Wan 2.2 Complete Training Tutorial - Text to Image, Text to Video, Image to Video, Windows & Cloud

[FurkanGozukara]

Wan 2.2 Complete Training Tutorial - Text to Image, Text to Video, Image to Video, Windows & Cloud

Full tutorial: https://www.youtube.com/watch?v=ocEkhAsPOs4

Wan 2.2 training is now so easy. I have done over 64 different unique Wan 2.2 trainings to prepare the very best working training configurations for you. The configurations are fully working locally with as low as 6 GB GPUs. So you will be able to train your awesome Wan 2.2 image or video generation LoRAs on your Windows computer with easiness. Moreover, I have shown how to train on cloud platforms RunPod and Massed Compute so even if you have no GPU or you want faster training, you can train on cloud for very cheap prices fully privately.

📂 Resources & Links:

Download the One-Click Installer & Configs: [ https://www.patreon.com/posts/Musubi-Tuner-Trainer-App-Configs-137551634 ]

Qwen Image Model Training Tutorial (Prerequisite): [ https://youtu.be/DPX3eBTuO_Y ]

SwarmUI & ComfyUI Setup Guide for Windows: [ https://youtu.be/c3gEoAyL2IE ]

SwarmUI Installer and Model Downloader : [ https://www.patreon.com/posts/SwarmUI-Install-Download-Models-114517862 ]

ComfyUI Installer : [ https://www.patreon.com/posts/ComfyUI-Installers-105023709 ]

SwarmUI & ComfyUI Setup Guide for RunPod & Massed Compute: [ https://youtu.be/bBxgtVD3ek4 ]

Upload / Download Big Files Guide for RunPod & Massed Compute: [ https://youtu.be/X5WVZ0NMaTg ]

⏱️ Video Chapters:

00:00:00 Introduction to Wan 2.2 Training & Capabilities

00:00:56 Installing & Updating Musubi Tuner Locally

00:02:20 Explanation of Optimized Presets & Research Logic

00:04:00 Differences Between T2I, T2V, and I2V Configs

00:05:36 Extracting Files & Running Update Batch File

00:06:14 Downloading Wan 2.2 Training Models via Script

00:07:30 Loading Configs: Selecting GPU & VRAM Options

00:09:33 Using nvitop to Monitor RAM & VRAM Usage

00:10:28 Preparing Image Dataset & Trigger Words

00:11:17 Generating Dataset Config & Resolution Logic

00:12:55 Calculating Epochs & Checkpoint Save Frequency

00:13:40 Troubleshooting: Fixing Missing VAE Path Error

00:15:12 VRAM Cache Behavior & Training Speed Analysis

00:15:51 Trade-offs: Learning Rate vs Resolution vs Epochs

00:16:29 Installing SwarmUI & Updating ComfyUI Backend

00:18:13 Importing Latest Presets into SwarmUI

00:19:25 Downloading Inference Models via Script

00:20:33 Generating Images with Trained Low Noise LoRA

00:22:22 Upscaling Workflow for High-Fidelity Results

00:24:15 Increasing Base Resolution to 1280x1280

00:27:26 Text-to-Video Generation with Lightning LoRA

00:30:12 Image-to-Video Generation Workflow & Settings

00:31:35 Restarting Backend to Clear VRAM for Model Switching

00:33:45 Fixing RAM Crashes with Cache-None Argument

00:35:13 Dual Model (High & Low Noise) Training Setup

00:36:54 Preparing Hybrid Datasets (Images + Videos)

00:37:40 Manually Editing Dataset TOML for Resolution Control

00:39:53 Setting High Noise Model Paths for Dual Training

00:41:50 Optimization: Block Swap vs CPU Offload

00:43:10 Generating Video with Dual-Model Trained LoRA

00:45:35 Massed Compute: Server Setup & Coupon Code

00:47:00 Connecting via ThinLinc & File Transfer Methods

00:49:12 Massed Compute: Fast UV Installation & Downloads

00:50:27 Loading Configurations on Massed Compute

00:52:18 Troubleshooting: Fixing Config Version Error

00:53:20 Dual Model Training Speed Analysis on Cloud

00:55:40 RunPod: Selecting the Correct Template & GPU

00:57:45 RunPod: Uploading Files & Extracting Archive

00:58:38 RunPod: Terminal Installation & Model Downloads

01:00:26 RunPod: Correct Pathing Syntax & Backslash Fix

01:01:28 Setting Dataset Paths on RunPod

01:03:34 Installing nvitop on RunPod Terminal

01:03:54 Speed Hack: Disabling Numpy Memory Mapping

01:06:00 Terminating Instances & Final Remarks

Greetings everyone! Today I am presenting an epic tutorial on how to train the Wan 2.2 model to generate extremely high-quality, realistic images and videos. This is currently the most advanced model for generating life-like textures and details.

In this comprehensive guide, I cover everything you need to know to train Wan 2.2 on your local Windows computer, as well as on cloud platforms like RunPod and Massed Compute. We utilize the SECourses Musubi Tuner with fully optimized, 1-click presets designed for every GPU range (from 6GB to 192GB VRAM).

🚀 What You Will Learn in This Tutorial:

Wan 2.2 Text-to-Image Training: How to train the Low Noise model for massive detail and realism.

Wan 2.2 Text-to-Video Training: Mastering Dual Model training (Low Noise + High Noise) for superior video consistency.

Image-to-Video Workflow: How to use your trained LoRAs to animate static images.

Cloud Training: Step-by-step guides for Massed Compute (ultra-fast disk speeds) and RunPod.

Performance Optimization: Using FP8 scaling, Block Swapping, and CPU offloading to train on consumer GPUs.

Inference & Upscaling: Using SwarmUI and ComfyUI to generate and upscale content to 4K resolution.

💡 Key Features of Our Workflow:

Auto-Resume & Speed: New UV package installers for lightning-fast setup.

Presets for All GPUs: Configurations included for 6GB, 12GB, 24GB, 48GB, and 80GB+ cards.

Dataset Automation: Auto-resizing and captioning for both image and video datasets.

Video Transcription

• 00:00:00 Greetings everyone. Today I am going to show  you Wan 2.2 trainings. This will be an epic  

• 00:00:08 tutorial. I will cover so many topics, so check  out the video descriptions to learn all. So,  

• 00:00:14 what I am going to show you today? I will show  you how to train Wan 2.2 model to generate  

• 00:00:22 images like these ones. You see, these are  extremely high quality and extremely realistic,  

• 00:00:28 with massive amount of details, textures,  amazing quality, and amazing realism. The  

• 00:00:35 Wan 2.2 is currently the most realistic plus  most advanced model for generating images.

• 00:00:44 But I will not show only how to train images.  I will show also how to generate videos,  

• 00:00:50 how to train video model, so many things. I  will show how to train on your local computer  

• 00:00:56 with our SECourses Musubi Tuner application.  I have prepared configurations for every GPU,  

• 00:01:04 starting from 6 GB GPUs to 192 GB GPUs. The  configurations are all set. They are separated  

• 00:01:12 into qualities, so you can pick your configuration  according to your GPU and start training right  

• 00:01:19 away and get the most amazing results without  doing all the research that I did for you.

• 00:01:25 I have literally done over 64 separate  trainings to find out the best workflow,  

• 00:01:33 find out the best parameters, and prepare these  presets. I have used a cloud machine with 8 B200  

• 00:01:41 GPUs to complete these trainings, analyze them.  Moreover, I have developed this Gradio based  

• 00:01:46 application that you will be able to load the  configuration and start training right away.  

• 00:01:52 Furthermore, I will show how to do training on  RunPod. So if you don't have a powerful GPU,  

• 00:01:57 you will be able to train on RunPod. And I will  show how to do training on Massed Compute. Again,  

• 00:02:02 if you don't have a powerful GPU, you will be  able to train on Massed Compute. For example,  

• 00:02:07 you see with this GPU we are able to train with  like 70 minutes with only 2 dollars in 1 hour.

• 00:02:14 Everything is ready. Some people were asking  me how to train on RunPod and Massed Compute  

• 00:02:19 the Qwen image models. This tutorial  covers the Wan 2.2 training. However,  

• 00:02:25 all of the models that we support in SECourses  Musubi Tuner are exactly same. So you can train  

• 00:02:31 the Qwen image models, Wan models, or the future  models hopefully that will get added like Flux 2,  

• 00:02:37 like Z Image base model, will be  exactly same as in this tutorial.

