Ultimate ComfyUI and SwarmUI on RunPod Tutorial with Addition RTX 5000 Series GPUs and 1 Click to Setup

Ultimate ComfyUI & SwarmUI on RunPod Tutorial with Addition RTX 5000 Series GPUs & 1-Click to Setup

Full tutorial link > https://www.youtube.com/watch?v=R02kPf9Y3_w

If you want to use ComfyUI or SwarmUI with ComfyUI backend on RunPod cloud platform, this is the ultimate tutorial that you will find to step by step install ComfyUI and SwarmUI on RunPod and use each one of them. RunPod is a great platform to scale your AI generation or if you are a GPU poor, rent the very best GPUs and leverage the AI in your profession. ComfyUI is the ultimate ecosystem right now for Image and Video generation models and with SwarmUI interface leveraging ComfyUI, you can become master for gen AI. So learn how to install ComfyUI on RunPod step by step and run it. Then learn how to install SwarmUI on RunPod step by step and learn how to use it. Then learn how to give installed ComfyUI backend to SwarmUI and leverage its features and ultimate performance and optimizations. Moreover, the installers I made installs Torch 2.7, CUDA 12.8, xFormers, Sage Attention, Flash Attention, Accelerate, Triton, DeepSpeed, ComfyUI manager and moıre.

🔗ComfyUI Installer Zip File Download ⤵️

▶️ https://www.patreon.com/posts/Advanced-ComfyUI-1-Click-Installer-105023709

🔗SwarmUI Installer and Model Downloader Zip File Download ⤵️

▶️ https://www.patreon.com/posts/SwarmUI-Installer-AI-Videos-Downloader-114517862

▶️ Download & Upload Models Tutorial (wget) : https://youtu.be/X5WVZ0NMaTg

▶️ CausVid LoRA V2 Tutorial : https://youtu.be/1rAwZv0hEcU

▶️ CausVid Main Tutorial : https://youtu.be/fTzlQ0tjxj0

▶️ SwarmUI Master Tutorial : https://youtu.be/HKX8_F1Er_w

🔗 SECourses Official Discord 10500+ Members ⤵️

▶️ https://discord.com/servers/software-engineering-courses-secourses-772774097734074388

🔗 Stable Diffusion, FLUX, Generative AI Tutorials and Resources GitHub ⤵️

▶️ https://github.com/FurkanGozukara/Stable-Diffusion

🔗 SECourses Official Reddit - Stay Subscribed To Learn All The News and More ⤵️

▶️ https://www.reddit.com/r/SECourses/

Video Chapters

00:00:00 Introduction to ComfyUI & SwarmUI Installation on RunPod

00:00:22 Advanced Features of the One-Click Installer (Sage Attention, xFormers, Blackwell)

00:01:03 Demonstration on RTX Pro 6000 GPU & VRAM Optimization

00:01:52 Introducing the High-Speed Unified Model Downloader

00:02:06 Starting the Tutorial: Downloading & Preparing the Installer Files

00:02:45 Registering and Setting Up Your RunPod Account & Billing

00:03:04 Selecting the Optimal RunPod GPU (RTX Pro 6000) and Server

00:03:54 Critical Step: Configuring the RunPod Pod Template (PyTorch 2.2.0)

00:04:20 Setting Volume Disk, Container Disk, and Exposing HTTP Ports

00:05:09 Deploying the Pod and Accessing the Jupyter Lab Interface

00:05:46 Uploading and Extracting the ComfyUI Installer on RunPod

00:06:16 Running the ComfyUI Installation Command in the Terminal

00:07:22 How to Start ComfyUI with Custom Launch Arguments (e.g., --gpu-only)

00:08:13 Overview of Included Libraries (Manager, Flash Attention, DeepSpeed, Triton)

