

  • 00:29:34 First of all you need to click either cancel run or this cell execution.

  • 00:29:40 So let's click cancel run.

  • 00:29:41 Let's clear all of the outputs.

  • 00:29:44 Clear all outputs.

  • 00:29:45 Then copy the path of this LoRA and paste it here.

  • 00:29:50 Then I need to remove this # so it will get uncommented.

  • 00:29:56 And now the Automatic1111 Web UI will look this folder for LoRA files.

  • 00:30:02 However we have now some residual RAM.

  • 00:30:05 You see it is using 3 GB RAM.

  • 00:30:08 Unfortunately I couldn't find how to clear this RAM usage.

  • 00:30:12 Even though I have installed this library, it is still using some RAM when it shouldn't

  • 00:30:19 use any.

  • 00:30:20 But let's try and see if it will work.

  • 00:30:23 So we will start another instance of the Automatic1111 Web UI.

  • 00:30:29 This time it won't install the libraries again because it was already installed and we didn't

  • 00:30:35 restart our session or we didn't terminate our session.

  • 00:30:39 Okay it is getting started.

  • 00:30:41 This start should be very fast.

  • 00:30:43 Yes it is getting started in like 30 seconds.

  • 00:30:45 Maybe under one minute.

  • 00:30:47 I am not sure.

  • 00:30:48 Let's wait.

  • 00:30:49 Yeah it took like 32 seconds and now we got our new link.

  • 00:30:54 Let's open it.

  • 00:30:55 Okay it is loaded with the SDXL base model file and the VAE.

  • 00:31:01 Okay it is not using much RAM.

  • 00:31:02 That is very good.

  • 00:31:04 So let's use our new LoRA.

  • 00:31:06 So I am going to use this prompt.

  • 00:31:08 This is my LoRA activation tokens.

  • 00:31:11 If you watch the other tutorial.

  • 00:31:13 you will learn what is it.

  • 00:31:14 Let's select our sampler.

  • 00:31:17 Let's make the resolution 1024x1024.

  • 00:31:19 This is the resolution that I trained.

  • 00:31:22 Then I need to add my LoRA.

  • 00:31:23 So click LoRA and now you see that LoRA is loaded.

  • 00:31:27 Let's click it.

  • 00:31:28 Let's generate 12 images and see which one will be very good.

  • 00:31:33 Also I will enable After Detailer and I will just use this prompt when inpainting the face.

  • 00:31:41 So let's copy this, paste it and let's delete this part.

  • 00:31:45 Okay so this will inpaint the face.

  • 00:31:47 I will also make the inpainting denoise strength to 50% like this.

  • 00:31:53 Okay let's hit generate and let's watch the progress in here.

  • 00:31:58 After this I will show ControlNet.

  • 00:32:01 How to use ControlNet.

  • 00:32:02 When the first time we generate after we restarted our Automatic1111 Web UI it will take a lot

  • 00:32:09 of time.

  • 00:32:10 It is also loading the LoRA weights.

  • 00:32:11 Okay it is started.

  • 00:32:13 The first time is slow as expected, but not very slow.

  • 00:32:16 This is a 32 rank dimension having LoRA.

  • 00:32:20 This was trained for SDXL on a free Kaggle account.

  • 00:32:24 As I said watch this tutorial to learn how you can train your own LoRA's.

  • 00:32:28 They can be style, they can be person, your face, anything.

  • 00:32:33 Okay in the first generation it is taking time at the VAE.

  • 00:32:37 Then it will cache the VAE and it will become much faster.

  • 00:32:42 Okay let's see the second one.

  • 00:32:45 But first it is inpainting the face because we are using After Detailer.

  • 00:32:49 It is also working very well as we are seeing right now.

  • 00:32:52 This is the first time run.

  • 00:32:53 Then it will move to the next.

  • 00:32:56 Okay it moved to the next one.

  • 00:32:57 Now it is much faster.

  • 00:32:58 You can see the speed comparison.

  • 00:33:01 By the way if you are wondering where are the outputs are saved.

  • 00:33:05 They are saved inside the Kaggle working.

  • 00:33:08 Inside Stable Diffusion Web UI.

  • 00:33:11 Inside outputs folder.