• 00:02:42 Moreover, I have updated all of our presets  to the highest quality with lowest amount of  

• 00:02:49 time to generate highest quality images and  videos. So all the image generation and video  

• 00:02:56 generation presets are updated and ready  for you to use. To update these presets,  

• 00:03:01 I have done so many different tests. I have  compared all the results of these tests to  

• 00:03:08 generate videos or images. So many tests have  been made. This tutorial is a product of a massive  

• 00:03:15 research. So many different parameter testing, so  many different workflow testing, and all is ready  

• 00:03:20 for you to use right away with 1 click installers.  So you see from bad videos to good videos.

• 00:03:26 This tutorial covers text to image  training with only from images,  

• 00:03:30 text to video training with only images or  plus videos, but videos are not necessary,  

• 00:03:36 or image to video training. I have done so many  testing and training only text to video yields the  

• 00:03:43 best results for text to video generation or image  to video generation. The presets are separated,  

• 00:03:49 so if you are not interested in video generation,  you can only use the text to image generation  

• 00:03:55 workflow. Still, it is working on text to video  generation or image to video generation as well.

• 00:04:01 So all these configs are interchangeable. There  

• 00:04:04 are several tricks. If you want to  generate highest quality images,  

• 00:04:08 you use the text to image configs. If you  want to generate highest quality videos,  

• 00:04:13 you use Wan 2.2 text to video configs. If you want  to generate image to video, still I recommend to  

• 00:04:20 use text to video configs because this is working  better than image to video. It is weird, I know,  

• 00:04:25 but this is my findings. How do I know? Because  I have done so many different testings as you  

• 00:04:32 are seeing. This is a research of a week doing  a lot of comparisons, doing a lot of testing,  

• 00:04:38 experimentation, and all the presets are ready for  you to use right away with the highest quality.

• 00:04:45 So I will show everything: how to install, how  to set up, how to start training both Windows,  

• 00:04:51 RunPod, Massed Compute. And everything is  literally with 1 click install. Let's begin.  

• 00:04:56 So as usual, I have prepared an amazing post  where you will find everything to follow this  

• 00:05:02 tutorial. However, before starting this tutorial,  you need to watch this main tutorial. So open  

• 00:05:08 this tutorial. This tutorial is our Qwen image  models training tutorial. This is a masterpiece.  

• 00:05:16 This tutorial covers so many different topics  as you are seeing right now. So I recommend  

• 00:05:22 you to watch this tutorial and learn how to use  SECourses Kohya Musubi Tuner premium application.

• 00:05:30 So I will download the latest zip file. This  zip file includes our latest installers and  

• 00:05:36 configurations. I will move it into my existing  installation folder. Then I will extract the  

• 00:05:43 files here and I will overwrite existing  files. Don't forget that you can also do a  

• 00:05:48 fresh installation. After extraction, double  click Windows install and update.bat file.  

• 00:05:55 This will update the application to the latest  version. Moreover, now we are using uv package  

• 00:06:01 installer along with pip, therefore it is ultra  fast. So you see my update already completed.

• 00:06:08 Then double click Windows download training  models.bat file. This file will ask you download  

• 00:06:14 which training models. For this tutorial, we need  Wan 2.2 text to video training. If you want to  

• 00:06:21 train Wan 2.2 image to video, then you need to  download it with option 5. Since I already have  

• 00:06:27 downloaded the models, it will just verify their  hash values, and once the verification completed,  

• 00:06:33 we will be ready. This way we are ensuring  that our models are 100 percentage accurately  

• 00:06:39 downloaded, so we will never have any weird bugs,  issues, or problems. If the files were missing,  

• 00:06:46 it would have downloaded them with the maximum  speed with your network connection support.

• 00:06:51 So now we are ready to begin training. I will  double click Windows start up.bat file. This  

• 00:06:58 will start the latest version of the SECourses  Kohya Musubi Tuner premium application. You see  

• 00:07:05 currently we are at the version 24. This is  where you see the versioning. Then go to Wan  

• 00:07:11 models training. So we are going to use this  tab. Whenever you are going to reload a model,  

• 00:07:16 I recommend to refresh, go to the accurate tab,  and then begin loading. Then click this icon  

• 00:07:23 and go back to your installation folder, enter  inside Wan 2.2 training configs. Now there are  

• 00:07:29 3 configuration folders: Wan 2.2 image to video,  Wan 2.2 text to image, and Wan 2.2 text to video.

• 00:07:37 So what are the differences between these 3? From  the naming as you can see, Wan 2.2 image to video  

• 00:07:44 trains image to video model. However, there  is only single configuration because if you  

• 00:07:49 want to only generate images, it is not the  model that you need. This model is for only  

• 00:07:54 generating videos. However, I don't recommend to  use this model for training even if you are going  

• 00:07:59 to use image to video models because training  with Wan 2.2 text to video works better than  

• 00:08:06 training with Wan 2.2 image to video, even  for image to video presets. Even for image  

• 00:08:13 to video generation. So for video generation, I  recommend to always use Wan 2.2 text to video,  

• 00:08:20 and for only image generation, I  recommend to use Wan 2.2 text to image.

• 00:08:24 So let's begin with Wan 2.2 text to image. Now  you will see bunch of configurations like this.  

• 00:08:30 The first numbers are for GPU  VRAM. So if you have 6 GB of GPU,  

• 00:08:36 you need to pick it. If you have 24 GB of GPU,  then you need to pick it. It also shows the  

• 00:08:42 quality. The quality 1 is the maximum quality.  Then the quality as the number increases the  

• 00:08:48 quality get worse. So quality 1 is better than  quality 2, quality 2 better than quality 3,  

• 00:08:54 and so on. Since I have RTX 5090, I am going  to use 32 GB VRAM configuration. However,  

• 00:09:02 you will see that there are 2 versions: 32  GB of FP8 scaled and no FP8 scaled. The FP8  

• 00:09:09 scaled has a big advantage. It is usually faster  because it uses lesser VRAM. Moreover, it uses  

• 00:09:16 much lesser RAM memory. So if your RAM memory is  limited, then always go with FP8 scaled version.

• 00:09:23 So let's open nvitop to see our GPU and RAM  memory. pip install nvitop. nvitop. And this is  

• 00:09:33 my current usages. So I have 60 GB of empty RAM  memory. My first GPU is fully empty. Therefore,  

• 00:09:42 I will pick the very best configuration inside  Wan 2.2 trainings. Wan 2.2 text to image 32 GB.  

• 00:09:50 And it is loaded. You can also click this icon  to be sure it is loaded. Then it is same as in  

• 00:09:57 our Qwen image training tutorial. What I am  going to change here is the training dataset.

• 00:10:02 For text to image training, you only need to use  the images. For text to video training, if you  

• 00:10:09 only use images, it is still working perfect as  I have shown in the beginning of the tutorial.  

• 00:10:15 But you can also include videos. I will show  that later. So I will pick my training dataset,  

• 00:10:21 but first let's prepare it. This is exactly same  logic as in the Qwen images, so this is my folder.  

• 00:10:28 I will delete the other files to show you what  happens as a quick rewinding. But watch the Qwen  

• 00:10:35 image training tutorial. So this is my trigger  word ohwx and this is my parent path. Don't worry,  

• 00:10:42 it will automatically resize them. So this is  it. I copy pasted, I can also pick from here.

• 00:10:48 The best working resolution is 960 to 960.  How do I know? I have literally tested so  

• 00:10:56 many different configurations. So when you  see these configurations 1024, 1280, 1328,  

• 00:11:05 these are the resolutions that I have tested.  And the best yielding resolution is 960 to 960.  

• 00:11:12 Therefore, I recommend that. Then don't forget  to click generate dataset configuration and  

• 00:11:17 it is ready. So when you return back to your  images folder, you should see the txt files.  

• 00:11:23 These are the captions and when you open  one of them you will see only the folder  

• 00:11:27 name is used. If you want to use custom  captions, you need to change these files.