00:09:15 Using Sage Attention vs. xFormers for Optimal Performance

00:09:42 Connecting to the Live ComfyUI Web Interface

00:10:27 Introducing the Unified Model Downloader for ComfyUI & SwarmUI

00:11:01 Running the Model Downloader and Configuring the Download Path

00:11:43 Downloading FLUX Models for a ComfyUI Image Generation Test

00:13:22 Configuring the ComfyUI Workflow with Newly Downloaded FLUX Models

00:15:01 First Image Generation Test: Analyzing the Incredible Speed

00:15:53 Terminating ComfyUI and Preparing for SwarmUI Installation

00:16:25 Installing SwarmUI and Setting it Up via the Cloudflared Link

00:16:50 How to Configure SwarmUI to Use ComfyUI as a Self-Starting Backend

00:17:50 Downloading Wan 2.1 Video Generation Models with the Downloader

00:19:22 Importing Presets and Setting up a Wan 2.1 Video Generation in SwarmUI

00:20:21 First Video Generation Test with Wan 2.1 (GGUF Version)

00:21:26 Analyzing GGUF Performance and Deciding to Test the FP16 Model

00:22:54 Downloading the FP16 Wan 2.1 Model for a Performance Comparison

00:24:01 Installing Extra Features like RIFE Frame Interpolation

00:24:41 Re-configuring SwarmUI to Use the FP16 Model with 16-bit Precision

00:25:52 Analyzing FP16 Performance and VRAM Usage (57GB)

00:26:28 Final Experiment: Using the Official Wan 2.1 Preset for Best Results

00:27:28 Final Results, Conclusion, and How to Get Further Help

ComfyUI: A Powerful Tool for Creative Workflows

ComfyUI is an open-source, node-based interface for Stable Diffusion, enabling users to create complex generative art and image processing workflows. Its modular design allows for intuitive drag-and-drop connections, making it accessible for artists and developers. With extensive customization, plugin support, and a vibrant community, ComfyUI empowers users to experiment with AI-driven creativity efficiently.

Some background music by NoCopyrightSounds : https://gist.github.com/FurkanGozukara/681667e5d7051b073f2e795794c46170

Video Transcription

• 00:00:00 Greetings everyone. Today, I am going to show you  how to install the latest version of ComfyUI on  

• 00:00:06 RunPod. I have been getting asked how to install  ComfyUI and SwarmUI on RunPod properly. I have  

• 00:00:15 prepared one-click installers for ComfyUI and  SwarmUI, and the difference of this installation  

• 00:00:22 is that it supports Sage Attention, xFormers,  Flash Attention, Accelerate, DeepSpeed, Triton,  

• 00:00:30 whatever the latest inference library there are,  it supports all of them. Moreover, I will show how  

• 00:00:37 to install SwarmUI on RunPod and use the installed  ComfyUI as a backend with extra parameters to  

• 00:00:47 optimize it to the maximum level. You see, like  GPU only, Sage Attention parameter. My installer's  

• 00:00:55 advantage is that it supports Blackwell GPUs  as well. So I will show everything on RTX Pro  

• 00:01:03 6000 GPU, which is the 96 gigabyte version of  the RTX 5090. It is a little bit faster and it  

• 00:01:13 has a huge amount of VRAM. So I will show you how  to leverage this 96 gigabytes of VRAM as well,  

• 00:01:21 like keeping the entire model in the GPU and  have the maximum performance. But don't worry,  

• 00:01:28 you can also use my installer and you can follow  this tutorial to use everything on different GPUs  

• 00:01:34 like B200, H100, RTX A6000 ADA, L40, L40S, RTX  4090, RTX 5090. My installer is basically working  

• 00:01:46 on everything. I'm also using the official  PyTorch template, so the initialization of  

• 00:01:52 the machine is super fast. Moreover, I will be  introducing you to my model downloader so that  

• 00:01:58 you will be able to download models extremely fast  and easy on RunPod. So let's begin the tutorial.

• 00:02:06 So as usual, I have prepared a really great post  where you will find all of the information and  

• 00:02:12 installers. Let's begin with installing ComfyUI.  Then I will install SwarmUI and I will show you  

• 00:02:19 both ComfyUI and SwarmUI installation and how  to use on RunPod. So I have downloaded my zip  

• 00:02:26 file. What I need to do is just, I need  to move it into a drive. I can use also  

• 00:02:31 downloads folder and extract files into there as  a beginning. Enter inside the extracted folder and  

• 00:02:39 double-click RunPod instructions read.txt file.  This will give you the instructions to follow.  

