Skip to content

FLUX Tools Outpainting Inpainting Fill Redux Depth and Canny Ultimate Tutorial Guide with SwarmUI

Jump to bottom
FurkanGozukara edited this page Oct 17, 2025 · 1 revision

FLUX Tools Outpainting, Inpainting (Fill), Redux, Depth & Canny Ultimate Tutorial Guide with SwarmUI

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

FLUX Tools Outpainting, Inpainting (Fill), Redux, Depth & Canny Ultimate Tutorial Guide with SwarmUI

image Hits Patreon BuyMeACoffee Furkan Gözükara Medium Codio Furkan Gözükara Medium

YouTube Channel Furkan Gözükara LinkedIn Udemy Twitter Follow Furkan Gözükara

FLUX.1 Tools by BlackForestLabs changed the #AI field forever. They became the number 1 Open Source community provider after this massive release. In this tutorial, I will show you step by step how use FLUX.1 Fill model (inpainting model) to do perfect outpainting (yes this model used for outpainting) and inpainting. Moreover, I will show all features of FLUX Redux model to do style transfer / re-imagine 1 and more than 1 images combination. Furthermore, I will show you step by step how to convert input image into Depth or Canny maps and then how to use them on #FLUX Depth and Canny models. Both LoRA and full checkpoints of FLUX Depth and Canny.

🔗 Full Instructions, Configs, Installers, Information and Links Shared Post (the one used in the tutorial - Public - no-paywall) ⤵️

▶️ https://www.patreon.com/posts/tutorial-instructions-links-public-post-106135985

Preparation of this tutorial took more than 1 week and this will be the very best and easiest to follow tutorial since it is made with famous #SwarmUI. SwarmUI is as easy and as advanced as Automatic1111 SD Web UI. Biggest advantage of SwarmUI is that, it uses ComfyUI as a back-end. Therefore, It is extremely fast, VRAM optimized and supports all of the newest SOTA models as soon as they are published.

So in this tutorial I will show you how to setup SwarmUI and FLUX Dev tools on your Windows Computer, Massed Compute, RunPod and Kaggle. I will step by step explanatin and show you every tips and tricks that you need to properly do style transfer, re-imagine, inpaint, outpaint, depth and canny with FLUX.

Video Chapters

00:00:00 Introduction to step by step FLUX Dev Tools tutorial

00:02:21 Windows: main part of the tutorial for learning everything about SwarmUI and FLUX Tools

00:03:20 How to do a fresh install and update of SwarmUI

00:04:45 How to set automatic back-end update enabled for ComfyUI

00:05:09 How to update your existing SwarmUI and ComfyUI back-end into the latest version properly

00:05:45 How to download newest FLUX Tools models (Redux, Fill, Depth, Canny)

00:06:50 Updated features and mind blowing speed of SwarmUI Unified Model Downloader

00:08:38 How to start SwarmUI again after the installation

00:09:59 How to use Redux (style transfer and re-imagine) FLUX Tool model

00:10:55 How and when Image Prompt screen appears on SwarmUI

00:11:35 How exactly Model Apply Start works

00:12:45 How to see / follow what is actually happening - see logs

00:13:22 How to use Grid system to find best parameters / settings for your usage case

00:14:21 How to use FLUX Redux model to combine multiple images into a new image - style transfer

00:15:34 How to use automatic face inpainting to improve face quality and consistency with face

YOLO model and segment feature

00:17:10 Merging 3 images at once into a new image with FLUX Redux adapter model

00:19:01 How to use FLUX Canny and Depth models

00:19:21 Beginning to show how to super easy use FLUX Canny model

00:21:20 Pay extreme attention to the aspect ratios when using Canny or Depth model

00:24:32 How to use Canny LoRA model

00:25:18 How to use FLUX Depth model

00:26:50 How to use FLUX inpainting (FLUX.1 Fill) model for inpainting task

00:27:33 How to properly use image editor of SwarmUI for inpainting

00:29:05 Best configs of FLUX fill model inpainting

00:30:28 How to make grid comparison to find best parameters for your case

00:30:57 How to remove / erase masking to fix the mask

00:31:51 Full guide for how to outpaint / expand images better than all paid apps for free with FLUX Fill model

00:34:35 Which prompting strategy works better for out-painting images

00:38:55 Where to find and follow us (SECourses) on Discord, Reddit, LinkedIn and GitHub

00:40:01 Massed Compute install and setup

00:40:43 Selecting SECourses image and using our 0.31$ coupon (50% lower price)

00:42:25 Setup ThinLinc client and connect cloud Massed Compute machine and synch and transfer your files

00:45:24 How to update existing SwarmUI and download new FLUX Tools on Massed Compute

00:49:08 How much time downloading over 80 GB of models takes on Massed Compute

00:49:18 How to start SwarmUI on a shared link to use it in your computer, phone, table but run in Massed Compute

00:49:55 How to utilize multiple GPUs in SwarmUI - e.g. dual GPU setup

00:50:40 How to terminate machine to not use any credits

00:51:05 RunPod install and setup for FLUX Tools and SwarmUI

00:52:28 Which template to use on RunPod for SwarmUI and FLUX Tools

00:52:56 How to set a RunPod port to connect via RunPod port based proxy system through your browser on any device

00:53:51 How to install SwarmUI on RunPod - fresh installation

00:55:52 How to install onnxruntime on RunPod to fix onnxruntime error that you get

00:56:21 How to download new FLUX Tools models on RunPod for SwarmUI

00:59:27 How to start using SwarmUI after FLUX Tools models downloaded and all installed, refresh models

01:00:27 Kaggle install and setup for FLUX Tools and SwarmUI

Video Transcription

  • 00:00:00 Greetings everyone. Today I am going to show you  how to use the newest FLUX models on SwarmUI very  

  • 00:00:05 easily with the maximum performance and  the quality and the best configurations.  

  • 00:00:11 I will begin with showing the FLUX Redux model,  re-imagine input images like this case. Moreover,  

  • 00:00:18 I will show you how to combine multiple images  with FLUX Redux model and re-imagine them like  

  • 00:00:25 this case. Input images are 3 images and this is  the output of combining 3 images with the FLUX  

  • 00:00:30 Redux model. Furthermore, I will show you how to  convert any image into a Canny map, then use the  

  • 00:00:36 FLUX Canny model and generate new images from the  input images like this one. This is an amazing  

  • 00:00:42 image. Moreover, I will show you how to convert  any input image into the depth map then use the  

  • 00:00:49 FLUX Depth model and generate new images based  on the depth of that input image. Furthermore,  

  • 00:00:55 I will show you how to do in-painting properly  and in-paint any part of any image with such an  

  • 00:01:02 amazing quality that you have never seen before  with the inpainting inpainting model of the FLUX  

  • 00:01:07 Tools. And finally, I will show you how to do  outpainting properly. This was something that  

  • 00:01:13 I have been waiting for. The outpainting quality  of the FLUX Dev is non-challenged. There is no  

  • 00:01:20 other even paid services to outpaint like  this one. And this is all running locally  

  • 00:01:26 on your computer. What if you are a GPU poor,  if you don't have a powerful GPU? Don't worry,  

  • 00:01:32 I will show you how to install and set  up SwarmUI with FLUX Tools on RunPod,  

  • 00:01:37 on Massed Compute, which is my favorite platform  with amazing price and amazing speed. And finally,  

  • 00:01:45 I will show you how to use SwarmUI and FLUX Tools  on a free Kaggle account. If you don't want to pay  

  • 00:01:52 to the RunPod or Massed Compute, Kaggle is way  to go. This tutorial will be a public tutorial,  

  • 00:01:58 you will be able to learn how to use FLUX  Tools totally free without any paid content,  

  • 00:02:03 without any paywall. And I will also show how to  generate grids to find the best hyperparameters  

  • 00:02:10 for your cases. I already have prepared some best  hyperparameters. But when you learn how to use the  