• 00:11:32 Okay, our training dataset is ready. Now we need  to set our output and other stuff same as in the  

• 00:11:38 Qwen image training tutorial. So I will save  them into my test trainings folder like this  

• 00:11:44 where the models will be saved. This is the model  LoRA models file names. Currently fine tuning is  

• 00:11:50 not supported so we can only train LoRA. Okay,  then the model settings. You need to pick the  

• 00:11:55 files for text to image training, we are only use  the low noise. There are 2 versions of Wan 2.2:  

• 00:12:02 high noise and low noise. High noise is the  initial generation which roughly generates  

• 00:12:08 the base of the image or video. The low noise  is the details. For generating only images,  

• 00:12:14 we are only using low noise. This is the higher  quality. So I will pick the accurate models from  

• 00:12:21 downloaded models training models Wan. So  this is the low noise text to video model.  

• 00:12:27 Like this. Then since other files are there, I  will just copy paste the folder paths like this,  

• 00:12:34 like this. But you can individually pick them.  We are not going to use Clip Vision, this is not  

• 00:12:38 used for Wan 2.2 training. And you see there is  high noise. This is used for dual model training  

• 00:12:44 which I will explain after this. So you don't need  to change anything else here. Don't change them.

• 00:12:50 In the training settings, you need to set your  training epochs as I have explained. Since I  

• 00:12:55 have 28 images, I am going to use 200 epochs. If  you have 100 images, then you can reduce this to  

• 00:13:02 100. That is the logic. But up to 50 images, I  recommend to train at least 200 epochs. Moreover,  

• 00:13:08 there is save every N epochs. So we are saving  every 20 epochs. It will save 10 checkpoints that  

• 00:13:15 I can compare later. Then don't forget to save  your configuration. So let's save this as like  

• 00:13:22 this. LoRA tutorial text to image TOML. Okay, I  will save it. Then I will click start training.

• 00:13:30 Then follow the CMD window. If you get  any errors, you need to report me that.  

• 00:13:36 So currently I have an error. Let's see where we  have the error. Maybe we forgot something. Yes,  

• 00:13:41 I have forgotten the VAE path. This can happen.  So let's open all the panels like this and  

• 00:13:49 search for VAE. Okay, the VAE is set here.  Okay, I can see the error that I have made,  

• 00:13:55 so I need to fix it. It will be like this. And  this is also error. Okay, it will be like this.  

• 00:14:00 So pay attention to the model paths save. You can  use the folder icon start training. And you see  

• 00:14:07 we are also supporting torch compile. I am using  your Visual Studio installation automatically.  

• 00:14:14 Therefore, it is super important for you to  follow the requirements tutorial which you  

• 00:14:20 must. So the requirements tutorial are listed  here in the post. Follow this requirements  

• 00:14:26 tutorial first. When you follow the Qwen image  models training, you will know this already.

• 00:14:31 Okay, let's follow our training. So now  it is loading model into the RAM memory,  

• 00:14:36 then it will load into the GPU memory. If you  don't have sufficient amount of RAM memory,  

• 00:14:42 training both models, both low and high noise,  is harder. And it is even not mandatory. You can  

• 00:14:48 also train only low noise model and even generate  videos. It will be a little bit lower quality,  

• 00:14:55 but it will work. Don't worry. So the training  is about to start. Yes, training starting. Okay,  

• 00:15:01 training started. It is not using my entire GPU  memory. This is good. You need to have some free  

• 00:15:07 memory. You will also see that your VRAM usage  is not static like this. This is because how  

• 00:15:12 Kohya developed the Wan 2.2 model training. It  clears the VRAM cache after every step because  

• 00:15:19 it is designed to be work with both models at  the same time during training. So the speed  

• 00:15:25 is around 5 seconds per IT. It will get a little  bit slower. This is how Kohya displays currently.  

• 00:15:31 But it is pretty decent. It is taking like 8  hours on my GPU. If I need faster training,  

• 00:15:37 then I can reduce the resolution or I can increase  the learning rate and I can do lesser epoch. So  

• 00:15:44 let's say you want 100 epochs, then you need to  multiply your learning rate with 2.5 and reduce  

• 00:15:51 the maximum epochs number. However, you will  lose some quality. So this set learning rate  

• 00:15:58 and number of epochs are the best overall. It is  a choice. It is a trade off. Either you need to  

• 00:16:04 reduce the resolution to speed up or you need to  lower the epoch count and increase your learning  

• 00:16:12 rate. So this will train text to image LoRAs for  us which you can use for text to video as well.

• 00:16:18 How we are going to use the trained models?  I will stop the training. So for using the  

• 00:16:23 trained models, we are going to use SwarmUI. The  SwarmUI link is here. So let's also proceed with  

• 00:16:29 it. If you don't know how to use SwarmUI  already, you need to follow this tutorial.  

• 00:16:34 You see this is the tutorial link in  the top of the post. When you open it,  

• 00:16:40 copy the zip file there. So let's download the  latest zip file from here. Then I will move it  

• 00:16:46 into my existing SwarmUI installation like this.  Copy the file there. Right click and extract files  

• 00:16:54 here and overwrite all. Then as a next step, we  will update our ComfyUI installation. So go to  

• 00:17:00 ComfyUI link from here. Download the latest  zip file. You see there are 2 versions with  

• 00:17:06 password protected one. If your antivirus, this  happens with Windows Defender, causes issues,  

• 00:17:13 download with the password protected, then move  your ComfyUI zip file into your folder or you can  

• 00:17:19 make a fresh installation. As usual, make an file  extraction overwrite all files. Then to update or  

• 00:17:27 install Windows install or update comfyui.bat  file. We have moved the installation to uv  

• 00:17:34 packages. So now it is much faster than before.  Both update and installation. So you see it is  

• 00:17:40 all updated. Then return back to your SwarmUI. In  the SwarmUI, we have Windows update swarmui. So  

• 00:17:48 double click it run. It will update SwarmUI and  start your application. Everything was already  

• 00:17:54 explained in other tutorials in more details,  so you should watch them. Okay, SwarmUI update  

• 00:18:00 succeed and it started. You will also get this  error. This is not important, just ignore that.

• 00:18:06 First of all, you need to get the latest presets.  So you can use import, choose file, go back to  

• 00:18:13 your SwarmUI installation and select the latest  preset from here. Then overwrite existing presets.  

• 00:18:20 Alternatively, while the application is running,  you can use the preset delete import. This will  

• 00:18:26 delete your existing presets and just refresh it.  Then click refresh and everything is here. The  

• 00:18:33 presets are important because I have updated the  presets for video generation for Wan 2.2 and also  

• 00:18:40 image generation. So use the latest presets. Now  how we are going to use our trained Wan 2.2 image  

• 00:18:48 generation LoRAs. So first of all, quick tools  reset params to default. Then from the presets,  

• 00:18:53 the preset that you need to use is Wan 2.2  generate realistic images. Click here and  

• 00:18:58 direct apply. This will let you generate 960  to 960 images. So let's type our prompt. Select  

• 00:19:07 your aspect ratio whichever you want. For example,  this one. And then you need to select your LoRA.

• 00:19:14 But there is one more thing. You need to have  models downloaded for this preset to work. So  

• 00:19:19 how you are going to download accurate models?  You see Windows start download models app.bat  

• 00:19:25 file. This will start the model downloader  latest version. We have upgraded this. Also  

• 00:19:30 it is using uv package. So now it is even faster  to start. Okay, the downloader started. So which  

• 00:19:37 models you need to download? You can download  complete image generation and editing bundle.  

• 00:19:42 Or you can download Wan 2.2 core 4 steps  bundle. This is my recommended bundle. So  

• 00:19:49 go to bottom and download all models inside  this bundle. This way you will have all the  

• 00:19:56 models that you need. Even the upscaler models  that you need to use to get maximum quality.  

• 00:20:03 You can follow the download process from the  CMD window. So if I am missing any models,  

• 00:20:09 they will get downloaded. If I already have some  models, they will be verified and it will skip  

• 00:20:14 the download. So currently it is downloading  some of my missing models, but I have the  

• 00:20:19 necessary models for image generation which is  Wan 2.2 text to video low noise. So it is set.