  • 00:33:13 Inside text to images folder.

  • 00:33:15 And we will see our outputs.

  • 00:33:17 This is working exactly as it would work in your computer.

  • 00:33:20 When you click here.

  • 00:33:22 You will see all of the generated images.

  • 00:33:25 Here you can download them.

  • 00:33:27 So to download them, go to here.

  • 00:33:29 Right click and click download.

  • 00:33:30 And it will download the image.

  • 00:33:33 Let's open it so I click it.

  • 00:33:34 And you see this is the previous image we generated.

  • 00:33:38 The new images will also appear here when I refresh.

  • 00:33:42 You see the new one has arrived.

  • 00:33:43 Okay the speed is very decent.

  • 00:33:45 It is taking only 17 seconds to generate each image.

  • 00:33:49 Then 18 seconds.

  • 00:33:51 Of course, the face inpainting is also taking time.

  • 00:33:53 It is taking like 10 seconds.

  • 00:33:55 So total image generation is being like 28 seconds right now.

  • 00:34:01 This is 1024x1024.

  • 00:34:03 So this is a really high resolution when you consider Stable Diffusion SD 1.5 models.

  • 00:34:09 You can also watch the progress here.

  • 00:34:11 So all of the images have been generated.

  • 00:34:13 The LoRA is working.

  • 00:34:14 This was a very simple prompt.

  • 00:34:16 So therefore, the results are not very good.

  • 00:34:18 But the LoRA is working.

  • 00:34:20 Now we can begin testing the ControlNet of Stable Diffusion.

  • 00:34:26 How are we going to do that?

  • 00:34:28 So I will disable After Detailer and I will open ControlNet.

  • 00:34:31 This is the ControlNet interface.

  • 00:34:34 To learn more about ControlNet please watch this master tutorial.

  • 00:34:39 Hopefully I will also make a very comprehensive tutorial about Stable Diffusion XL ControlNet

  • 00:34:45 options.

  • 00:34:46 But they are pretty much similar.

  • 00:34:49 So let's also clear all outputs here.

  • 00:34:51 And let's load a beginning image.

  • 00:34:54 So I will use my this image.

  • 00:34:56 I will use this as a Canny.

  • 00:34:59 So let's select Canny.

  • 00:35:00 And you see it is automatically seeing the Diffusers XL Canny full model.

  • 00:35:05 Where you can download these models yourself if you are not my Patreon supporter?

  • 00:35:09 They are all shared in this Hugging Face repository.

  • 00:35:13 The link is in the GitHub readme file.

  • 00:35:15 That is why it is really important to read this readme file.

  • 00:35:19 Okay Canny is selected.

  • 00:35:21 Canny model is selected.

  • 00:35:23 Let's enable preview.

  • 00:35:25 So when I click this, it will show me the preview of this image as a Canny preprocessed.

  • 00:35:31 You see like this.

  • 00:35:32 Let's also enable Pixel Perfect.

  • 00:35:34 And we are ready.

  • 00:35:35 Now we can generate a new image based on this image.

  • 00:35:40 So I will use my own LoRA and this prompt.

  • 00:35:44 And let's generate one image.

  • 00:35:46 Maybe two images.

  • 00:35:47 And see the results.

  • 00:35:49 It should be as in this format.

  • 00:35:53 Okay let's look at the output of the command line.

  • 00:35:57 Equivalent of the command line.

  • 00:35:59 So it is going to load the SDXL config for ControlNet.

  • 00:36:03 Since this is first time using.

  • 00:36:06 By the way it is going to use more RAM because now it is combined with LoRA and the ControlNet

  • 00:36:13 Okay we got a connection error.

  • 00:36:16 Okay.

  • 00:36:17 Okay, so what can we do since we have a connection error?

  • 00:36:20 We can reload the browser and see if it is working.

  • 00:36:25 Okay it is not working.

  • 00:36:27 So we need to restart the Web UI instance.

  • 00:36:31 So I will click cancel run.

  • 00:36:32 This is all because of the limited RAM it provides.

  • 00:36:37 Yeah we are having a RAM problem and it is not able to cancel run.

  • 00:36:41 So sometimes you may encounter similar issues.