  • 00:02:16 grid system of the SwarmUI, you will get even  better images. So as usual, I have prepared an  

  • 00:02:23 amazing public post. This post is 100% public.  You don't even need to login into the Patreon,  

  • 00:02:31 register into Patreon to see this post. As I  promised, this is going to be a fully public  

  • 00:02:36 tutorial. I have updated my SwarmUI playlist.  So click this link. And if you are new to the  

  • 00:02:44 SwarmUI, you can watch them from beginning to the  end. This is the oldest one. This is the newest  

  • 00:02:50 one. And each one of them is a public tutorial.  Since there weren't any changes. For installation  

  • 00:02:57 and setup of the SwarmUI, I recommend you to  watch these tutorials. Because I am going to  

  • 00:03:03 make this one, the one you are watching right now,  shorter than the other ones at the installation  

  • 00:03:10 part. They are also individually linked here. So  I will quickly begin with installation and update  

  • 00:03:18 if you already have SwarmUI. The files that you  need are located in the attachments. Moreover,  

  • 00:03:26 I will show how to use the FLUX tools, how  to download models and install on RunPod and  

  • 00:03:32 Massed Compute as well. However, I am not going  to show how to use at that part of the tutorial.  

  • 00:03:39 So to learn how to use FLUX tools, you need to  watch this part. This Windows usage part. Then  

  • 00:03:46 you can look at the Massed Compute and RunPod part  to how to install them. So the Windows installer  

  • 00:03:54 is here. You see Windows Installer bat file.  Remember, to be able to use SwarmUI on Windows,  

  • 00:04:00 you need to have .NET 8 from here installed. So I  am going to install into R drive into swarm_tools  

  • 00:04:08 folder. Double click. More info run anyway.  It will automatically start the installation.  

  • 00:04:13 This is same as other tutorials so it will  be fast. I agree. Customize settings. Modern  

  • 00:04:19 dark. Just yourself on this PC. ComfyUI local.  I am not going to download anything. Next. Yes,  

  • 00:04:26 I am sure. Install now. So the installation is  getting completed. In the logs you can always see  

  • 00:04:32 what it is doing. It is installing the necessary  libraries in the initial install. And there is one  

  • 00:04:39 very important thing. In the backends. In the  ComfyUI. It is automatically added here. You  

  • 00:04:45 should have always update to have always updated  ComfyUI backend. Because I am getting a lot of  

  • 00:04:53 complaints that the newest models are not working  because their ComfyUI back-end is not updated. So  

  • 00:04:59 when this option is selected it is going to update  the ComfyUI backend to the latest version whenever  

  • 00:05:06 you restart your SwarmUI. So the installation has  been completed. Now I will show how to update your  

  • 00:05:12 existing SwarmUI installation before beginning  this tutorial. It is inside here. Enter inside  

  • 00:05:19 the SwarmUI installed folder and double-click  start the windows update.bat file. It will update  

  • 00:05:25 the SwarmUI to the latest version. Moreover, if  your ComfyUI is not automatically updated for any  

  • 00:05:32 reason, in the ComfyUI update folder you will see  ComfyUI update.bat file. Double-click it. It will  

  • 00:05:39 update it to the latest version. Now the SwarmUI  is ready. To use the FLUX Tools, what you need?  

  • 00:05:46 You need to download the FLUX Tools models. 28  November update. Which files you need? Whichever  

  • 00:05:53 you want to use. But I recommend to download all  of them and test all of them. So the FLUX.1 Redux  

  • 00:06:00 dev file goes into this folder. The other one goes  into this and this one goes into this. However,  

  • 00:06:06 I recommend to use SwarmUI unified downloader  which I am going to use. So click here. I am  

  • 00:06:12 keeping this downloader very up to date and I am  adding the models as you request them. Go to the  

  • 00:06:19 bottom attachments. Download the latest zip file.  Move that zip file into the parent folder of your  

  • 00:06:27 SwarmUI folder like this. Right-click and extract  all. I am going to extract. So you see they are  

  • 00:06:33 extracted into a new folder. So right-click and  cut them. Move them into here. This folder. Then  

  • 00:06:40 you are ready. So double-click Windows download  models.bat file. More info run anyway. It will  

  • 00:06:47 list you all of the possible options. So for this  tutorial what I am going to do is. I am going to  

  • 00:06:53 use FP16. It will be automatically casted into FP8  when I am using because my GPU is not sufficient.  

  • 00:06:59 So option 1 comma. However if you have RAM limited  you should download the FP8 FLUX dev model. The  

  • 00:07:07 scaled may not work with the newest FLUX tools.  So you should download option 10 if you don't have  

  • 00:07:15 64 GB of RAM memory. Then I am going to download  FP8 version of the T5. This is scaled version as  

  • 00:07:22 you are seeing. Option 6. Then I am going to  download the newest FLUX tools. They are all  

  • 00:07:29 here as you are seeing. Option 13, 14, 15, 16, 17,  and 18. You can also download the automatic face  

  • 00:07:39 inpainting yolo model which is here. Option 11.  And we are ready. Then hit enter and it will start  

  • 00:07:46 downloading all of the models with the maximum  possible speed of your internet connection. I  

  • 00:07:53 have optimized the installer for this task. You  see it is downloading with over 150 megabytes  

  • 00:07:59 per second. This is on my personal computer. If  you are installing on the cloud services it is  

  • 00:08:05 even over 1000 megabytes per second. So it depends  on your internet speed and your hard drive speed.  

  • 00:08:12 But this will download with the maximum possible  speed. It is not like downloading from your  

  • 00:08:18 browser. This will be way faster way way faster.  Okay. So all of the models have been downloaded  

  • 00:08:25 under 20 minutes. Folder verified. Downloads  have been completed. Now we are ready to start  

  • 00:08:32 using the FLUX tools with the SwarmUI locally on  our Windows computer. So close this. Go back to  

  • 00:08:39 your SwarmUI installation. This is the folder and  to launch the SwarmUI we are going to use launch  

  • 00:08:45 windows.bat file as usual. It is going to start  the SwarmUI. So we are both updated to the latest  

  • 00:08:53 version of SwarmUI and the ComfyUI backend. And  we have downloaded the FLUX tool models and the  

  • 00:09:00 FLUX dev model. I'm also hopefully going to add  FP8 version of these FLUX tools when they become  

  • 00:09:07 available. FP8 version is especially important for  low RAM users. Because otherwise it will load into  

  • 00:09:16 RAM memory and then it will cast it into the FP8  and it will take time if you don't have a high RAM  

  • 00:09:23 memory. So I'm going to begin with showing you how  to use the FLUX Redux models. As I have shown, I  

  • 00:09:32 have made a lot of a lot of research over 3 days,  hundreds, thousands of images. So I have prepared  

  • 00:09:40 best working configurations. They are inside the  SwarmUI unified downloader. You see best working  

  • 00:09:47 configurations. If you don't want to watch this  tutorial you can just look at this file to see  

  • 00:09:53 the best configurations. So I'm going to use this  image to generate similar images as a beginning.  

  • 00:09:59 How are you going to use the FLUX redux model?  Drag and drop the image into the prompt bar here.  

  • 00:10:06 So you see, I have dragged and dropped the image  that I want to use as a base into the prompt bar  

  • 00:10:13 like this. Not init image or anything else. CFG  scale is 1. The sampling I let them default. It  

  • 00:10:20 works best. However, I find that FLUX guidance  scale 6 is working best. This is what I recommend.  