• 00:20:26 Now I need to set my LoRA. So these are the LoRAs  that it finds. It is compatible with Wan 2.2. My  

• 00:20:33 trained LoRA is inside here. Wan 2.2 low noise.  This is my trained LoRA. This is 200 epochs  

• 00:20:42 training results. As in the Qwen image training  tutorial, you should compare your checkpoints  

• 00:20:47 with grid system and verify which model,  which checkpoint is the best. Like 20 epochs,  

• 00:20:54 40 epochs, 60, 80. So this is 200 epochs. And  let's generate 4 random images. Currently this  

• 00:21:01 is using our ComfyUI installation. It  was updated. It is using Sage Attention,  

• 00:21:07 my first GPU. Everything was previously  explained in other tutorials. Moreover,  

• 00:21:12 you should follow your VRAM usage from nvitop  to see what is happening. I have the sufficient  

• 00:21:17 amount of RAM memory and VRAM. If your GPU  is weak like 6 GB GPU, don't worry. Since  

• 00:21:23 we are using the ComfyUI, it is automatically  doing block swapping and also CPU offloading,  

• 00:21:29 RAM offloading. Therefore, no matter what GPU  you have, it will work. This is the beauty of  

• 00:21:35 the SwarmUI with ComfyUI backend. So it is  doing all the optimizations that you need.

• 00:21:41 The image generation in the base resolution will  be pretty fast as you are seeing right now. Okay,  

• 00:21:46 this was the first image. The face is not that  great. We should do either face inpainting or  

• 00:21:51 we are going to do upscaling which I will show  in a moment. However, this took only 45 seconds.  

• 00:21:58 You can also reduce the number of steps if you  want. But this model is not a small model. This  

• 00:22:04 is a big model. 14 billion parameters. This  model is much more powerful than the Z Image  

• 00:22:11 turbo model if you ask my opinion. Okay, this is  another image. So generate few images like this.  

• 00:22:16 Okay. Let's say I liked this image. So I will  click reuse parameters. It will set its seed.  

• 00:22:22 Then I will go to presets and I will apply the our  upscale. You see there is 2x upscale direct apply.  

• 00:22:30 And I will regenerate. Let's see the difference.  Furthermore, you can change the base resolution  

• 00:22:36 of this model. Up to 1280 to 1280. 1280 to 1280.  This model is able to generate images. Sometimes  

• 00:22:46 it may add some hallucinations to the right and  left border. But you can increase the resolution,  

• 00:22:52 the base resolution to 1280 to 1280. I will  show after this how to do that. But let's  

• 00:22:59 see the upscaled result. So as a reminding, let  me remind you where you put your trained LoRAs.  

• 00:23:04 They are put inside SwarmUI inside models inside  LoRA folder. This is where you need to put your  

• 00:23:10 trained LoRAs. Inside LoRA folder. Okay, now  it is upscaling. The upscaling will take more  

• 00:23:15 time obviously because we are doubling the image  resolution. So as I said, don't worry if you don't  

• 00:23:22 have a powerful GPU, it will still work with  automatic block swapping and CPU offloading.  

• 00:23:28 Everything will be handled by the ComfyUI.  No matter what GPU you have, it will work.

• 00:23:33 And here it is. So let's make a comparison.  This was the base image. You see this was the  

• 00:23:38 base image. Then this is the upscaled image.  As you can see, it fixed the face. It added  

• 00:23:44 huge amount of details, quality. There is a  significant difference between the base image  

• 00:23:50 and the generated upscaled image. Again, as I  said, you can increase the base resolution even  

• 00:23:58 up to actually 1920 to 1080. However, it may add  some hallucinations. Let's try this and to see  

• 00:24:06 the difference. But in my experimentation what I  did was I selected my model which is low noise.  

• 00:24:14 So here low noise. So click this hamburger menu,  edit metadata, and change this to 1280 1280. Okay,  

• 00:24:23 like this. Save. Then select another model and  select your model. It will update resolutions  

• 00:24:29 accordingly. So when you select your aspect ratio  now, it will set it accordingly to 1280 to 1280  

• 00:24:37 base resolution. These are all optional. You can  do whichever you want. And generate another image  

• 00:24:43 with 1280 to 1280 total resolution. The aspect  ratio is like this. But let's see also 1920 to  

• 00:24:51 1080 result as well. Okay. Yeah, not bad.  Not as good as upscale but not bad at all.  

• 00:24:58 Now with this 1280 to 1280 base resolution, if I  also do upscale which I did in the example images,  

• 00:25:07 it will become much better. For example, let  me show you from the history row and Wan 2.2.  

• 00:25:15 I have shown in the beginning of the tutorial but  let's see again. So you see these images are huge,  

• 00:25:21 massive. If I open, this is 3360 to 1920 pixels.  Therefore, it has huge amount of details. I mean  

• 00:25:34 look at these details. So you can generate  massive images as well with this strategy.  

• 00:25:40 Set the base resolution and upscale. Okay, this  is 1280 to 1280. Now I will reuse parameters  

• 00:25:46 and enable my upscale preset one more time. It  is here direct apply and generate. Now it will  

• 00:25:53 upscale this image into twice resolution.  Of course, it will take time. However,  

• 00:25:58 it will be the maximum quality. So this image  full HD was generated in 73 seconds. This image  

• 00:26:07 1280 to 1280 was generated in 55 seconds. This  big image was generated in 2 minute 25 seconds.  

• 00:26:16 I mean these are the expected times because  we are generating really high quality images,  

• 00:26:21 really high resolution images with perfect  accuracy. And this is literally the first  

• 00:26:26 generation that we did. So I can generate  more and pick the better ones and as you  

• 00:26:30 do more generations you will understand  how the model works, what are the basics.

• 00:26:35 Okay, so the upscale of the 1280 to 1280  completed and let's look at the difference  

• 00:26:41 one more time. So this is upscaled and  this was the base. So from this base,  

• 00:26:47 we upscaled into this masterpiece. The  resolution is just amazing. The details,  

• 00:26:53 the realism, everything is mind blowing compared  to the other models. So can you use this model,  

• 00:27:01 this only images trained low noise model  for video generation? Yes. It will be  

• 00:27:07 a little bit lower quality, but yes. Let me  demonstrate that. So I will copy this prompt.  

• 00:27:13 You can also use this for image to video as well,  don't worry. Everything is same. So let's refresh.  

• 00:27:19 Reset params to default. And there are several  options that you can use. For the highest quality,  

• 00:27:26 you need to use Wan 2.2 high quality text to video  20 steps. This will take some time. On my GPU it  

• 00:27:34 takes like 6 to 7 minutes. Or you can use the  lightning LoRA based one. You see Wan 2.2 text to  

• 00:27:41 video. So let's see the lightning LoRA first, then  let's see the other one. Then let's proceed to  

• 00:27:47 dual model training. So direct apply. You see it  is going to set the low and high noise LoRAs like  

• 00:27:54 this automatically for you. Copy paste your prompt  and select your trained LoRA. So this was the low  

• 00:28:00 noise LoRA. Then generate. This will be pretty  fast because it is only total 4 steps. And again,  

• 00:28:09 for this preset to work flawlessly, you need to  download this bundle. Or you will be having issues  

• 00:28:16 with the models finding them, setting them.  I am doing everything automatically for you.

• 00:28:21 Okay, so the generation started. First it is  starting to use the high noise Wan 2.2 model.  

• 00:28:28 Then it will use the low noise model. However,  we should change the prompt to into a video  

• 00:28:35 of. So it will be a little bit better. When you  switch presets like this, since your base model  

• 00:28:41 will change, it will take some time for loading  reloading models, but everything is automatically  

• 00:28:46 and properly made by the ComfyUI. So you shouldn't  have any problems. And as you can see, it is able  

• 00:28:54 to 100 percentage utilize my GPU. It is using  my GPU 575 watts at once. So it is maximum power  

• 00:29:05 usage. Okay, it is starting to generate. Since  we trained our LoRA only on low noise model, this  

• 00:29:12 will work a little bit lower quality than training  on dual models which I will show. However,  

• 00:29:18 you can use this LoRA on image to video presets  as well. I will show that too after this. Okay,  

• 00:29:24 so the video has been generated. Let's see  the video. Yes, I can see the resemblance  

• 00:29:30 and accuracy. It is pretty good as you can  see. This is base video. If I upscale it,  

• 00:29:35 it would be better, but I can certainly see the  resemblance especially as it gets closer. Let me  

• 00:29:42 show you the final frame perhaps like this. And  it only took how many minutes. It only took 111  

• 00:29:49 seconds. So it takes lesser than 2 minutes. If I  want better quality, then I need to switch to this  

• 00:29:56 preset. This will do real 20 steps with real  CFG scale and it will be much higher quality.