  • 00:36:45 Sometimes the only solution could be restarting the entire session.

  • 00:36:49 Be prepared for it.

  • 00:36:51 When you also combine more things you may also get such out of RAM errors.

  • 00:36:58 Okay it was restarted.

  • 00:36:59 Okay very nice.

  • 00:37:00 Now our RAM usage is dropped a lot.

  • 00:37:03 It is entirely cleared.

  • 00:37:05 Let's clear all of the output.

  • 00:37:07 Clear all outputs.

  • 00:37:09 Restart our Automatic1111 Web UI session.

  • 00:37:12 Remember if you click this or click this, you will start from beginning.

  • 00:37:16 You will install everything from beginning.

  • 00:37:18 Okay it is going to start the instance pretty quickly.

  • 00:37:22 Okay we got the new Gradio live link.

  • 00:37:24 Let's open it.

  • 00:37:25 It is getting loaded.

  • 00:37:26 Alright so let's go to ControlNet.

  • 00:37:29 Reload the image.

  • 00:37:31 Maybe let's use this one.

  • 00:37:33 Not that one.

  • 00:37:34 So I will select Pixel Perfect.

  • 00:37:36 Then I will select Canny.

  • 00:37:37 Okay it is selected.

  • 00:37:39 Let's say photo of a very famous and very handsome actor.

  • 00:37:46 Or let's say very handsome male actor.

  • 00:37:50 Let's change this to 1024 1024.

  • 00:37:54 Let's change the sampling method and let's generate 4 images.

  • 00:37:59 Okay everything looking good.

  • 00:38:00 Make sure that you have enabled the ControlNet.

  • 00:38:03 And selected the correct model and the preprocessor.

  • 00:38:06 Let's watch the status here.

  • 00:38:08 So the first time it will cache into RAM.

  • 00:38:11 I am spending huge time to prepare these tutorials.

  • 00:38:15 These easy to use scripts for you.

  • 00:38:18 So I appreciate that if you support me on Patreon or with other ways that you wish.

  • 00:38:23 You can also support me with cryptocurrency if you wish.

  • 00:38:26 So I am open to every support or you can support me by purchasing my Udemy course.

  • 00:38:31 Or giving me a Coffee or sponsoring me on GitHub.

  • 00:38:35 All of them are available in the GitHub repository.

  • 00:38:39 All of the links are here.

  • 00:38:40 If you also go to this Stable Diffusion link and Star our repository I would appreciate

  • 00:38:46 that very much.

  • 00:38:47 We are almost 1,000 stars.

  • 00:38:49 It is all thanks to you.

  • 00:38:50 You can also fork this, watch this and sponsor me from here.

  • 00:38:54 Okay the image is being generated.

  • 00:38:57 And we can see the progress here.

  • 00:38:59 It is not also using a lot of RAM.

  • 00:39:02 So it also freed up the previously 3 GB RAM.

  • 00:39:05 If you remember we had 3 GB RAM usage even though it should have been zero.

  • 00:39:10 Okay the images are getting generated.

  • 00:39:13 By the way ControlNet developer is doing a lot of updates all the time.

  • 00:39:18 At this moment he could have made updates.

  • 00:39:21 Let's look at the discussion thread from here.

  • 00:39:25 So it is not perfect yet.

  • 00:39:27 But he is working very hard.

  • 00:39:29 You see his last reply is 1 hour ago.

  • 00:39:33 So therefore, when you are watching this tutorial, it could be much better.

  • 00:39:38 All right, we are getting some results.

  • 00:39:41 So it should be much faster in the second image.

  • 00:39:44 The first image took 25 seconds.

  • 00:39:46 Actually the first image took 31 seconds.

  • 00:39:49 Then 25 seconds.

  • 00:39:51 Then it is 26 seconds.

  • 00:39:53 Yeah.

  • 00:39:54 And we are using the best Canny model.

  • 00:39:56 Diffusers XL Canny full.

  • 00:39:58 There are also smaller models hosted on the Hugging Face.

  • 00:40:01 You can also download and use them.

  • 00:40:03 We are generating 4 images.

  • 00:40:05 Okay we got the results.

  • 00:40:07 Let's see.