  • 00:10:28 And turn off the init image. We are not using  it. Control let everything is turned off. In  

  • 00:10:34 the advanced sampling, since I have lower VRAM,  I am going to make the precision as FP8. So it  

  • 00:10:43 is going to convert the base model into the FP8 on  my GPU, but if you are using it on a high VRAM GPU  

  • 00:10:51 like 48GB I recommend FP16. And after I did load  image into here as a prompt, you will see that  

  • 00:10:59 a new screen appeared. Image prompting. You  see it is enabled. So in here we are going to  

  • 00:11:05 select FLUX.1 Redux dev.safetensor file. This  is super important. Base model is FLUX dev and  

  • 00:11:12 this is the style model. Now which style model  merge strength you need? I find that 0.4, 0.5,  

  • 00:11:22 and 0.6 are the best ones. I cannot say which one  will work best. So I recommend to test all three  

  • 00:11:29 of them. I usually find that 0.6 is working best.  But you need to compare. Model apply start. This  

  • 00:11:37 is super important. Since it is percentage  based but it only can works with integers  

  • 00:11:42 because it is depending on the steps count. You  need to calculate which step you want to start.  

  • 00:11:48 If you want to start with the step 1 then you  can enter something like this. It will start  

  • 00:11:53 with the step 1. If you want to start with the  step 2 and if you are doing 40 steps then it is  

  • 00:11:59 2 over 40. So it is this value. If you set this  model apply start at this, it is going to start  

  • 00:12:07 at the step 2. However I find that starting model  start at 0 is working better. So it is up to you.  

  • 00:12:13 You can compare it. Like enter 0.01. It will start  at the step 1. If you enter 0.05 it will start at  

  • 00:12:22 the step 2 according to the number of steps  you are doing. But I find that 0 is better.  

  • 00:12:29 And finally we need to define our prompt.  What we want to generate similar image to  

  • 00:12:35 this. I am going to generate a dragon bird.  Something like this. And the seed is random.  

  • 00:12:42 Let's generate the first image. So if you want to  follow what is happening. Go to the server. Go to  

  • 00:12:48 the logs. Go to the debug from view. And you  will see what is happening. It is downloading  

  • 00:12:53 the clip vision automatically since this is  first time running on a fresh installation.  

  • 00:12:59 It started the generation. I have RTX 3090 and it  was 1.6 seconds per it. If you have RTX 4090 it  

  • 00:13:08 would be like 3 times faster. So this is the image  generated. I think it is looking pretty pretty  

  • 00:13:16 amazing. So as I said. To be sure which style  model merge strength is best. You can use this.  

  • 00:13:23 It is so easy. Go to the tools. Grid Generator.  And enter here: Style Model Merge Strength. So  

  • 00:13:31 try 0.4, 0.5, 0.6. And generate a grid. You can  also give a name. Like test1. And generate grid.  

  • 00:13:40 OK. So the grid generation has been completed.  Click here to open it also from loadgrid config.  

  • 00:13:46 You can select Load grid config and open. Let's  open. So the grid is here. The first one is 0.4.  

  • 00:13:54 This is the generation. 0.5. And 0.6. So you can  compare. You can pick the best one. In this case,  

  • 00:14:02 I think the best working one is 0.5. Because I  see some errors in the other two. Like here. Or  

  • 00:14:11 here. So I like the 0.5 best. But it is up to you.  Whatever you want. And also I have entered a very  

  • 00:14:19 very basic prompt. Second usage case of the Redux  is combining multiple images. So I am opening a  

  • 00:14:28 new tab. And I am going to combine three images.  So the first image will be this one. This is an  

  • 00:14:34 image that I have generated with my trained model.  So I am going to combine this image with a style  

  • 00:14:41 image. Move it here. Then. I am going to move the  style image which is this one. However to keep the  

  • 00:14:49 face consistency. I am going to use the trained  model as a base model. Which is here. This one.  

  • 00:14:57 And if you wonder how you can train yourself  with such quality. This is the tutorial that  

  • 00:15:03 you need to watch. Watch this tutorial. And you  will be amazingly train yourself and generate  

  • 00:15:10 amazing quality images as that one. So what else  I need. When combining styles you need to prompt  

  • 00:15:17 them accordingly. So this is a real image. A  person. And this is an anime style image. So  

  • 00:15:24 this is the anime style image that I am going to  use. Therefore I am going to say anime ohwx man  

  • 00:15:31 in a jungle. Then I am going to inpaint the face.  For inpainting face I am going to use YOLO model.  

  • 00:15:39 Segment YOLO. And the model that I am going to use  is man face. The values is like 0.6, 0.5. So the  

  • 00:15:50 0.6 is how much denoise I want to have. And I am  going to say anime ohwx man. The image prompting  

  • 00:15:58 is ready. It is selected. Let's make the strength  as 0.5. Everything else is same as before. And now  

  • 00:16:06 we are ready to generate. So hit generate and just  wait. Okay. The image has been generated. However,  

  • 00:16:13 you see that I have changed the YOLO face  segmentation model because the other one  

  • 00:16:21 gave an error. I have reported that. So I have  used the classic after detailer trained YOLO face  

  • 00:16:30 model. This face model information exists in the  main post that I have shown you in the beginning.  

  • 00:16:38 So here the image that has been generated. If the  resemblance is not that great. You can modify your  

  • 00:16:45 prompt. You can generate more images. Or you can  even use the variation seed and set it like 0.05.  

  • 00:16:54 So it is going to change the image only 5%. So  you know the drill. You need to generate more  

  • 00:17:00 images to get the perfect accuracy. It is still  very decent. However, it can be way better if I  

  • 00:17:06 want and if I generate more images. So what about  merging 3 images at once? The logic is totally  

  • 00:17:16 same. All you need to do is adding the third  image. You see. Now I can merge 3 images. You  

  • 00:17:23 don't even need to change the prompt. But you  can also change the prompt. So let's make the  

  • 00:17:28 seed minus 1 to get a random one. And I am going  to add the car to the prompt with a car. Since  

  • 00:17:37 this is anime, these two styles are anime, I am  adding the anime word here. And let's generate  

  • 00:17:43 again. To see how it is going to merge 3 images  into the one image with the FLUX Redux. As you  

  • 00:17:50 combine more images, the speed drops. Currently  it is 2.25. It was 1.97 when there were 2 images.  

  • 00:18:00 By the way, this YOLO model is also automatically  downloaded by the SwarmUI Unified Downloader that  

  • 00:18:07 I have made. Currently inpainting the face. It is  already good. But it is going to be better. So if  

  • 00:18:13 your face is better initially, you will get even  more better image when you inpaint the face. So  

  • 00:18:21 the base initial image that you get matters most.  And the image has been generated. As you can see,  

  • 00:18:29 all 3 images are now merged into a single  new concept. And this was the third image  

  • 00:18:37 that we have used. When you pay attention,  you will notice that the third image is also  

  • 00:18:43 excellently put and combined into the image. So,  you see this was its style, its environment. And  

  • 00:18:51 we can see that environment is put there into  the final composition of the image. So you can  

  • 00:18:57 even combine more images. This is how the Redux  model works. Now it is time to move how to use  

  • 00:19:04 the FLUX Canny DEV model and the FLUX Depth DEV  model. You can also use their LoRAs, however,  

  • 00:19:12 their LoRAs are extremely low quality, so I  don't recommend to use them. I compared them,  

  • 00:19:19 but I will still show them how to use. Let's begin  with using the FLUX Canny model. And finally,  

  • 00:19:27 after those two, I am going to show how to do  inpainting and outpainting. So keep watching.  

  • 00:19:33 To use the FLUX Canny dev model, we need a Canny  preprocessed image. And also, we need to select  

  • 00:19:40 the accurate model. So let's open a new tab. And  I am going to load the FLUX Canny dev model load  

  • 00:19:46 now. When you select it, it will also load it when  you generate. But I'm just speeding up the things.  

  • 00:19:52 And you see the image prompt has been disappeared.  So as I said, you need to have preprocessed  

  • 00:19:59 Canny image. How we gonna do it? To be able to do  it, I am going to use ControlNet of the SwarmUI.  

  • 00:20:07 So the ControlNet is here. First of all, you  need to have installed ControlNet preprocessor.  

  • 00:20:12 Because we are going to use the preprocessor of  the ControlNet. Click this. And it will ask you.  