• 00:30:04 What about using this on image to video. Can I do  that? Yes. So I will do reset params to default.  

• 00:30:12 Go to presets. With image to video again we have 2  quality presets. You can use this high quality or  

• 00:30:20 you can use Wan 2.2 image to video 4 steps. This  is also really good quality. So direct apply. Then  

• 00:30:27 let's select our prompt like this. I don't need  to copy this of course this time. Just copy paste  

• 00:30:33 the prompt. Then as an image for example let's use  this image. So I will go to init image choose file  

• 00:30:41 then choose this then resolution close use closest  aspect ratio. And the preset was not selected.  

• 00:30:48 Yes, I need to select the preset so refresh. Reset  params to default. Then refresh models to be sure  

• 00:30:55 you have everything. Then presets and let's go  to presets. Okay this one direct apply. Yes,  

• 00:31:02 you see it has selected the accurate model. This  is important apply. Then let's go to init image  

• 00:31:08 choose file select our photo closest aspect ratio.  Okay. Now I also need to select my LoRA. So my  

• 00:31:15 LoRA was let's see this one. This is only trained  with the images on low noise and generate. This is  

• 00:31:24 also 4 steps. So therefore it will be pretty good.  However, since we are switching a lot of models,  

• 00:31:30 it may leave some left over. Therefore what  I am going to do is I will cancel this first.  

• 00:31:35 Then server backend and restart all backends.  This will refresh my memory usages like this.  

• 00:31:43 You see it cleared everything. If you are going to  do a lot of model switching, you can restart your  

• 00:31:48 backend to clear your VRAM and RAM memory and then  generate. Now it should work better. Otherwise you  

• 00:31:56 need to trust the ComfyUI RAM management. So I  recommend to restart backends when you are going  

• 00:32:01 to do big models switching like this. Because  we switched from LoRA text to video to image  

• 00:32:08 to video. Okay, it is nice. It is loading into RAM  then starting generation. This should be also take  

• 00:32:15 lesser than 2 minutes especially the second and  third generations. By the way this was an image  

• 00:32:21 prompt. Therefore for video prompting you need  to do better. But I am just showing an example.  

• 00:32:28 So this prompt is not made for video generation.  This was made for image generation. Therefore it  

• 00:32:35 may not be perfect for video generation. However  still we should see the you know the quality and  

• 00:32:42 video. I have tested that training text to  video model LoRA improves image to video as  

• 00:32:49 well. Training image to video LoRA doesn't improve  that well. It is weird I know. Maybe it is because  

• 00:32:55 I have trained with only images static images  not with videos. But this is really improving  

• 00:33:00 image to video as well. So you can train an image  LoRA from text to video model with just images and  

• 00:33:07 it will improve your image to video generations  as well. Okay, it failed for some reason maybe.  

• 00:33:14 Let's see the logs. No I don't see any error.  Why it is showing an error. Okay, I think I  

• 00:33:21 was out of RAM memory so the ComfyUI crashed  probably because I am already using 2 GPUs.

• 00:33:29 RAM memory on my computer with all other stuff  that is open. Okay, it crashed again due to out of  

• 00:33:37 RAM memory. Therefore, I have added this argument:  --cache-none. When you add this into your backend,  

• 00:33:45 it will cache none of the models. So this is  really useful when you are working with Wan 2.2  

• 00:33:52 low and high noise models together at the same  time if you don't have sufficient amount of RAM  

• 00:33:58 memory, especially system RAM memory. So this  way, it will use a model, generate, it will  

• 00:34:04 then offload it, completely deleted, and load the  next model. This will minimize your RAM usage,  

• 00:34:11 also VRAM usage, with delay of loading models  from the disk. But since my disks are fast,  

• 00:34:18 it is not an issue. So it fixed the  issue. There is no more RAM leakage  

• 00:34:23 or RAM accumulation and we are about to  get the video ready. It is pretty fast.

• 00:34:30 Okay, we got the video. So this was from image  to video. As we get, you know, distant from the  

• 00:34:37 camera, the resemblance gets decreased, but  this was only static images trained model,  

• 00:34:44 not the auto model. However, it is decent  and it took only 2.4 minutes generation. So,  

• 00:34:52 how we are going to train on dual models?  Not only with the single model. If you want  

• 00:34:58 to generate videos, this is what you should do.  So I will just disable my backend to minimize my  

• 00:35:04 VRAM usage. Then go back to your SECourses  Musubi Tuner. So for dual model training,  

• 00:35:13 we are just going to change our preset. So I will  open a new tab here. Go to 1 Models Training,  

• 00:35:20 open folder. Go back to Wan 2.2 training  configurations. So we are going to use  

• 00:35:24 text to video configs. These configurations train  both high noise and low noise models at the same  

• 00:35:32 time. However, this will use more VRAM and this  will use more RAM memory since you need to have  

• 00:35:38 both of the models loaded into your RAM memory.  Therefore, be careful. I recommend to begin with  

• 00:35:45 fp8 scaled versions. Then once you verified it  is working, you can move to non-fp8. So let's  

• 00:35:52 begin with fp8 scaled. Again, always save your  configurations like this. Then we need to set  

• 00:36:00 the output directory. Don't forget that. So  I can just copy paste the directories here.

• 00:36:07 Training dataset. Again, I am only using static  images. If you use videos, the videos will use  

• 00:36:13 more VRAM memory. Therefore, be very careful.  If you use videos, then you need to have more  

• 00:36:20 VRAM memory, RAM memory. You need to do more  block swapping. You may need to reduce your  

• 00:36:25 training resolution. Everything changes. So first,  train with only static images, verify results,  

• 00:36:33 get some results, get some test done, then you can  include videos into your training dataset. You can  

• 00:36:39 have multiple subfolders for different training  datasets and for each dataset have different  

• 00:36:47 resolutions. So let me demonstrate an example.  So what I did is I generated another folder like  

• 00:36:54 this, 1_ohwx_video. However, I will be needed  to set different captions, not automatic ones.  

• 00:37:00 So I will copy paste it here. As I said, first I  recommend to train with only static images then  

• 00:37:06 include videos. But I will show how to include  videos. And generate dataset configuration. Now,  

• 00:37:13 it will generate datasets like this. You need to  edit this TOML file. Why? Because you will see  

• 00:37:22 that it is using the same resolution for both  images and videos. What does this mean? This  

• 00:37:28 means that if your images are high resolution,  the videos will be too high resolution. Either  

• 00:37:32 you need to auto downscale your videos to lower  resolution or you need to edit this TOML file.  

• 00:37:40 So I will open this TOML file. It is saved  inside this folder. You need to do this manually.

• 00:37:47 So this is the generated TOML file. You can see  that there are 2 different datasets. The first  

• 00:37:53 one is the image dataset and the second one is  the video dataset. And in the video dataset,  

• 00:37:58 it is going to use some frame extraction method  and it is going to use some methods like frame  

• 00:38:03 extraction head, target frames, number of repeats,  max frames. So I am going to set resolution here  

• 00:38:09 as well and change the resolution like 480 to  480. So that it will use this resolution for  

• 00:38:17 this dataset. You can also change the resolution  of the image dataset from here. So you can have  

• 00:38:23 different dataset resolutions. And if you are  wondering what are these other options for  

• 00:38:28 video datasets, it is all written here with  details. When you scroll down, you will see  

• 00:38:33 dataset preparation details. Open this and read  here. However, as I said multiple times now,  

• 00:38:41 train with only images then verify your results  then do training with video dataset. Moreover,  

• 00:38:49 now since we have 2 different folders, you will  see that it generated the captions like this  

• 00:38:54 for the video dataset. So you need to change the  caption of the video as well. I didn't test this.  

• 00:39:01 Probably this will work best with just trigger  word or you can describe the video action as  

• 00:39:07 well in these captions. However, for using the  static images, I recommend only trigger word  

• 00:39:13 as a caption. So I will delete this and I will  regenerate dataset configuration. Yes. Now it will  

• 00:39:20 use only my images dataset. It shows 56 double  times but it is okay. There is actually 28 images.