• 00:02:45 Please use this link to register to RunPod. I  appreciate that. Then click sign up, register,  

• 00:02:52 put some credits. To putting some credits, go to  billing and set your account, put some credits.  

• 00:02:57 Then click "Pods" section here. This is the latest  interface of the RunPod. Today I am going to show  

• 00:03:04 you on RTX Pro 6000 GPU, but you can use any  GPU. You can also set some additional filters  

• 00:03:12 from here, like the location or the disk type.  Let's go with the NVMe disk. This is faster. You  

• 00:03:20 can also filter by RAM. Moreover, you can also  use community cloud, but these are way slower,  

• 00:03:25 way, way slower for installation. But they are  also cheaper, but for the demonstration purposes,  

• 00:03:32 I will use the RTX Pro 6000. So what's up with  this GPU? This GPU has 96 gigabytes of VRAM. It is  

• 00:03:40 also a modern GPU. It is like RTX 5090 as a speed.  The price is decent, not great as the Massed  

• 00:03:49 Compute prices, but it is decent. Then as a next  step, click "Change Template". This is important.  

• 00:03:54 You should always look at the templates that I  write here. You need to use them. For ComfyUI,  

• 00:04:01 we are going to use this template. Don't worry,  we are going to use latest libraries with this  

• 00:04:06 template, so it will not matter. Okay, click  "Change Template". You need to type here like  

• 00:04:11 PyTorch, then select the PyTorch 2.2.0. This  is the official template, therefore it will be  

• 00:04:17 really fast. And then click the "Edit Template"  and set the volume disk. Volume disk is the place  

• 00:04:23 where all the models will be downloaded. So set  it according to your needs, which models you are  

• 00:04:29 going to download. I will go with 200 gigabytes.  Moreover, you can set increased container disk.  

• 00:04:36 This is the temporary disk where the libraries  will be installed, but 20 gigabyte is sufficient.  

• 00:04:41 Now we are also going to expose some of the  HTTP ports to connect ComfyUI from the proxy.  

• 00:04:47 So export 3000, 3001 and 3002. For SwarmUI, we  will use the Cloudflared, but you can still use  

• 00:04:54 these ports if you wish, like 7860, 7861. You can  expose as many as you want, then set overrides. So  

• 00:05:03 we are done with setting the template and editing  the template. Just click "Deploy on Demand". Then  

• 00:05:09 click "My Pods". You can also click here. Wait  for your pod to be initialized. You see, this is  

• 00:05:15 a really, really good pod because these machines  are newly built because of this GPU. This GPU is  

• 00:05:21 very new and you see the pod has been initialized.  The initialization of the pods are really fast on  

• 00:05:28 RunPod, but the installation of libraries are  usually slow. Wait until you see Jupyter Lab  

• 00:05:34 become green. If it doesn't become green, still  just click and see if it is working because  

• 00:05:39 sometimes it may not. And our Jupyter interface  has been opened. Then move the downloaded zip file  

• 00:05:46 into here like this. You can also use this icon,  "Upload files" to upload it. Wait for upload to  

• 00:05:52 be completed. You will see in the bottom here.  It is uploaded. Right-click, "Extract Archive".  

• 00:05:58 So first we are going to install the ComfyUI  with latest libraries including Flash Attention,  

• 00:06:03 Sage Attention, xFormers, everything, and it  supports all of the GPUs. Then click the RunPod  

• 00:06:09 instructions.txt file, double times and copy this  installation command. Then click this plus icon,  

• 00:06:16 "Terminal", paste it. If it doesn't paste, you  can right-click and paste. It will ask you to  

• 00:06:20 allow and you can allow it, but with Ctrl+C,  Ctrl+V, it is working on Windows. And now you  

• 00:06:26 need to wait for installation to be completed  patiently. The installation is the slowest part  

• 00:06:31 on RunPod because of their disk speeds, because  of the disks they use. So we just need to wait.