  • 00:20:17 Click OK. It is going to install and make it  ready. You can watch the status from the logs  

  • 00:20:24 as usual with the debug option. You will see what  it is doing. And it is installing the ControlNet  

  • 00:20:30 preprocessor right now. So let's just wait until  the backend has been loaded. If it doesn't load,  

  • 00:20:36 you just can restart. But it usually works. OK.  So it has been installed. Now. First, choose the  

  • 00:20:44 file that you want to use as a Canny model. So  I am going to use this dragon bird, duck bird,  

  • 00:20:51 whatever you can call it. Then, which preprocessor  is best? So I have compared the preprocessors and  

  • 00:20:58 I find that Pyre Canny preprocessor is working  best. After selecting it, click preview. And it  

  • 00:21:05 will generate the preprocessed Canny preprocessed  image. It will first download the preprocessor,  

  • 00:21:11 then it will preprocess it, so it will take  some time. And also always it will load the  

  • 00:21:16 model first. OK. So this is the preprocessed  image. There is one another very crucial thing  

  • 00:21:22 that you need to be careful of when you are  generating the preprocessed Canny or depth image  

  • 00:21:28 for the FLUX tools models. You see this  image, this input image has 3 to 2 aspect  

  • 00:21:34 ratio. Therefore. If I generate the Canny  map or depth map with aspect ratio of 1 to 1,  

  • 00:21:41 let's see what happens. So I'm going to select  the Canny preprocessor from here, this one,  

  • 00:21:46 and preview. And let's see the result. And this  is the result. So you see that image turned into  

  • 00:21:54 this one. This is looking very bad. So how to  fix it? Pay attention to the aspect ratio here  

  • 00:21:59 or the resolution and select the aspect ratio if  it exists here. Or set custom resolution to the  

  • 00:22:07 any resolution that you need. So since this is  3 to 2, I make it 3 to 2 and I will preview and  

  • 00:22:13 generate again. And you will notice that the newly  generated Canny image is way better. So this was 1  

  • 00:22:20 to 1 and this is 3 to 2. You see 3 to 2 is looking  way better, way more accurate. This applies to the  

  • 00:22:28 depth preprocessor as well. So pay attention to  the input image aspect ratio when you are using  

  • 00:22:37 preprocessors to generate Canny or depth map to  use them inside FLUX tools. And this is the depth  

  • 00:22:44 map of the same image with 3 to 2. Now we are  not going to use the ControlNet. We are going  

  • 00:22:51 to use this as an input image. So drag and drop it  into here and then use as init. Then turn off the  

  • 00:23:00 ControlNet. Don't forget it. It doesn't work with  ControlNet. And you will see the init image is  

  • 00:23:05 ready. Now what you need to do is set this as init  image creativity 1. Don't forget that. And you  

  • 00:23:12 don't need to change anything else here. I have  compared everything. All of this information is  

  • 00:23:18 shared in this txt file. Then what else you need  to do? You need to set the FLUX CFG scale to 30.  

  • 00:23:28 Like this. This scroller will go up to 10 but  you can type it here as 30. And that's it. This  

  • 00:23:35 is working best. I have compared all of them.  Don't worry. And let's generate an image but  

  • 00:23:42 let's first write our prompt a dragon bird like  this. Ok generate. Now it is going to generate an  

  • 00:23:49 image based on this. Because this is used as an  init image and the FLUX canny model is selected.  

  • 00:23:56 The generation speed is even faster than the  regular dev model. You see 1.35 seconds per  

  • 00:24:04 it. And the image has been generated. This is the  generated image. It's a decent quality compared  

  • 00:24:10 to the simple prompt that we have entered. And  more importantly this was the base image. You see  

  • 00:24:18 it is matching it very accurately when you think  about how canny model works. So it is keeping the  

  • 00:24:24 structure, the environment. It's a pretty accurate  image. And when we use the LoRA model now you will  

  • 00:24:31 see the difference. So to be able to use the LoRA  model I am going to use the same seed parameter  

  • 00:24:37 here. And I am going to select FLUX 1 dev. I am  going to go to LoRAs and I am going to select  

  • 00:24:43 FLUX canny dev LoRA. Which LoRA strength you  should use. Unfortunately none of them is working  

  • 00:24:50 very good but let's use like 0.97 and generate.  So remember as a model we select FLUX 1 dev and  

  • 00:25:00 as a LoRA we select the canny dev LoRA. LoRA canny  is also exactly same speed there is no speed delay  

  • 00:25:08 added when we are using it. And this is the image  of the LoRA canny. Unfortunately it is very bad  

  • 00:25:15 therefore I do not recommend it. Now how we are  going to use the FLUX depth model. So let's cancel  

  • 00:25:22 this LoRA. Let's go to the models and let's select  the FLUX depth. You see here depth dev model. You  

  • 00:25:30 can also see the selected model here. Right click  and load now. To be able to use the depth model  

  • 00:25:35 we need to have depth image. So to get the depth  image I am also going to use the ControlNet again.  

  • 00:25:42 Go to ControlNet and select the preprocessor  depth. You see there are several depths I have  

  • 00:25:48 compared all of them. The depth anything and depth  anything version 2 working best. Let's go with  

  • 00:25:54 depth anything version 2. This is a little bit  more accurate. Click preview. First time it will  

  • 00:26:01 download the depth model then it will generate  the preview as usual. So the depth mask of the  

  • 00:26:08 image has been generated. Drag and drop it here  and use as init. Then turn off the ControlNet.  

  • 00:26:15 So what is changing? Nothing. Everything is  same as the FLUX canny. And let's generate  

  • 00:26:22 the image and let's see the output of the depth  model. The speed of depth model is same. You see  

  • 00:26:28 1.35 second IT on RTX 3090 Ti. Okay so the image  has been generated and depth model is working way  

  • 00:26:37 better way way better than the canny model. So  this is how you use depth and canny models of  

  • 00:26:44 the FLUX tools. Now as a next step I am going to  show the FLUX inpainting. As an inpainting model  

  • 00:26:51 we are going to use FLUX 1 fill dev model here.  This one. Load now. You see the fill dev model  

  • 00:26:57 is selected. This is used for both inpainting and  outpainting of the images. Now this part is really  

  • 00:27:05 important because the interface of the SwarmUI is  a little bit harder to use for image inpainting  

  • 00:27:12 or image editing overall. But I am going to show  you everything. Don't worry. So first of all we  

  • 00:27:19 need to choose our editing image. To choose your  editing image I am going to use this image as an  

  • 00:27:25 example. You see this was for Invoke AI. This was  also generated with my DreamBooth / Fine-tuned  

  • 00:27:32 model. So click edit image and it will open the  image editor window. So I am using my mouse wheel.  

  • 00:27:42 You really need your mouse wheel otherwise it will  be super hard. You see the orange window here it  

  • 00:27:49 is the camera. So for inpainting it is pretty  straightforward. You need to have mask here. So  

  • 00:27:56 select this mask. Make sure that it is selected  the mask here. Then you need to use paint brush  

  • 00:28:03 and mask the part where you are going to inpaint.  Remember this is totally dependent on your input  

  • 00:28:10 image resolution. Therefore if your input image  is high resolution it is going to use huge amount  

  • 00:28:15 of VRAM. Don't forget that. All right. Then what  we are going to do is let's make the seed random  

  • 00:28:22 and I am going to select the paint brush. You  see there is now new icons at the bottom. Which  

  • 00:28:29 is radius. I am going to make the radius bigger  and I will zoom with my mouse wheel. And I am  

  • 00:28:36 going to mask it like this. Select the paint brush  again and mask all the area that you want. You can  

  • 00:28:41 make a perfect mask this is up to you. Now we are  ready to do inpainting. What you want to inpaint  

  • 00:28:48 I am going to inpaint into like this. SwarmUI  text on a board. Like this. SwarmUI text on a  

  • 00:28:56 board. You can also add some prompts. Like. Let's  say glowing and burning. SwarmUI text. So what  