• 00:39:27 Okay, we did set the dataset. This is  how you prepare your video training  

• 00:39:32 datasets. Then we set the other things same  as the previous training. So in the models,  

• 00:39:39 I will just copy paste the paths here, here,  here. And additionally, for this model,  

• 00:39:46 you need to set the high noise model as well. So  it is here. So I will set it like this. You see,  

• 00:39:53 high noise. And that's it. Now I will save my  configuration and click training. It should  

• 00:40:01 start and work right away. However, you need  to have more RAM memory for this to work and it  

• 00:40:08 will use a little bit more VRAM memory as well.  So let's see what happens. The training window  

• 00:40:15 is here. You will see that it will load both low  noise and high noise models into the your system.  

• 00:40:24 You should see them in the CMD window. So first  it is loading the low noise then it is loading  

• 00:40:30 the high noise. Low noise is for details,  high noise is for the initial generation,  

• 00:40:35 the base structure. We train only low noise for  image generation. However, if you want to generate  

• 00:40:42 videos, you should train on both of them. You can  also train them individually one by one. However,  

• 00:40:48 its quality is lower than training this dual  model configuration. So you can load the image  

• 00:40:54 configuration, image training configuration,  change low noise model to high noise model. It  

• 00:40:59 will work exactly same and train high noise  model. If your system is not able to train  

• 00:41:05 both models at the same time, you can follow  that strategy. But what I recommend is rent  

• 00:41:11 a machine from the RunPod or Massed Compute  which I will show and train there and use.

• 00:41:16 Okay, it started. As you can see, both of the  models is running at the same time now. You will  

• 00:41:23 see that it will fluctuate the VRAM usage and how  it changes model is like 87 percent. Okay, you see  

• 00:41:31 the RAM usage increased because now it is loading  the other model and it is switching between them.  

• 00:41:37 Okay, you see. It is working nice. It is using  almost my entire RAM memory right now so I should  

• 00:41:43 decrease my RAM memory usage. I can also reduce my  block swap count because I still have some VRAM.  

• 00:41:50 So where you increase or decrease the block swap?  So let's open all panels, search for swap and  

• 00:41:57 you will see that there is block swap. So you can  change block swaps. Currently since it is fitting  

• 00:42:04 into my VRAM entirely with fp8 scaled, I am not  doing any block swap. But if it doesn't fit in  

• 00:42:10 your case, you need to increase this like 10, 20,  15. Moreover, I can use offload inactive DiT to  

• 00:42:18 CPU. This requires more RAM memory and it is not  compatible with blocks to swap. So when you do  

• 00:42:24 block swap, you cannot use this unfortunately.  Therefore, it is the fastest training that I  

• 00:42:30 can get right now. Its speed is very good as  you can see. It is also faster than training  

• 00:42:34 only image model right now because I am doing fp8  scaled training. If I were doing not fp8 scaling,  

• 00:42:41 I would be needed to do block swapping and block  swapping reduces your speed. So this dual model  

• 00:42:47 training takes like 6 hours on my GPU, maybe  even faster. This is how you train dual model.

• 00:42:54 And how you use dual model? Exactly same as the  single image training. I only change my LoRA to  

• 00:43:01 use it. So let's enable back our backend and let's  open a new tab. Copy our prompt. Quick tools,  

• 00:43:10 reset params to default presets. And let's use  this high quality preset this time. Direct apply.  

• 00:43:16 Type our prompt like this. Okay. Make sure to  not include the LoRAs in your prompt when you  

• 00:43:22 copy paste. Okay, it is like this. And I need to  select my LoRA. So my LoRA is this one. High noise  

• 00:43:30 low noise auto. So this was trained exactly as I  have just shown you. This is the highest quality  

• 00:43:36 and let's generate. So this generation will  take significantly more time compared to the  

• 00:43:43 fast generation which is Wan 2.2 text to video 8  steps. Actually this is 4 steps. I forgotten to  

• 00:43:50 change its name. Yes, this is 4 steps. So it will  take more time but this is the highest quality  

• 00:43:57 that you are going to get with video generation,  with text to video generation. For image to video,  

• 00:44:02 it doesn't make that much big difference. You can  perfectly use this preset. But for text to video  

• 00:44:09 generation, it makes really big difference  still because none of the recent LoRAs are,  

• 00:44:15 the lightning LoRAs are still as good as the image  to video LoRAs. I am using the latest LoRA which  

• 00:44:22 was published only 3 days ago. You see they  are 17 December 2025 LoRAs but it is as I say.

• 00:44:30 So the generation started. As you can see, it  should be done in like 6 minutes if I remember  

• 00:44:37 correctly. We will see. Okay, so we got our video  generated. It is not the best video. The quality  

• 00:44:43 is good but the prompt is bad. So therefore I  need to edit my prompt and generate new one.  

• 00:44:50 The prompting do matters a lot. So play with your  prompts to get the best results. Remember all the  

• 00:44:58 generations will be inside your SwarmUI, inside  output, inside local, inside raw folder. So you  

• 00:45:06 will see all of your generations here. They will  be saved by the date, latest date folder. You will  

• 00:45:13 see and find your images and videos in this  folder if you need them later. Furthermore,  

• 00:45:19 you will see them in the history tab as well  from the raw and with the folders like this.

• 00:45:25 So as a next step, I will show how to train on  Massed Compute then I will show on RunPod. So  

• 00:45:32 for Massed Compute, enter inside the extracted zip  file folder and open Massed Compute instructions  

• 00:45:38 read txt. Reading this file is mandatory  and watching the Windows tutorial part are  

• 00:45:43 mandatory. Please use this link to register. I  appreciate that very much. Then go to billing  

• 00:45:50 and set up some credits. They accept crypto  payments as well. Then go to deploy. In here,  

• 00:45:57 I recommend you to use RTX Pro 6000 Blackwell GPU.  This is working really great. From the category,  

• 00:46:04 select creator, select SECourses. Then enter  your coupon. This is important. We are going  

• 00:46:10 to use Category Creator, Image SECourses  and discount coupon SECourses. You see the,  

• 00:46:16 this very powerful GPU is 1.8 dollars  per hour. When I apply my coupon,  

• 00:46:23 it becomes 1.35 dollars. If you want  even more powerful GPU, faster training,  

• 00:46:29 select H200 NVL. This is even more powerful than  RTX Pro 6000. It is 2.6 dollars. With our coupon,  

• 00:46:38 it becomes 1.95 dollars. This is the fastest GPU.  So you can use either of them. Let's use H200 for  

• 00:46:46 this demonstration, for this tutorial. Deploy.  Now we need to wait for machine to initialize.

• 00:46:53 To connect our machine, we are going to  use ThinLinc Client. Download it. If you  

• 00:46:57 followed other tutorials, you already know. After  downloading, just next, next, next. Then you need  

• 00:47:03 to set your shared folder from local devices.  You see clipboard synchronization and drives.  

• 00:47:09 And in here I have a shared folder with read  and write permission. So I can copy my small  

• 00:47:14 files and I can connect. For big files, you  need to watch this tutorial. This will show  

• 00:47:21 you how to upload and download big files to  Hugging Face. You can also use Google Drive  

• 00:47:26 or OneDrive. So other cloud services as well. You  know how to use them. But you cannot use ThinLinc  

• 00:47:33 Client to transfer big files. So let's just  wait for initialization to be complete. Okay,  

• 00:47:39 now it is running. You may refresh page time to  time to update but it was auto updated for me.  

• 00:47:45 Then click here and copy login URL. Open ThinLinc  Client. Copy username for first time users and  

• 00:47:52 copy password and connect. Then continue. Then  click start. It will start the machine and  

• 00:48:00 connect from the ThinLinc Client. Remember this  is running on a remote machine, not on my machine.

• 00:48:05 Now I need to transfer my file, my installer file.  You can use Patreon login from this browser or you  

• 00:48:13 can use the shared folder which I prefer. So I  will copy my file into my shared folder. You can  

• 00:48:21 also put your training images here. Go back to  home from here. Go to Thin Drives. This is your  

• 00:48:26 shared folder. Do not run anything inside here.  Run everything from downloads folder. Whatever it  

• 00:48:33 is. It doesn't matter. The transfer speeds are not  great with ThinLinc Client. Therefore you should  

• 00:48:38 use the other method that I have just shown you,  the link. You see this link for big files. Okay,  

• 00:48:44 I will copy paste my images. Drag and drop or copy  paste both works. And our latest SECourses Musubi  

• 00:48:53 Tuner Premium Installer. So it is here. Drag and  drop into downloads folder. Wait for files to be  

• 00:48:59 transferred. You see it is being transferred right  now. Then go to downloads folder. Extract the  

• 00:49:05 Musubi Tuner. Extract here. And I will open the  Massed Compute instructions. Copy this command.  