  • 00:40:08 Okay this is the first result.

  • 00:40:10 You see it is almost same picture.

  • 00:40:14 And this is second result.

  • 00:40:15 Third and the fourth.

  • 00:40:17 And this was the used Canny preprocessed map.

  • 00:40:20 I think this is because of the selected options error.

  • 00:40:25 I also made several test sets.

  • 00:40:27 So let's test with my prompt is more important and compare the results.

  • 00:40:32 I will use the same seed.

  • 00:40:34 So I click this icon.

  • 00:40:35 It will use the same seed.

  • 00:40:37 You see.

  • 00:40:38 Let's see the results.

  • 00:40:39 Okay this time the results are pretty significantly different.

  • 00:40:43 So let's look at them.

  • 00:40:44 First, second, third and fourth.

  • 00:40:48 How well it can keep the shape of the input is amazing.

  • 00:40:53 You see.

  • 00:40:54 Almost perfect hands.

  • 00:40:56 Almost perfect hand composition.

  • 00:40:58 The face direction.

  • 00:40:59 It is superb.

  • 00:41:01 So after I have investigated all of the ControlNet options hopefully I will make an amazing tutorial

  • 00:41:08 for ControlNet for Stable Diffusion XL.

  • 00:41:10 Let's say you want to use ControlNet for SD 1.5 based versions.

  • 00:41:15 So what you can do?

  • 00:41:17 Let's clear all outputs.

  • 00:41:19 Let's stop this cell execution.

  • 00:41:22 Instead of starting from zero I will delete the temporary folder models.

  • 00:41:29 However you can just click this and restart from beginning.

  • 00:41:33 So I will type here a new command.

  • 00:41:36 Which will be equal to rm -r and the path.

  • 00:41:41 Actually I will add this into this command.

  • 00:41:44 So you will be able to use this command when you are downloading from my Patreon.

  • 00:41:51 Therefore you won't be needed to restart your session.

  • 00:41:55 It will delete the older files and redownload them.

  • 00:41:59 And this should get better.

  • 00:42:01 Let's see.

  • 00:42:02 Okay it is showing like this.

  • 00:42:03 But let's execute and download all of the Stable Diffusion 1.5 based models.

  • 00:42:08 And see if we will get sufficient space.

  • 00:42:12 Okay it says that it already exists.

  • 00:42:14 It is fine.

  • 00:42:16 And okay, yeah.

  • 00:42:18 The problem is I know the problem.

  • 00:42:20 So I also need to add this command to here.

  • 00:42:24 So let's delete this cell.

  • 00:42:27 Then add this here.

  • 00:42:29 And also add this here.

  • 00:42:32 Alright.

  • 00:42:33 And now let's clear all of the outputs.

  • 00:42:36 And execute this again.

  • 00:42:37 You won't need to do these changes when you are watching this tutorial.

  • 00:42:42 Just download the notebook file and you will have this command ready.

  • 00:42:47 So it did delete the older files from the disk space.

  • 00:42:52 And then it is going to download the SD 1.5 ControlNet model files.

  • 00:42:57 But it is still showing the disk space utilization high.

  • 00:43:01 This shouldn't happen actually.

  • 00:43:03 Because we deleted the older file.

  • 00:43:06 I think it didn't fully delete them.

  • 00:43:08 Yeah this is weird.

  • 00:43:10 But let's see what happens.

  • 00:43:11 If we get out of disk space error I will just start from the scratch.

  • 00:43:17 I think it did put them into the trash bin.

  • 00:43:22 Which shouldn't happen but it looks like it happened.

  • 00:43:25 Let's see what will happen when we fully use it.

  • 00:43:27 The download speed is amazing.

  • 00:43:29 So it is not a problem if we start from the beginning.

  • 00:43:32 Okay we are almost full.

  • 00:43:34 Yeah I think we will get out of error.

  • 00:43:37 Okay it says that you are approaching the limit of disk space outside Kaggle working.

  • 00:43:42 Continued use may cause your notebook to crash.

  • 00:43:45 Let's see if it will crash or it will give us some extra space.

  • 00:43:48 Okay it downloaded all of the models.

  • 00:43:51 Also our removal didn't remove the space.