  • 00:29:03 else we need to change. Again the FLUX guidance  scale is 30 for inpainting. Init image creativity  

  • 00:29:11 is 1. And what else we need. None of these are  selected. And let's use a certain seed because I  

  • 00:29:18 am going to test something else and show you it's  a newest feature. Okay 40 steps. And the sampling  

  • 00:29:25 is default. And init image creativity is 1. And  the FLUX guidance scale is 30. So these are the  

  • 00:29:32 best settings as you are seeing right now. Let's  generate. The inpainting speed is also same as  

  • 00:29:38 you are seeing. And the inpainted image has been  generated. It only inpainted the masked area as  

  • 00:29:46 you are seeing right now. Okay to show the other  example which is mask shrink growth which lets the  

  • 00:29:53 AI to use the surrounding area more accurately.  I am going to inpaint my hair. So let's brush my  

  • 00:30:02 hair. This will be a good example. Okay let's make  it this. Okay let's control-z. Okay. Okay let's  

  • 00:30:11 mask the hair as much as possible. When you are  holding the control key with your mouse wheel you  

  • 00:30:17 can change the radius like this. Okay let's make  the hair like red hair. Something like this. And  

  • 00:30:24 let's generate an image then I will try the mask  shrink growth. Actually let's make a comparison.  

  • 00:30:30 So red hair and go to the tools. Go to the grid  and let's change the mask shrink growth from here.  

  • 00:30:38 So let's search for mask shrink growth. Let's try  this as 0 versus like 128. Shrink test. If you  

  • 00:30:48 get something like this it is not important and  generate grid. Let's see what we are going to get.  

  • 00:30:54 Oh by the way we didn't remove this masking.  So let's just interrupt all sessions. And to  

  • 00:31:00 unmask it there is erasers radius let's erase this  mask now generate grid. Don't forget we'll see the  

  • 00:31:07 masks on the screen here. Yeah I know it's hard to  see but the mask will be here on your screen. Okay  

  • 00:31:13 the grid has been completed let's open it and  let's see the difference. So this is mask shrink  

  • 00:31:19 growth is 0 and this is 128 and when I look at the  difference I think 128 is looking more natural as  

  • 00:31:29 you are seeing right now. So it is up to you. You  can always make comparisons and see which one is  

  • 00:31:35 working better but I like it. It is really really  cool and this is based on the mask that we have  

  • 00:31:42 made and it is really looking natural if you ask  my opinion. So it is up to you. You can compare  

  • 00:31:49 and see which one is working better. So finally I  am going to show you how to outpaint the images.  

  • 00:31:56 Which is one of the most important one. So let's  open a new page. For this one we are using FLUX  

  • 00:32:02 1 fill dev model. Because this is both inpainting  and outpainting. I will show the example on this  

  • 00:32:08 image. So drag it and drop it that image here  and I will use as init. You can also pick the  

  • 00:32:15 init image from choose file from here. Init image  creativity is set 1. This is important. These  

  • 00:32:21 other options are not making any difference with  init image that I have found. FLUX guidance scale  

  • 00:32:28 is again 30 and the more steps you do you get  better results. I have tested that. So we are  

  • 00:32:34 going to do edit image again and this yellow  part is our camera. So to outpaint it I need  

  • 00:32:44 to move this image out of the camera so it will  fill the empty camera area. How am I going to do  

  • 00:32:52 that? First of all I need to select the image from  here. You see I have selected the image. Go to the  

  • 00:32:57 move and then drag and drop the image out of the  camera like this. So I am going to zoom in and I  

  • 00:33:04 will be sure that it is properly positioned like  this. Okay. So you can even outpaint like 40%,  

  • 00:33:11 45%. You can also do small steps. Small steps  would work better. But I just noticed that this  

  • 00:33:18 is not the image that I wanted to outpaint this  was already outpainted. Which one was the init  

  • 00:33:25 image that I wanted to use. It was this one. This  was the original image. So let's click edit image  

  • 00:33:32 again and let's click the move and select the  image. Let's move it. Sorry that this will be  

  • 00:33:38 like two times showing but you understand the  logic. Okay and here. Okay let's zoom in. Okay.  

  • 00:33:46 Yeah I know it is not that so easy but I am going  to outpaint it like 50% right now. Then how we  

  • 00:33:53 are going to outpaint? Click mask and select the  paint brush. Set your radius and mask this corner  

  • 00:34:03 area. Yeah I know it is hard to make a straight  line. It is not that important. And that's it. Now  

  • 00:34:08 you are ready. A bird dragon. So we can make  this random seed or whatever we want. I will  

  • 00:34:15 generate like 8 images. The FLUX guidance scale  is 30 as we set and the image creativity is 1.  

  • 00:34:22 Then are we ready? Yes. Make sure that the grid  is not selected. Generate. Now it will generate  

  • 00:34:28 8 images. Alright the generations has been  completed and there are several very important  

  • 00:34:35 things. The first generations that I have made  with a bird dragon generated unwanted objects  

  • 00:34:43 like this. Why? Because we are really expanding  the picture hugely. Huge percentage. Therefore,  

  • 00:34:52 it is adding unwanted objects like this. If you  define the prompt with an extreme details. I mean  

  • 00:35:00 something like this. I have used the chat GPT  to generate this. Then it is adding additional  

  • 00:35:07 good images as well. If you are looking something  like this. Yes you can use. However. If you want  

  • 00:35:13 to make the image generation something like this.  Like this. Like this. I mean look at this. This  

  • 00:35:20 is just perfect. You can not get this in the paid  Photoshop, in Midjourney, wherever, whichever the  

  • 00:35:28 platform or other model you are using. All of them  is just perfect. Just amazing. So the key to get  

  • 00:35:35 this is not using any prompt. Just empty. When you  leave it empty, the FLUX is amazingly generating  

  • 00:35:45 the rest of the image. Just perfect. Just amazing.  So pick the best one that you liked. I mean all of  

  • 00:35:51 them is just amazing. I can use every one of them.  So I am going to use this one. However this image  

  • 00:35:57 is now incomplete. It is not like the full image.  So how we gonna make it full? Drag and drop this  

  • 00:36:05 image into this area. And then click these three  dots and click store full canvas to history. Then  

  • 00:36:14 it will make it a complete image like this. Just  purely amazing. Now you can continue expanding  

  • 00:36:21 from right from top and make this image fully  complete. So to do that, we need to click X and  

  • 00:36:28 click edit image and the process is exactly same.  Select the image. Click move. Let's move it a  

  • 00:36:36 little bit bottom like this to expand it. But you  see that it is just extremely powerful to expand  

  • 00:36:43 in any direction. So let's make it expand like  this. Then select mask. Click paint tool. Mask the  

  • 00:36:52 border like this. And continue generating images  without any prompt. And it will automatically  

  • 00:36:59 expand our image. This is just mind blowing  quality. You see it started the expansion. And it  

  • 00:37:07 is already looking good from the preview image. I  can see that. If you don't like the preview of the  

  • 00:37:13 image, you don't need to wait fully. Just click  this X icon. And it will cancel the generation as  

  • 00:37:19 you are seeing right now. Then click generate and  it will continue generating. So I will show one  

  • 00:37:25 more time how we were making the final image.  Drag and drop to the camera view. Click three  

  • 00:37:31 dots. Store full canvas to history. Then the image  is ready. Let's just download the image more and  

  • 00:37:39 download. So from what to this image. Original  starting image was this one. So let's see them in  

  • 00:37:47 the same image. Like this. This image. Into this  image. Zoom in. Okay. Magnificent. So this was the  

  • 00:37:55 original image. And this is the final image. Look  at the expansion. There is no such outpainting  

  • 00:38:02 exists as good as this one. We have outpainted  the image into the perfect quality image. The  

  • 00:38:09 environment. The wings. I mean it is just perfect.  From this image to this image. I mean look at this  