• 00:06:37 Finally, the installation has been completed. It  took more than 25 minutes. This is why I recommend  

• 00:06:45 Massed Compute instead of RunPod because it  would take maximum few minutes on Massed Compute.  

• 00:06:50 Moreover, if you don't want to install every time  like this, I really recommend you to watch RunPod  

• 00:06:58 permanent network storage tutorial to learn  how to use network storage system of RunPod  

• 00:07:05 so that you can continue using your previously  installed machine. So since the installation has  

• 00:07:11 been completed, you can verify the install logs  with scrolling down, probably not necessary,  

• 00:07:16 but let's start our ComfyUI installation. For  starting ComfyUI installation, you need to copy  

• 00:07:22 this. You see, like this. If you want to add  more new parameters to your ComfyUI, you can  

• 00:07:29 add it from here. In our post, you will see some of  the arguments like this. You see, like --gpu-only.  

• 00:07:37 So for example, let's use this --gpu-only. Copy  this, open a new terminal, and let's right-click,  

• 00:07:45 paste, and it will ask you, allow, and hit enter.  Now it will start the ComfyUI on the port 3000 and  

• 00:07:53 we will be able to connect ComfyUI from the 3000  proxy port. Since the installation takes huge  

• 00:08:00 time, I really recommend you to download models at  the same time, which I will show you in a moment  

• 00:08:07 how to download models with our unified model  downloader. My installation has everything that  

• 00:08:13 you would need, even ComfyUI manager, Accelerate,  Flash Attention, Sage Attention, xFormers,  

• 00:08:20 Deep Speed, Triton, whatever that you may need. So  this is a perfect installation of ComfyUI with all  

• 00:08:28 the libraries, and it's a clean installation. It  is the latest version of installation. You can see  

• 00:08:34 the printed messages here. It shows everything.  So this is a really, really clean and amazing way  

• 00:08:41 of installing ComfyUI on RunPod. Our installation  zip file also has installer for Windows and Massed  

• 00:08:47 Compute as well. So I am not only covering RunPod,  but I am covering Massed Compute and Windows as  

• 00:08:53 well. Moreover, the RunPod installation installs  on Ubuntu, so if you are a Linux user, you can use  

• 00:09:00 this installation. You see it shows that CUDA 0,  NVIDIA RTX 6000 Pro Blackwell Workstation Edition,  

• 00:09:08 using xFormers attention. Oh, by the way, if you  want to use Sage Attention, you need to add it here  

• 00:09:15 like --use-sage-attention and it will use the  Sage Attention instead of the xFormers. We are  

• 00:09:23 still waiting for it to start. The installation,  the model loadings, the start, all are extremely  

• 00:09:29 slow on RunPod because of their disk speeds.  Moreover, I also install GGUF support,  

• 00:09:35 so it is coming with GGUF. Okay, we can see that  the server started on the 3000 port. So how we are  

• 00:09:42 going to connect? Return back to Pods, click this  and click "Connect". Select the HTTP service 3000  

• 00:09:49 port and you see my ComfyUI started on the RunPod  3000 port. So how we are going to install SwarmUI?  

• 00:09:58 By the way, you see there are a lot of examples.  Let's go with the image generation. It says that I  

• 00:10:03 need to download this model from workflow. Perhaps  we can use FLUX. Okay, FLUX, FLUX Dev. Let's  

• 00:10:10 generate an image of the FLUX Dev. You see it will  download this model, but we also have a special  

• 00:10:16 downloader and it is downloading into my computer,  not into the RunPod. So therefore, let's continue  

• 00:10:22 with SwarmUI installer and also model downloader.  The link will be in the description of the video.