  • 00:43:55 So I won't add that line in the Kaggle notebook.

  • 00:43:58 However it is allowing us to use 80 GB so there is some probably flexibility.

  • 00:44:05 Let's restart our Web UI.

  • 00:44:08 So probably we didn't need to stop our..

  • 00:44:11 Oh we had to stop our Web UI instance because otherwise we wouldn't be able to execute this

  • 00:44:17 cell.

  • 00:44:18 That means when you are watching this tutorial, you can download both of the model files at

  • 00:44:23 the same time since it allowed us to use 80 GB instead of max 73 GB which is nice.

  • 00:44:31 Okay it is getting started and it started.

  • 00:44:34 Let's open it.

  • 00:44:35 Since this time we are going to use ControlNet models based on Stable Diffusion 1.5 based

  • 00:44:41 models I am going to change the base model.

  • 00:44:44 I am going to use Realistic Vision 5.1.

  • 00:44:48 You can also use the base SD 1.5 model as well.

  • 00:44:52 It is also automatically downloaded.

  • 00:44:53 Okay let's see message.

  • 00:44:55 Okay it says that it was a runtime error.

  • 00:44:59 Okay let's try again because sometimes we may get this error.

  • 00:45:03 We may get such errors.

  • 00:45:05 Okay let's just wait.

  • 00:45:07 Okay it is loading the VAE.

  • 00:45:09 We will also change the VAE.

  • 00:45:11 Okay model loaded I think.

  • 00:45:13 Yeah we got another error.

  • 00:45:16 I think this is because of the VAE.

  • 00:45:17 Maybe let's first set the VAE none.

  • 00:45:21 Okay then change the model.

  • 00:45:24 This could be due to we are out of HDD.

  • 00:45:29 Actually it shows the HDD here as green.

  • 00:45:32 Maybe because of the it is displaying output not temporary.

  • 00:45:36 Okay it is loading.

  • 00:45:39 You see I will not cut these parts of the video because you may also encounter same

  • 00:45:44 problems and you will learn how to solve them.

  • 00:45:47 This is really important.

  • 00:45:48 This is why you shouldn't skip my tutorials.

  • 00:45:52 You should watch them entirely.

  • 00:45:54 Okay this time it is loaded.

  • 00:45:55 So first I made the VAE none, then reloaded the model and then let's select the best VAE

  • 00:46:02 from the VAE selection.

  • 00:46:04 Okay it is loaded.

  • 00:46:05 Now we are ready.

  • 00:46:07 So let's use 1.5 based ControlNet models.

  • 00:46:11 I will use the same image.

  • 00:46:14 You see enable Pixel Perfect.

  • 00:46:16 Okay then select Canny.

  • 00:46:18 Let's see which model it will select.

  • 00:46:19 Okay by default it has selected the SD 1.5 version based Canny model as you are seeing

  • 00:46:26 right now.

  • 00:46:27 Let's also make the my prompt is more important and let's use the same prompt.

  • 00:46:32 You can click this but I wonder if it will change the Stable Diffusion checkpoint.

  • 00:46:36 Let's try.

  • 00:46:37 This should load the last used prompt.

  • 00:46:40 Okay it didn't load.

  • 00:46:42 Yeah so I will download the last generated image.

  • 00:46:47 I just click this refresh icon to it to expand the folder output.

  • 00:46:53 Okay I think the last file is 23.

  • 00:46:56 Yeah they are not sorted by name, so let's download.

  • 00:47:00 Then let's go to png info.

  • 00:47:02 Load the image.

  • 00:47:03 We should see the prompt here.

  • 00:47:05 Yes this was the prompt.

  • 00:47:08 Let's type it.

  • 00:47:09 Okay and since this is Realistic Vision, which one we should use.

  • 00:47:14 Okay let's use 512 to 512.

  • 00:47:19 Let's see the message here.

  • 00:47:20 Okay it is going to load the ControlNet model.

  • 00:47:23 It is loading the model, loading everything.

  • 00:47:26 The generation will be much faster of course because it's a lower resolution.

  • 00:47:30 And we got the output and this is the Realistic Vision output with this Canny.

  • 00:47:37 You can also see the Canny preprocessor.