  • 00:38:17 quality. Look at this outpainting. This image.  To this image. It is just mind-blowing. This  

  • 00:38:23 image. To this image. Now we have seen everything  regarding the FLUX tools. The usage is exactly  

  • 00:38:31 same on RunPod and Massed Compute. First, I will  show how to install and start using on Massed  

  • 00:38:38 Compute. Then I will show how to install and start  using on RunPod. I am not going to repeat the  

  • 00:38:44 parts that I have shown in the Windows tutorial.  So it is mandatory to use the Windows tutorial  

  • 00:38:51 part. Now we will begin with installing on Massed  Compute. Before starting the Massed Compute part,  

  • 00:38:57 at the top of our main post, if you join our  Discord, I appreciate that you see our Discord  

  • 00:39:04 link is here, when you click it, you will see our  Discord page, we have over 9000 members and I am  

  • 00:39:10 replying every member, whether they are Patreon  supporter or not. Also, if you go to Stable  

  • 00:39:15 Diffusion and Generative AI GitHub repository  from this link, I appreciate that if you Star it,  

  • 00:39:22 if you Fork it, and if you Watch it. If you also  sponsor me, I would appreciate that. Moreover,  

  • 00:39:28 we have a Reddit, you see the link is here, click  it. And if you follow our Reddit, I appreciate  

  • 00:39:34 that very much. Please become follower, subscriber  of our Reddit page, because I am sharing so much  

  • 00:39:41 information here. And finally, you can follow  my LinkedIn account here. So you see this is  

  • 00:39:48 my LinkedIn account, when you open my LinkedIn  account, you will see my account, you can follow  

  • 00:39:55 me on here as well. I appreciate that. Now we are  beginning the cloud part of the tutorial. Okay, we  

  • 00:40:02 are starting to cloud part of the tutorial. This  part is especially useful for people who doesn't  

  • 00:40:09 have a powerful GPU and who wants to scale their  generation speed. So type Massed Compute, search  

  • 00:40:16 for Massed Compute Instructions. The instructions  of the Massed Compute is here, the installation is  

  • 00:40:22 exactly the same, the initial installation, only  the FLUX tools will be different. I recommend you  

  • 00:40:28 to use this link to sign up, I appreciate that.  So after signing up, you can go to the billing and  

  • 00:40:35 set your billing. You see the billing is here,  set like $10 and $10 here, it is up to you to  

  • 00:40:41 set. Once you have the setup, go to the deploy  and in here we are going to select our category  

  • 00:40:48 creator and we are going to select SECourses.  This is our image. This is the most important  

  • 00:40:54 part. What I recommend is RTX A6000 GPU and our  coupon, which is SECourses and then make the GPU  

  • 00:41:03 quantity according to your needs. If you select 1,  it will work fine. If you select 2, you can speed  

  • 00:41:08 up your generation 2x. So select 1 and deploy. And  verify. And then click deploy. However, there are  

  • 00:41:15 no available GPU right now because Massed Compute  is extremely popular because of the features it  

  • 00:41:22 has. However, I am talking with the team and they  are going to add more GPUs hopefully very soon.  

  • 00:41:27 Alternatively, you can use the ALT config if it  is available. It only has lesser RAM memory. The  

  • 00:41:34 coupon works and it is $0.30 per hour after the  coupon applied. If it is also not available, you  

  • 00:41:39 can use other GPUs. Like L40s if it is available  or L40. So you can use any of them. However,  

  • 00:41:47 the coupon will work only with RTX A6000 ALT  config and RTX A6000. So currently I am going to  

  • 00:41:53 use RTX A6000 NVLink. There is 1 available GPU.  The coupon will not work on that unfortunately.  

  • 00:42:01 And click deploy. So once the machine has been  generated, it will start initialization part.  

  • 00:42:07 We are updating our image. Hopefully very soon  it will be up-to-date with CUDA latest version,  

  • 00:42:14 the latest drivers, latest installations. And  the initialization part will become way faster.  

  • 00:42:20 We are working on that. So when you are watching  this tutorial, hopefully it will be way faster to  

  • 00:42:24 initialize. And to connect the initialized virtual  machine, the cloud machine, we are going to use  

  • 00:42:30 the ThinLinc client. It is here. Click that link.  If you did set up this, you are ready. But if you  

  • 00:42:36 didn't, download according to your platform.  You see you can use it on the Windows, Mac,  

  • 00:42:40 Linux, wherever you are using. After download,  just click on it. Click yes. Click next. Click  

  • 00:42:47 next. Install. That's it. And click finish.  Now with the ThinLinc client to connect, if you  

  • 00:42:53 want to synchronize the small files between your  computer and the Massed Compute virtual machine,  

  • 00:42:59 click options. Go to the local devices, uncheck  all, and just check the drives, details. In here,  

  • 00:43:05 remove whatever is available. Click add. So you  are going to add an existing folder from your  

  • 00:43:11 computer. You can also copy paste the folder  path to here, which I recommend. So generate  

  • 00:43:16 a folder in one of your drives. For example,  I have generated this folder. I copy this,  

  • 00:43:22 paste it here. Then I will select the permission  read and write so it can work both way, upload and  

  • 00:43:27 download and click OK. As an optimization,  if your ThinLinc client is working slow,  

  • 00:43:33 uncheck this, select tight, select custom, 9,  allow JPEG compression 0, and set the color level  

  • 00:43:40 medium and SSH compression. This is the speed of  your ThinLinc client slightly, however, it will  

  • 00:43:46 reduce the quality of the image. So I'm going to  make this auto select for now. And just like this.  

  • 00:43:52 And OK. Now we need to wait initialization  to be completed. One of the advantages of  

  • 00:43:57 the Massed Compute is that not only the devices  are great, the speed is great, but also our image  

  • 00:44:03 is specially crafted. It has all the things that  you need, like SwarmUI automatically installed,  

  • 00:44:09 one trainer automatically installed. Forge web UI  automatically installed. And many other things.  

  • 00:44:13 We are adding things that you are demanding from  us. We also have some base model automatically  

  • 00:44:20 installed. It is coming up with the image you  are going to see in a moment. And hopefully we  

  • 00:44:25 are going to update it even into a better version  very soon. So the machine has been initialized.  

  • 00:44:31 So copy the IP address, paste it here like this.  You see, copy the username, paste it here like  

  • 00:44:37 this and copy the password and paste here then  connect. Click continue. Click start. And we are  

  • 00:44:44 into ThinLinc client. What if, if you want to  turn off your computer and connect again, just  

  • 00:44:49 click X and turn off the ThinLinc client. Open  the ThinLinc client again and copy the password  

  • 00:44:55 again. It will remember the last settings. Do not  click end existing session because it is going to  

  • 00:45:01 terminate all of the running applications. You  will not lose the existing data on the server,  

  • 00:45:07 but all the running applications will be  terminated. You only need that option when  

  • 00:45:11 you are having synchronization problems or when  you are not able to access the machine. So this  

  • 00:45:17 is the interface. This is going to get updated. It  is going to become better. The SwarmUI is already  

  • 00:45:23 installed. So to update the SwarmUI, first run  this one. You see, this is for update. When you  

  • 00:45:30 hover your mouse, you will see that it is run  stable SwarmUI update. It says stable SwarmUI,  

  • 00:45:35 but it is the latest SwarmUI. So double click it.  It will update the SwarmUI to the latest version.  

  • 00:45:41 Also it will update the ComfyUI backend to the  latest version since it is automatically selected.  