• 00:49:12 While I am inside this folder, open in terminal.  This is important. And paste. You need to be  

• 00:49:18 in the accurate folder. Then it will start the  installation. You will see that installation is  

• 00:49:23 lightning fast. Actually let's see in real time.  So it is going to use uv package installation.

• 00:49:30 I mean look at the speed. Look at the  installation speed. It will take like  

• 00:49:33 30 seconds. Maybe 1 minute maximum. So it is  almost done. Also we need to download training  

• 00:49:40 models. So in here you see there is this command.  python3 download_train_models. Again while I am  

• 00:49:46 inside this folder, open in terminal. Paste. And  it will ask you which models to download. So let's  

• 00:49:53 download text to video option 4. The download will  begin. The downloader is also optimized. It will  

• 00:50:00 be pretty fast. You see 500 megabytes per second.  So it is 4 gigabits per second. 4 gigabits. So it  

• 00:50:08 will be done almost in no time. The installation  also completed and the trainer started. You see  

• 00:50:15 installation take like 60 seconds. The rest  is exactly same as in the Windows tutorial.  

• 00:50:22 You will open your configuration. And there  is one advantage because configuration is  

• 00:50:27 set for Massed Compute folders directly. So  I will download the Wan 2.2 training configs,  

• 00:50:34 text to video configs. Whether you want to  use dual GPU or whichever GPU. So this one  

• 00:50:41 is using 140 gigabytes of GPU. Let's see our  VRAM. nvidia-smi. Okay we have 140 gigabytes.  

• 00:50:49 Therefore I am going to use this config. I  hope it fits. You see it is maximum. Then  

• 00:50:54 I need to set my output folder and other thing.  I think output folder is already set. So I need  

• 00:51:00 to set my training images or videos. So  extract here. Okay. So copy this path and  

• 00:51:08 copy paste here. Then don't forget to click  generate dataset configuration. And let's see  

• 00:51:14 if the model download is completed. Yes almost  completed. You see 3 over 4. It is verifying.

• 00:51:22 So literally with like 5 minutes setup, you  can begin training in Massed Compute. Like 5  

• 00:51:29 minutes. Maybe 4 minutes. Once you do this like 1  or 2 times, you will get used to it. You will be  

• 00:51:35 immediately setup a machine and start training  right away. Okay high noise model is getting  

• 00:51:42 verified. I am also verifying files. Therefore  you will never have any issues with downloaded  

• 00:51:47 models. This is super important because sometimes  models get corrupted and it is causing a lot of  

• 00:51:52 issues. Okay it is downloading the last model.  It is 50 seconds remaining time. Okay so all  

• 00:51:58 the downloads have been completed. Now I did  set the training dataset and start training.  

• 00:52:05 It should work if the folder paths are accurate  if I remember but probably we need to change one  

• 00:52:11 more thing. Not ready yet. So in the 1 Models  Settings, yes. You see this is 23 version so I  

• 00:52:18 need to make this 24. You also need to change  that. The output folder should be okay. Yes  

• 00:52:26 it is okay. And start training. Now let's see.  Oh by the way if this was a 2 model training,  

• 00:52:33 yes it was. I need to also change this. So  dual model. Okay. Now it will give another  

• 00:52:38 error and I will click training again. So first  it is caching the training dataset. Okay it  

• 00:52:45 is loading. Okay I will stop and click start  again. Yes now it should work perfectly fine.  

• 00:52:51 Let's follow both of that. So let's zoom in.  Another zoom. Okay. Let's see in real time. So  

• 00:52:59 starting training. Loading models. You see loading  models are amazingly fast. Blazing fast on Massed  

• 00:53:06 Compute. This is why I like Massed Compute over  RunPod. The disk speed on Massed Compute is like  

• 00:53:12 10x, maybe 20x faster than the RunPod. Okay 77  gigabytes. Okay now it is training the low noise  

• 00:53:20 then it will begin the high noise. Okay training  is starting. 4 steps done on the first model then  

• 00:53:27 it will load the second model. The logic of dual  training is that it switches models between. Okay  

• 00:53:33 why it is taking time? I think the first step is  taking time. Okay. Now it did begin. Yes. So the  

• 00:53:43 training started. We will see the training speed  gets better. Okay it is using 80 gigabytes right  

• 00:53:50 now. But the training speed is getting better.  You see 2.2 second. It will get faster than 1 IT  

• 00:53:57 per second. But it is still switching models.  So what we can do, we can disable this and  

• 00:54:04 see if it works. So let's disable this to get  even faster. Did we load the inaccurate model?  

• 00:54:10 Yeah this should be disabled. I will update  this configuration for you so you won't have  

• 00:54:15 this issue. Now it should be even faster than  before because it will not offload each model.  

• 00:54:21 It will load both of the models into the GPU.  Let's see the speed. Of course you can enable  

• 00:54:27 fp8 scaled. It will take even lesser VRAM so you  can load it into like 80 gigabytes GPUs as well.

• 00:54:35 Okay training started. You see it is 1.5  IT per second using only 106 gigabytes of  

• 00:54:45 GPU. I will update the configuration so  you won't have this issue. But this is  

• 00:54:49 the training speed. It is going to  take only 67 minutes. So with like  

• 00:54:54 2 dollars, you will be able to train very best  Wan 2.2 model. And it's a dual model training,  

• 00:55:02 not single model. You are seeing the speed right  now. 1.35 IT per second. It is amazing. And it  

• 00:55:09 will take only 68 minutes and we are doing 5600  steps. Not a low number of steps. So it is working  

• 00:55:17 just amazing. Yes I have updated the config as  140 gigabytes and it is using 107 gigabytes GPU.

• 00:55:25 So now I will show how to train on RunPod. The  rest of the Massed Compute is same. You just  

• 00:55:32 need to download your LoRA files and use  on your computer or you can use in Massed  

• 00:55:37 Compute as well. If you wonder how to use  on Massed Compute, on our channel SECourses,  

• 00:55:44 we have a video for this. The video is this one.  Generate AI art 10x faster. The ultimate SwarmUI  

• 00:55:50 and ComfyUI cloud tutorial. You can watch this  tutorial to learn how to use SwarmUI on Massed  

• 00:55:57 Compute or on RunPod. This tutorial covers both  of them. So for training on RunPod, we have RunPod  

• 00:56:05 instructions. Open that file. Always you need  to follow this file for RunPod. Please register  

• 00:56:12 RunPod from this link. After registering and  logging in, go to billing and set up some credits.  

• 00:56:20 Then once you have the credits, go to pods.  If you want to use permanent storage, I have  

• 00:56:26 a tutorial for that. It is here. If you want  to learn how to upload and download big files,  

• 00:56:31 the tutorial is here. And if you want to learn  RunPod usage, there is another tutorial here.

• 00:56:36 So I will show on RunPod with RTX 5090. One of  the most commonly used one. But you can use bigger  

• 00:56:45 GPUs like B200 for even faster training. So this  is the disk type and this is the RAM GPU filters  

• 00:56:51 that I use. Let's select RTX 5090. Now a lot of  people are getting confused here. You are not  

• 00:56:58 going to use this template. You are always going  to use the template that I write in this file.  

• 00:57:05 Always. Because my installers are optimized for  whatever it is written here. Otherwise you will  

• 00:57:10 get errors. So change template. Select PyTorch  2.2.0. This is official template. Click edit  

• 00:57:18 and increase the disk size like 200. It is up to  you. If you want to save more checkpoints, you  

• 00:57:24 need to increase this. Set overrides and deploy on  demand. Now we need to wait for machine to become  

• 00:57:32 ready but it should be very fast because this  is official template. This is very lightweight  

• 00:57:37 template. So let's see. Time to time click here.  Okay it is ready. Jupyter Lab. If this doesn't  

• 00:57:43 open, you need to refresh, click some several  times. If still doesn't work, get a new machine.