  • 00:47:39 So this is the Canny preprocessor of the ControlNet extension.

  • 00:47:43 I don't know if the preprocessor output is changing based on either it is SD 1.5 model

  • 00:47:50 or SDXL model.

  • 00:47:51 I think it wouldn't change but I haven't compared.

  • 00:47:54 So I don't think so anything left except one thing.

  • 00:47:57 What if if we want to download all of the generated images.

  • 00:48:01 To do that I have a pretty easy script here.

  • 00:48:05 So to be able to execute this cell, you see I am not able to execute it right now because

  • 00:48:11 it is running another cell at the moment.

  • 00:48:14 So I cancel run.

  • 00:48:16 After canceling run you may get such messages.

  • 00:48:18 It is not important.

  • 00:48:19 Clear all outputs.

  • 00:48:21 Then this cell execution will begin.

  • 00:48:23 What this cell is going to do is it is going to zip entire outputs folder into the Kaggle

  • 00:48:31 working directory with the name of generatedimages.zip file.

  • 00:48:35 So you see it is zipped.

  • 00:48:37 It is 62 megabytes because we didn't generate too many images.

  • 00:48:41 So I click here and click download and I will download all of the images.

  • 00:48:46 If you download multiple files it will ask you to allow multiple files download from

  • 00:48:52 somewhere around here.

  • 00:48:54 So pay attention to it and allow it.

  • 00:48:57 Otherwise you won't be able to download multiple files, download any files.

  • 00:49:02 So let's look at the inside of this downloaded folder.

  • 00:49:05 I will extract it.

  • 00:49:07 To extract it right click and extract.

  • 00:49:09 I am using Winrar but on Windows you can also directly extract or if you can't extract,

  • 00:49:15 you can install Winrar.

  • 00:49:17 So extract generated images.

  • 00:49:18 So they are here.

  • 00:49:20 You see text to image grids and text to image images.

  • 00:49:23 It has downloaded both grid images.

  • 00:49:25 You see the grid images we generated and let's go back and all of the images we generated

  • 00:49:32 in this tutorial are here.

  • 00:49:35 And finally, when you are done, you need to click this and it will terminate your session

  • 00:49:40 and delete all of the data if you didn't select persistence.

  • 00:49:45 Everything here will be deleted.

  • 00:49:47 Only the data sets you upload will remain.

  • 00:49:50 Okay let's terminate and we have turned off our session.

  • 00:49:54 Everything deleted and still I have 20 hours this week GPU time.

  • 00:50:00 Next week it will get reset.

  • 00:50:03 I hope you have enjoyed.

  • 00:50:04 Please support me on Patreon.

  • 00:50:05 This is super important.

  • 00:50:07 You can also support me with buy me a Coffee.

  • 00:50:10 You can also follow me on Medium.

  • 00:50:12 I am writing articles there.

  • 00:50:13 I am writing articles on CivitAI as well.

  • 00:50:16 You can support me there.

  • 00:50:17 You can follow me on DeviantArt.

  • 00:50:20 Also on our channel.

  • 00:50:21 We have amazing videos as you are seeing right now.

  • 00:50:25 You should watch all of them.

  • 00:50:26 If you want to learn more about Stable Diffusion, more about AI because I have mixed tutorials

  • 00:50:33 and videos but mostly about Stable Diffusion recently.

  • 00:50:37 You can also follow me on LinkedIn.

  • 00:50:39 I am not an anonymous person.

  • 00:50:41 I am a PhD computer engineer.

  • 00:50:44 So you can follow me on LinkedIn.

  • 00:50:45 You can connect with me on LinkedIn.

  • 00:50:48 I also have a Udemy course.

  • 00:50:50 If you purchase my Udemy course I would appreciate that very much.

  • 00:50:54 You can support me by purchasing my course.

  • 00:50:56 You see I have already 182 students and we are growing.

  • 00:51:01 I thank all of you who purchase my course or support me on Patreon.

  • 00:51:08 It is super important.

  • 00:51:09 You can also follow me on Twitter.

  • 00:51:10 I am sharing a lot of useful information here so you won't regret it.

  • 00:51:11 Hopefully see you in another amazing tutorial video.

Clone this wiki locally