• 00:10:27 So download the latest zip file and also read  whatever written here. Click this "Upload files"  

• 00:10:33 icon, upload the downloaded zip file, let's  refresh. You will see that it is uploaded. Yes,  

• 00:10:38 completed. SwarmUI model downloader zip  file, "Extract Archive". So let's open the  

• 00:10:44 model downloader first because while installing  ComfyUI, you may want to download your models.  

• 00:10:49 All you need to do is just copy this, open a  terminal like this, paste and hit enter and it  

• 00:10:55 will start the model downloader application which  supports downloading models both ComfyUI and both  

• 00:11:01 SwarmUI. So I will make example of both of them.  Okay, you see it started on this URL, click it,  

• 00:11:08 and I want to download into ComfyUI backend.  So I right-click here, copy path, then open the  

• 00:11:15 started Gradio application interface. Now I will  download into not SwarmUI models, but into here,  

• 00:11:22 ComfyUI. By the way, don't forget to put backslash  here and actually we need to download into models  

• 00:11:30 folder. So let's also type models here. Then we  can select ComfyUI folder structure. So whichever  

• 00:11:37 the model you want to download, let's download  the image generation model, FLUX model, FLUX Dev,  

• 00:11:43 official model. Let's click download. You will  see that it will queue and start download. Okay,  

• 00:11:49 it is queued, starting the download, downloading  file. We can see the download progress here. Okay,  

• 00:11:54 let me zoom out. Yes. So currently the  speed is 70 megabytes per second. This  

• 00:12:00 is a terrible speed because we are on the  cloud and on Massed Compute, I get 300, 500,  

• 00:12:05 sometimes gigabytes of second speed. This is a  terrible speed for RunPod. Unfortunately, RunPod  

• 00:12:12 download speed is also not great even though I am  doing all of the optimizations. So we just need  

• 00:12:17 to wait for download. It will download it into the  accurate folder. So while installing the ComfyUI,  

• 00:12:23 you can download the models that you wish with  this application. It has so many bundles like  

• 00:12:30 the Wan 2.1 bundle, FLUX models bundle, HiDream  bundle. Also image generation models, FLUX models,  

• 00:12:36 different ones, HiDream models or Stable Diffusion  1.5 models or Stable Diffusion XL models,  

• 00:12:43 Stable Diffusion 3.5 models. It also has  other models like YOLO face segmentation,  

• 00:12:48 image upscaling. It also has text encoder models,  video generation models like Wan 2.1, Hunyuan  

• 00:12:54 models, Fast Hunyuan Skyreels, SkyVid, GenMo. It has  some LoRAs like migration LoRA, Wan 2.1 CausVid  

• 00:13:01 with LoRA. It also has some LLM models that are  used in some of the workflows. It has VAE, Clip,  

• 00:13:07 Vision models, ComfyUI workflows. So as I said,  you can use this downloader meanwhile installing  

• 00:13:15 to download models to save your time. Moreover,  if you need any more models in this application,  

• 00:13:22 you can always message me and hopefully I will  include them into the downloader as soon as  

• 00:13:27 possible. If this comes you so much confusing,  you can just search from here like SDXL and it  

• 00:13:34 will show you all the SDXL. Or you can search for  FLUX and it will show you all the FLUX. So it has  

• 00:13:41 an internal search feature which you can use. This  application is really, really programmed well and  

• 00:13:47 working really well, but this download speed is  terrible for a cloud platform, unfortunately.

• 00:13:53 For FLUX, we need to download Clip as well.  So let's download this Clip model. Okay,  

• 00:13:58 downloading Clip. Let's also download  the VAE of the FLUX, which is here,  

• 00:14:03 FLUX VAE. Yes. Example from the ComfyUI  wiki. This already supports everything. Oh,  

• 00:14:10 we are missing the T5 XXL in our previous. So  we also need to download it from our SwarmUI,  

• 00:14:17 T5 XXL here, T5 XXL model, T5 XXL. Now it is  downloading the T5 XXL. We can see here. Yes,  

• 00:14:26 we just need to wait. Then we will be able to  generate it as we expect, let's say. So we have  

• 00:14:34 the FLUX Dev here. You see, you need to click here  to select. We have the T5 XXL, just refresh. Okay,  

• 00:14:41 here. Clip Large here. This is FLUX type,  default. We have the FLUX VAE like this. Make  

• 00:14:47 sure to select them. This is FLUX guidance. So it  is 3.5 by default. We have everything set here. We  

• 00:14:55 just need to run and it should now generate  as we want. Okay, it is loading the models.  