  • 00:45:46 Just wait. And it will take less than one minute.  Yes, it is updated and started. You see the Massed  

  • 00:45:52 Compute advantage is huge and we are keeping this  image updated, not like the RunPod. However, I am  

  • 00:45:58 not going to use it inside the ThinLinc client. I  am going to use it from my computer. To do that, I  

  • 00:46:04 will start it from here. But first, let's download  FLUX Tools models. You can download the FLUX tools  

  • 00:46:11 models from these links as I have shown in the  beginning of the tutorial. However, I recommend  

  • 00:46:16 your SwarmUI unified downloader. You see it is  here. Click it. Go to the very bottom of the post  

  • 00:46:21 where you will see the attachments. This is the  link. It is here. Download it. Move the downloaded  

  • 00:46:25 attachment into your synchronization folder  wherever you have set. It is inside here. You can  

  • 00:46:30 also download this directly in the ThinLinc client  as well. It is up to you. But I recommend to use  

  • 00:46:35 synchronization for smaller files. Because it is  working fast. So how you are going to access that  

  • 00:46:40 file? Open the home button from here. Go to thin  drives. Go to the shared folder. This is shared  

  • 00:46:46 between your computer and the cloud machine. Then  where is the downloader? The SwarmUI downloader  

  • 00:46:51 is here. It is the latest version. Move it into  the downloads. Never use anything inside this  

  • 00:46:57 synchronization folder. Always copy them into  the machine like this. Into the download sfolder.  

  • 00:47:03 Right click. Extract here. Enter inside the  folder. All of my scripts are in here. They have  

  • 00:47:08 the same format. Open the Massed Compute download  instructions txt file. Then copy this command.  

  • 00:47:14 This is the command that you need. Copy it.  Ctrl C. Open a terminal here. You see three dots  

  • 00:47:20 icon inside the folder. That is important. Open  terminal. Right click and paste. And it will give  

  • 00:47:26 you all the options that you want to download. So  if you want to replicate exactly as I did in the  

  • 00:47:33 Windows tutorial. What you need? Option 1. Since  this is a big machine with 48 GBs of GPU you can  

  • 00:47:40 download FP16 versions. Then the option 5. Then  the option 11. If you want to download the YOLO  

  • 00:47:47 models for face segmentation and auto inpainting  the face. Then 13, 14, 15, 16, 17. Actually you  

  • 00:47:56 don't need LoRA here because it is way lower  quality but it is up to you if you want to use  

  • 00:48:01 them on a different base model 16, 17, and 18. Now  look at the download speed. You will be shocked  

  • 00:48:09 by the speed. It is mind-blowingly fast. We are  seeing right now over 600MB per second. Sometimes  

  • 00:48:18 800MB. So you see it is going to take less than 1  minute to download 24GB of the base model. This is  

  • 00:48:26 why I like the Massed Compute. They are just the  highest quality cloud service that you can find.  

  • 00:48:33 They are also very responsive. They also have  staff members in our discord. We can also chat  

  • 00:48:37 with them. We can complain to them. We can ask  questions to them. And they are really really  

  • 00:48:43 helpful. So you see the base FLUX1 dev model took  only 38 seconds to download. 24GB of file only  

  • 00:48:51 downloaded in 38 seconds. So you can become ready  in less than 2 minutes after the machine has been  

  • 00:48:59 initialized. This is why the Massed Compute is  the number one. And we are almost done. We are  

  • 00:49:03 downloading the latest models. So the models have  been downloaded. It took less than let's calculate  

  • 00:49:10 3 minutes. All of the models are downloaded. I  am going to start the SwarmUI on the shared link.  

  • 00:49:17 You see in the left bottom there is run cloudflare  stable SwarmUI. This is going to start the SwarmUI  

  • 00:49:24 on cloudflare. We are going to see a link here.  Let me zoom. Okay you see here. This is the link  

  • 00:49:30 that you need to start. Click this link. It will  open a new page like this. Then copy this link and  

  • 00:49:38 move back to your computer. So now I can use the  SwarmUI in my computer browser. But it will run  

  • 00:49:46 on the Massed Compute. The rest is exactly same  as the Windows tutorial part. There is only one  

  • 00:49:53 difference that I recommend. Since this is a dual  gpu. Add another ComfyUI backend to be able to  

  • 00:49:59 generate multiple images. Copy this. Click ComfyUI  self starting. Click ok. And paste it here. You  

  • 00:50:06 see. I am duplicating. Then set the gpu id 1.  So now when you generate multiple images like  

  • 00:50:13 10 image it will use both of the GPUs at the same  time since there are 2 GPUs on this machine. When  

  • 00:50:21 I go to the server info I can see that. So how you  can backup your files. You can upload them to the  

  • 00:50:26 Hugging Face, onedrive, google drive. If they are  not very big you can also use the synchronization  

  • 00:50:32 folder. Once you are done with the usage you need  to terminate your machine. Do not forget that.  

  • 00:50:39 Because if you don't terminate it will keep using  your credits. You see everything else is as same  

  • 00:50:45 as using on the Windows computer here. Therefore  I will just skip the rest. And I will terminate  

  • 00:50:52 the machine to show you. How you are going to  terminate from here. If you click the stop.  

  • 00:50:56 It will not stop using your credits. So you need  to terminate your machine. And I will terminate.  

  • 00:51:02 And that's it. All the data has been deleted. Now.  I will show how to use the FLUX tools SwarmUI on  

  • 00:51:08 the RunPod. So in our public post search for  RunPod and go to the RunPod instructions. You  

  • 00:51:15 see here. I recommend to read here. Also if you  register with this link. I appreciate that. Click  

  • 00:51:21 this link. Register your RunPod account. Login.  Once you did login. Go to the billing. And set up  

  • 00:51:28 your billing as you wish. Then go to the pods. And  in here click deploy. Now. You can pick any GPU  

  • 00:51:35 to use. It is up to you. I am recommending secure  cloud because it is way faster than the community  

  • 00:51:40 cloud. But you can also go with community cloud if  you are going to use it for a long time. Because  

  • 00:51:45 the initial setup is slow but the rest is same  speed. So therefore it is up to you. If you are  

  • 00:51:50 short on time. Go with secure cloud. If you want  to use permanent storage from here. I already have  

  • 00:51:56 tutorials for them. You see RunPod permanent  network storage. You see RunPod massive tutorial  

  • 00:52:01 for RunPodCTL. So you can watch them. But I am not  going to show that right now. So deploy. Secure  

  • 00:52:07 cloud. I recommend US Texas 3. That is the fastest  one. But you can use any of them. And let's go  

  • 00:52:14 with L40s GPU. This is a very good GPU. It can  load FP16 since it is 48 gigabytes. But you can  

  • 00:52:22 also go with the RTX 4090. Let's go with dual L40s  GPU. So which template you need to use. The latest  

  • 00:52:31 recommended template is written here. So always  read the RunPod instructions. So change template.  

  • 00:52:37 Type PyTorch. And select the PyTorch version. For  example like this one 2.2.0. Then select the GPU  

  • 00:52:45 count. That how many you want to use. You can also  use single GPU. And click edit template. Set the  

  • 00:52:51 volume disk according to the models that you are  going to download. Let's download 200 gigabytes.  

  • 00:52:56 Also you need to set 7801 port. To connect the  SwarmUI interface. Set. Overrides. Don't forget  

  • 00:53:04 that. And we are going to deploy on demand. But  I see that this GPU is not available dual on us  

  • 00:53:11 texas 3. So let's go with 4090. Because this is  a tutorial proposal. So it doesn't matter very  

  • 00:53:17 much. And deploy on demand. And my pods. And it  will start the initializing the template. Since  

  • 00:53:24 this is an official template it is super fast  to be loaded. And we are going to install the  

  • 00:53:29 latest version. We are not dependent on any  pre-configured version. We are not using any  

  • 00:53:34 pre-compiled version. We are always using the  latest version. It is also super important to  

  • 00:53:38 use the latest version. So it is ready. Then click  connect. Connect the Jupyter lab interface. In the  

  • 00:53:44 left top you see there is upload icon. So how you  gonna install. The installation is same. In the  

  • 00:53:50 very bottom of the post you will see that there is  install linux.sh file. And install linux fix step  

  • 00:53:57 2. The instructions all written here. I am going  to show in a moment. So upload both of the files  

  • 00:54:03 into your workspace like this. From this icon.  Select them and upload. Then open a new terminal.  