• 00:57:49 Then I will upload my downloaded zip file. Wait  for upload to complete. You see in the bottom it  

• 00:57:54 shows. Then right click and extract archive.  Click refresh. Then open RunPod instructions  

• 00:58:00 read txt file. Copy this command. Open a new  terminal from here. And terminal. Make sure  

• 00:58:07 that you are in the workspace and copy paste  like this. Then it will start the installation.  

• 00:58:12 So run the commands inside workspace folder.  Open new terminal. Make sure that you are in the  

• 00:58:18 workspace wherever you have extracted the files.  The installation will be extremely fast compared  

• 00:58:25 to before. Why? Because we are using uv package.  While installation is continuing, let's download  

• 00:58:31 the model because model download may be slow.  Open another terminal. Copy paste. Select the  

• 00:58:38 model text to video. It will start downloading.  This is also super optimized. Let's see. You  

• 00:58:44 see it is downloading with like 300 megabytes per  second. Pretty good. The uv installation is also  

• 00:58:50 ultra fast. Before it was taking sometimes 10  minutes, 20 minutes, 30 minutes to install. Now  

• 00:58:56 it will take only few minutes to install. This  is our latest optimization. I will move my all  

• 00:59:03 installers to uv installation hopefully slowly.  So you will always install my applications much  

• 00:59:10 faster now compared to before. Also models  are getting downloading meanwhile. So after  

• 00:59:15 installation it will auto start the application.  This warning is not important. You will get  

• 00:59:20 this in all installers. The model merging and  verification on RunPod is sadly slow compared  

• 00:59:27 to Massed Compute because its disk speed  is very slow compared to Massed Compute.

• 00:59:32 Okay application installation completed and  started. I will open the Gradio live link.  

• 00:59:38 However model downloading is still continuing.  Meanwhile let's also upload our training images.  

• 00:59:44 So for upload click this icon. Select your  training images as a zip file. Upload. I  

• 00:59:50 prefer this way of uploading. So we can set up the  configuration while models are getting downloaded.  

• 00:59:56 Wait for upload to be completed. You see it is  slow. You can also use runpodctl for uploading.  

• 01:00:02 I explain that in another tutorial video. For  example you can watch this one. But it is not  

• 01:00:08 mandatory to learn or you can use this tutorial  to upload and download big files very fast. This  

• 01:00:14 is the recommended tutorial. And interface  is here. This is running on RunPod. 1 Models  

• 01:00:20 Training. So from the configurations, let's  pick the configuration we want to use. Let's  

• 01:00:26 use the text to image configuration. So you see  we have 32 gigabytes. Copy path. Right click and  

• 01:00:32 copy path. Put a backslash. Always RunPod  puts a backslash to beginning of path and  

• 01:00:39 paste it. Then click this. It will auto load the  config. You see it is loaded successfully. Then  

• 01:00:46 set your output folder. Let's save them inside  workspace a new folder: Trained LoRAs. Right  

• 01:00:54 click and copy path. Then paste it here. Put a  backslash to the beginning. This is important.

• 01:01:00 If you want to do Qwen training, it is same. We  only load Qwen training into here. So this is how  

• 01:01:06 you train on RunPod. You can change your save  name or save every n epochs. And the training  

• 01:01:12 dataset. So I have uploaded my images. Right click  and extract archive. And they are named like this.  

• 01:01:19 Remember this is mandatory. 1_ohwx. ohwx will be  our captions. 1 is the repeating. So I am going to  

• 01:01:28 give the path of the parent folder. Copy as, right  click and copy as path. Backslash, put the path  

• 01:01:35 then click generate dataset configuration.  It is done. Then the 1 Models Settings,  

• 01:01:41 we need to set the folder paths. So the models are  downloaded inside Training Models 1. So copy path  

• 01:01:49 and change them like this. And backslash. Okay.  No not like this. It needs to be like this. Yes.  

• 01:01:57 So this is the accurate pathing on RunPod. So like  this. Yes. Verify the paths. And like this. Okay  

• 01:02:06 all paths are set. This is single model so there  is no high noise. And what else is left? Nothing  

• 01:02:12 else is left. Save your configuration always.  So you can load later if any error happens,  

• 01:02:18 if you restart. Okay the models are still being  downloaded. We need to wait them to be completed.  

• 01:02:24 If you use RunPod permanent network storage  system, you won't be needed to wait. However  

• 01:02:30 currently we are not using. But it is also  very slow. I mean unbearably slow. And you  

• 01:02:36 see only 260 megabyte per second disk speed. On  Massed Compute this is 2 gigabytes per second.

• 01:02:43 Okay all models downloaded. It is verifying the  last one. So we can start training. So make sure  

• 01:02:50 to save and click start training. Then we will  follow what is happening on CMD window. This  

• 01:02:58 is where we see the actual logs. If there is any  error, you need to report from here. Unfortunately  

• 01:03:04 setting up RunPod, starting training is like  10x slower than the Massed Compute. But it has  

• 01:03:11 some other advantages like permanent storage or  more GPU options. However this is the reality.  

• 01:03:18 You see I click start training and I am still  waiting for it to read from the disk to start  

• 01:03:24 caching, loading models. Okay let's also install  nvitop. pip install nvitop and just type nvitop.  

• 01:03:34 So we can monitor the GPU usage as well. If  you don't see GPU here, that means that your  

• 01:03:41 machine is broken. It is going to start caching  the text encoder outputs. It is started. Nice.  

• 01:03:47 Now it will load model. Oh one more thing. We can  increase the model loading speed. So I will stop  

• 01:03:54 training. How? Open all panels and search for  loading and disable numpy memory mapping. This  

• 01:04:02 will increase your loading speed significantly  on RunPod. So start training. Don't worry it  

• 01:04:08 will just resume from wherever it is left. So we  didn't lose any time. This is really speeding up  

• 01:04:14 the model loadings on RunPod. Okay now we need  to wait again. Okay it is loading the model.  

• 01:04:20 Pretty fast. This optimization is making huge  difference on RunPod. This has been implemented  

• 01:04:27 after I have taught this issue to Kohya. But  still nothing fast as the Massed Compute.

• 01:04:34 Okay training starting. You can monitor the GPU  from nvitop. Yes started. You see the first step  

• 01:04:41 speed was very slow. Now 1.46, 2 second IT, 2.15  second IT. This is slower than the other GPUs  

• 01:04:50 because we are also doing some block swapping  since this is not fp8 scaled. So when we go  

• 01:04:57 to swap, let's see how many we are doing. We are  doing 10 block swapping. If you want even faster  

• 01:05:03 training, you can do fp8 scaled training and  you can make the block swap zero and it will be  

• 01:05:09 faster. But this is okay speed. It is going to  take like 4 5 hours. So it is bearable maybe a  

• 01:05:16 little bit more. But this is how you do training  on RunPod. The models will be saved inside my  

• 01:05:23 Trained LoRAs folder. You see it is already  generated and the dataset TOML file is saved  

• 01:05:29 there. The training TOML file also saved there  when you click training. So you will download  

• 01:05:33 LoRAs from here and use them. Or if you want to  use them on RunPod, we have a tutorial for how  

• 01:05:39 to use SwarmUI and ComfyUI on RunPod. It is here.  This is a very recent tutorial. Fully up to date.  

• 01:05:45 So you can follow this tutorial to learn how  to use SwarmUI and ComfyUI on RunPod and Massed  

• 01:05:52 Compute. So this is it. I hope you have enjoyed.  Don't forget to stop your machine, terminate your  

• 01:05:59 machine. On Massed Compute, you need to delete  your machine. Stopping your machine will not  

• 01:06:04 stop your credit spending. Don't forget that. But  before deleting, make sure to backup all of your  

• 01:06:11 data. On RunPod, you can stop your machine and  it will keep your data then you need to terminate  

• 01:06:17 your machine to completely avoid credit usage. But  stopping machine will make it very minimal amount  

• 01:06:24 of credit usage. If you have any questions, always  ask me. We have Z Image Turbo LoRA training, very  

• 01:06:31 up to date, working amazing, very lightweight.  We have Flux SRPO training updated. We have Qwen  

• 01:06:40 training that you need to watch. Hopefully more  videos are coming, more applications are coming.  

• 01:06:45 I am working 7 days every week literally like 10  hours a day. So stay subscribed, leave a comment,  

• 01:06:53 ask questions, join our Discord, ask questions,  message me from Patreon. Hopefully see you later.