• 00:15:01 After the initial load, it should be really fast  because we are keeping everything in the VRAM.  

• 00:15:06 You can download every one of these from the  downloader interface. The best way would be  

• 00:15:12 just download the FLUX models bundle and it will  download everything. And image has been generated.  

• 00:15:18 It is not the greatest quality because the base  model is not great. You can use fine-tuned models,  

• 00:15:24 you can use LoRAs. This was just a demonstration  and you can already see the speed. The speed is  

• 00:15:30 just mind-blowingly fast. Can we see the step  speed somewhere? Okay, the step speed is 3.66  

• 00:15:38 IT per second. This is just amazing. So it takes  less than six seconds to generate an image on this  

• 00:15:46 GPU with ComfyUI. The ComfyUI is hard to use, so  you really should find workflows and use them. But  

• 00:15:53 now I will move to the SwarmUI. Since my ComfyUI  is already running, I need to terminate it. For  

• 00:16:00 terminating it, I need to use this command. If you  don't start the ComfyUI, you don't need it. Now I  

• 00:16:06 have terminated it and my VRAM usage should  have been dropped. Yes. Now I will install  

• 00:16:12 the SwarmUI and show you how to use SwarmUI  with ComfyUI backend, which I recommend. Okay,  

• 00:16:18 double-click RunPod SwarmUI install instructions,  copy this entirely, open a new terminal,  

• 00:16:25 and it will automatically install the SwarmUI for  you. Since we have installed the ComfyUI already,  

• 00:16:31 it will be super fast to install. So just wait for  installation to start on the localhost. Okay, we  

• 00:16:38 got the Cloudflared and also we got the localhost  URL. Okay, the Cloudflared has started already. So  

• 00:16:45 agree, customize settings. This is important  because don't use just install, customize  

• 00:16:50 settings, select your template. This is yourself  on this PC. This is none because we already have  

• 00:16:57 ComfyUI. Next, I don't want to download any  models. Next, yes, I'm sure, install. It will  

• 00:17:03 be instant as you are seeing right now. Then go  to server, go to backends, ComfyUI self-starting.  

• 00:17:09 Now this is super important. Click okay. So where  is our ComfyUI installation? It is inside here,  

• 00:17:14 ComfyUI, and it is the main.py. So copy path and  paste it here like this. And which arguments we  

• 00:17:23 want? These are the arguments from the ComfyUI.  --use-sage-attention and also, since this is a  

• 00:17:30 very high VRAM GPU, let's also use --gpu-only  and copy it and save. Then it will start the  

• 00:17:37 backend from this ComfyUI. So it will be super  fast. It will be using the latest ComfyUI with  

• 00:17:43 all the libraries that we would need. Let's also  download some of the models. Let's test out Wan  

• 00:17:50 2.1. So for that, I will use the model installer  again. So let's copy this. I had closed it. Let's  

• 00:17:57 start again. Okay, it is started. Let's go to the  Gradio. Meanwhile, it is loading the backend. When  

• 00:18:03 you go to logs and debug, you will see that it is  installing the necessary additional libraries for  

• 00:18:10 SwarmUI, nothing else. So it will be really fast  compared to initial install of the ComfyUI. Then  

• 00:18:16 let's go to the bundles and let's download Wan  2.1 core models bundle. This may take a while,  

• 00:18:23 unfortunately, because of the download speed, but  the download has started. This time it is a little  

• 00:18:29 bit faster than before. And make sure that your  pod is not out of volume. This is where the models  

• 00:18:36 will be downloaded. And by default, you see that  it has recognized my SwarmUI models folder, so  

• 00:18:42 they will be downloaded into the accurate folder.  You can also give different folder from the server  

• 00:18:48 configuration. You can give your model root. So  if you want to use them from another folder, like  

• 00:18:53 from the ComfyUI models folder, you can change  the root model and it will work. This applies to  

• 00:18:58 Windows as well. Everything applies to Windows,  Linux, cloud, it is all same. And it is still  

• 00:19:04 installing the necessary packages and meanwhile,  we are downloading the models. So I really  

• 00:19:09 recommend you to download models at the same  time as installation on RunPod to save your time.