  • 00:54:10 So first install with this command. Copy paste  it. It will start installation. Unfortunately  

  • 00:54:16 RunPod is not as straight as the Massed Compute.  It is a little bit more harder to use. So when you  

  • 00:54:22 see this local URL. That means that it has been  started. Go to my pods. Then click connect HTTP  

  • 00:54:31 Port 7801. If it is not appeared. Refresh this  page. And click connect. And connect whether  

  • 00:54:37 it is orange or not. Then you will get to this  installer agree. Customize settings. Modern dark.  

  • 00:54:44 Whatever you wish. Next. ComfyUI local. I am not  going to download anything. Next. And I am sure  

  • 00:54:49 install. The installation speed totally depends on  the pod that you got. Some pods could be broken.  

  • 00:54:55 They could be very slow. So RunPod storage system  is not as fast as Massed Compute storage system.  

  • 00:55:03 However they have permanent storage system. And on  RunPod you can stop the machine. And it will use  

  • 00:55:08 very little amount of credits. And you can start  and again use wherever you are left. Unfortunately  

  • 00:55:14 this is not available on Massed Compute yet. But  however hopefully it is coming soon. So you can  

  • 00:55:19 monitor the installation status from here or on  here. From the usual cmd window. Just wait here  

  • 00:55:26 for installation. Initial installation has  been completed. You need to wait. Backends  

  • 00:55:30 are still loading. You need to wait until it  is fully loaded. Go to the logs. Debug. After  

  • 00:55:35 the initial installation it is going to install  the missing libraries. So wait until everything  

  • 00:55:41 is installed and backends are fully loaded. All  right. So the installation has been completed.  

  • 00:55:47 We can see that backends loaded. Now there is  second part of installation. This is for to  

  • 00:55:53 fix onnxruntime error. So this is not mandatory.  But if you want. Copy this. Go back to your pod.  

  • 00:56:01 Open a new terminal. And copy paste. It is going  to install the missing ONNX library. After that  

  • 00:56:09 go to the server and backends. Turn it off. Then  turn it on. So the changes will be applied. Now  

  • 00:56:16 to be able to use we need to download the models  we are lacking right now. So as I shown in the  

  • 00:56:22 beginning you can download them from here. However  I am going to use the SwarmUI unified downloader.  

  • 00:56:27 This is the best way of using them. Click here  to download. Then go to your workspace like this.  

  • 00:56:34 Click this upload icon. Select the downloaded  zip file. Wait until uploaded. Yes. Then right  

  • 00:56:40 click. And extract archive. And it will extract  the installation. Refresh by clicking here. Open  

  • 00:56:47 the RunPod download instructions txt file.  Then you will see the download command here.  

  • 00:56:53 Copy this. Open a new terminal. Paste it. And you  will get the options. Whichever you want. Since  

  • 00:57:01 this is a 24gb gpu I am going to download  the FP8 version of the FLUX dev as I have  

  • 00:57:09 shown in the windows part of the tutorial. But I  recommend you to get 48gb of GPU and download the  

  • 00:57:16 fp16 version. So we are going to download the  base fp8 version which is here not the scaled  

  • 00:57:23 one because I am not sure if FLUX tools would be  compatible with the scaled. So download 10. Then  

  • 00:57:29 should download the fp8 version of the T5. It is  option 6. And I did hit enter inaccurately. I was  

  • 00:57:38 going to type more but it started downloading.  After the download I will download the rest of  

  • 00:57:42 the models. So if you hit enter accidentally  we will continue. Since the downloader scripts  

  • 00:57:49 that I have made is super optimized you see even  on RunPod we are getting very very good speeds.  

  • 00:57:56 However if you download them in the regular way  they will be slower than this one because this is  

  • 00:58:01 super optimized with special techniques. Moreover  if you want to upload big files here do not use  

  • 00:58:09 this upload file because it will not work for big  files. Only for small files it will work. What you  

  • 00:58:15 can use alternatively is RunPodctl and also this  tutorial is what you need. Upload to Hugging Face  

  • 00:58:23 and download from there. Okay let's continue  to download. So it is continuing to asking us  

  • 00:58:28 which we are going to do. Lets download the best  face models face inpainting. Then the FLUX tools  

  • 00:58:35 which are 13, 14, 15, 16, 17, 18 and hit enter.  However I do not recommend the LoRA of the Depth  

  • 00:58:45 and Canny because as I have shown in the windows  tutorial they are very low quality compared to  

  • 00:58:51 the full checkpoint. But if you want to use Depth  and Canny models on a custom fine-tuned FLUX model  

  • 00:58:59 then you can download the LoRAs and see how they  are performing. So the rest of the models are also  

  • 00:59:04 getting downloaded. We can see the download speed  it is sometimes 900 megabytes sometimes 1 gigabyte  

  • 00:59:10 per second sometimes 300 megabytes. Still really  really fast because I have optimized them hugely.  

  • 00:59:16 You see this is an amazing speed for RunPod we  were never seeing such speeds on RunPod before  

  • 00:59:22 I did this optimization. All right the models  have been downloaded. Now we are ready to use.  

  • 00:59:27 Go back to your RunPod proxy which we have been  connected from here. Then go to models and click  

  • 00:59:34 this refresh icon and it will refresh all the  models. The usage is exactly same as on the  

  • 00:59:41 Windows nothing is different. Whenever you have a  problem read the instructions and you will solve  

  • 00:59:48 them. Maybe watch some of the tutorials. You  can also join our Discord channel and ask me  

  • 00:59:53 questions directly. Our discord channel is  at the very top of the post here. Moreover,  

  • 00:59:58 also follow our subreddit follow me on linkedin  or star and follow me on the GitHub repository  

  • 01:00:05 as well. So on RunPod what you can do? You can  stop your machine and you can keep your data and  

  • 01:00:12 you can start it again. Currently it will use a  very little amount of money per hour as you are  

  • 01:00:17 seeing right now. However if you don't want  any money to be used what you need to do is  

  • 01:00:23 terminate and everything will be deleted and it  is gone. Okay. As a last step I am going to show  

  • 01:00:29 you how to use SwarmUI with all of the supported  models and features on a free account Kaggle. So  

  • 01:00:37 if you don't want to pay to do paid cloud services  and you want to use 30 hours per week GPUs of the  

  • 01:00:45 Kaggle this is the part that you need. Search for  Kaggle in our main post. So control-F and Kaggle  

  • 01:00:54 and you will see free Kaggle account notebook  for GPU poor. Click this link and in this page  

  • 01:01:00 you will see the Kaggle notebook file. Download  it. This notebook file is kept updated. It has all  

  • 01:01:07 the instructions and private tutorial links that  will show you how to install and start SwarmUI  

  • 01:01:16 and how to download models that you need on the  Kaggle. So please use this Kaggle notebook. Once  

  • 01:01:23 you got the Kaggle notebook you need to register a  Kaggle account. It is for free. Verify your phone  

  • 01:01:29 number. Once you verified your phone number you  are ready. Click create new notebook wait until  

  • 01:01:35 the interface initialize. Interface initialized.  Click file, import notebook. Then drag and drop  

  • 01:01:41 the downloaded notebook file. Click import and for  the rest wait until settings updated. Click ok and  

  • 01:01:49 follow the instructions on this notebook. Follow  the videos on this notebook and you will be able  

  • 01:01:55 to use the SwarmUI on Kaggle like you are using  it on your computer. So this is end. Hopefully  

  • 01:02:02 see you in another amazing tutorial video. Please  don't forget to join our discord channel we have  

  • 01:02:07 over 9000 members and I am answering all the  questions. Please also follow us on our GitHub  

  • 01:02:14 repository the link is here. Fork it Star it  Watch it and if you sponsor I appreciate it.  

  • 01:02:20 Please also join our subreddit and please also  follow me on LinkedIn and hopefully see you later.

Clone this wiki locally