• 00:19:17 Okay, so the downloads have been completed.  You can also see in here and the backend has  

• 00:19:22 been loaded. So I am going to use CausVid LoRA  with Wan 2.1 and we already have presets. It  

• 00:19:29 is inside the SwarmUI model downloader. You will  see it when you extract the zip archive. You will  

• 00:19:34 see that there are amazing SwarmUI presets.  So let's import preset, select the preset,  

• 00:19:40 import. It is imported. Quick tools, reset  params to default, select "Fast CausVid With  

• 00:19:46 Image to Video" here. And from models,  let's go to models, let's refresh. Okay,  

• 00:19:52 it is here. Let's use "Image to Video GGUF". By  the way, this has downloaded Q6, but on this GPU,  

• 00:19:58 you can use the FP16 version as well. So since  this is a very powerful GPU, and from presets,  

• 00:20:05 let's uncheck this and direct apply. Everything  applied. By the way, this model native resolution  

• 00:20:11 is not 640 to 640. Let's edit metadata. It is  960 to 960 like this. Okay, and this is image to  

• 00:20:21 video model. Let's generate 81 frames and let's  go to the init image. Let's choose file. Okay,  

• 00:20:27 let's try this dinosaur. Let's make the init  image creativity zero, resolution closest aspect  

• 00:20:33 ratio. You see, this is the resolution. Okay,  I need to check this and uncheck this. Okay,  

• 00:20:38 like this. Now it is fixed. And yes, we  set the resolution. We also need to type  

• 00:20:43 a prompt. A running Tyrannosaurus Rex. Okay, like  this. And we are ready. So this will generate this  

• 00:20:52 video in eight steps. Oh, we have forgotten the  LoRA. So let's interrupt all sections. Go to LoRA,  

• 00:20:58 select this LoRA. Okay, we are now ready. So  this will generate this video in eight steps and  

• 00:21:05 it should be super fast on this GPU. We are also  keeping everything in VRAM. Let's see. pip install  

• 00:21:12 nvitop. nvitop. This is a 96 gigabyte  VRAM having GPU. We have to wait for model to  

• 00:21:19 be loaded. We are also using Sage Attention and we  are keeping everything on GPU. So the generation  

• 00:21:26 is starting. We could also install frame  interpolation as I have shown in the previous  

• 00:21:31 tutorials. It is all same. I really recommend you  to use SwarmUI if you are not an expert of the  

• 00:21:36 ComfyUI because it is really, really hard and it  gets broken. You need to find accurate workflow,  

• 00:21:43 download all the accurate files. And if you don't  know how you can download files to the RunPod,  

• 00:21:48 just type "SECourses wget" into the YouTube  search and you will see this tutorial.  

• 00:21:53 This tutorial shows you how to use wget command  and also amazing Hugging Face upload and download  

• 00:22:00 Jupyter Notebook. So I really recommend to watch  this tutorial if you want to learn how you can  

• 00:22:06 upload and download big files to the RunPod. Okay,  it is using 35 gigabytes of VRAM while the entire  

• 00:22:14 model is in the GPU. It is generating the video  right now in the 720p. So it's an HD resolution.  

• 00:22:23 Let's see the step speed in a moment. Okay, so the  step speed is 16 second / IT on this GPU. It will take  

• 00:22:31 less than two minutes to generate this natively HD  video. By the way, if we were using FP16 version,  

• 00:22:39 perhaps it would be faster. So let's also  download FP16 Wan 2.1. It is inside the Wan  

• 00:22:47 models. So let's turn this off. Let's go to the  video generation models, Wan 2.1 models, and I am  

• 00:22:54 going to download FP16 image to video here, this  one. Okay, let's click the download. Meanwhile,  

• 00:23:00 it is generating, it will download the model into  the accurate folder. It is 32 gigabytes of disk  

• 00:23:06 space. Make sure that your pod is not getting  out of disk, and we still have space. And it is  

• 00:23:12 downloading fast, nice. And we can also see the  difference of between two models, but GGUF Q6 is  

• 00:23:19 also a great quality model with great precision.  So yes, it is now 15 second IT. When you are using  

• 00:23:26 CausVid LoRA, you can just use eight steps and you  don't need to use Teacache or other optimization  

• 00:23:34 with Sage Attention, it works really, really  good and fast. And this is natively HD resolution  

• 00:23:41 video. Let's also install frame interpolation  after this. Okay, it may take a while. Okay,  

• 00:23:48 it didn't animate greatly this, but let's see  the difference with FP16. Maybe this prompt  

• 00:23:55 wasn't good enough too. Okay, and let's install  frame interpolation. Meanwhile, the FP16 model  

• 00:24:01 is getting downloaded. When you install Teacache,  it will reload the backends. And when you click  

• 00:24:06 "Display Advanced Options", you will see all the  settings we set, like here. Okay, FP16 version has  

• 00:24:13 been downloaded. We are still waiting backends  to be loaded. And by the way, I am also working  

• 00:24:19 on another tutorial for Wan 2.1 with VACE, and it  will be really great. Let me show you example of  

• 00:24:26 VACE. So you see, this was a video I generated by  using a dog running as a driving video and using  

• 00:24:35 this image. So this is way better compared to some  random animation. Okay, backend loaded. So let's  

• 00:24:41 change our model. To change our model, we will  change it from here, like this. And let's also  

• 00:24:47 refresh here, and let's also select the model  from here, new downloaded model. Moreover, from  

• 00:24:52 the advanced sampling, I will set the precision  as 16 bit and let's generate. I wonder how much  

• 00:25:00 VRAM it will use now since we will use the FP16  model. Okay, let's wait for model to be loaded,  

• 00:25:07 but it is really, really slow. For some reason,  it didn't start, so let's interrupt all sessions.  

• 00:25:13 Let's also enable RIFE and make it 2x. Okay, it  still shows two generations. I think we got a  

• 00:25:20 bug at somewhere. So let's refresh. Okay, I think  it did refresh. Okay, everything is looking good.  

• 00:25:27 Let's try again. Okay, it says one generation,  one queued. Nice. Let's go to the logs. Sometimes  

• 00:25:32 this can happen. And let's see. Okay, now it  is loading the model. I can see that. Good. Oh,  

• 00:25:39 we got an inaccurate resolution because of the  metadata of newly downloaded model. But first,  

• 00:25:46 let's just wait for initial generation or load.  This model loading taking forever on RunPod,  

• 00:25:52 unfortunately. Okay, finally, the model has been  loaded. You see it is using 57 gigabytes of VRAM  

• 00:25:59 and it is entirely loaded into the GPU. We are not  using block swapping or any optimization. This is  

• 00:26:06 the highest quality that you can get in terms of  the model precision. And let's see if there is a  

• 00:26:12 big animation. I think this image and this prompt  is not very animatable. Yeah, it is also doing  

• 00:26:19 RIFE frame interpolation and it is generated, but  the animation is still very limited. Let's disable  

• 00:26:28 the CausVid LoRA and let's change the resolution  to the accurate. So let's edit metadata and make  

• 00:26:35 it properly like this, save. Okay, change model.  Yes, it is fixed. We removed the LoRA. Okay,  

• 00:26:42 there is a prompt. Let's try this one. I  wonder how much difference it will make. Okay,  

• 00:26:46 this prompt also didn't make much difference. As  a final step, I am testing with Wan 2.1 image to  

• 00:26:53 video. So this is the preset. Just direct apply,  and I changed the frame count to 81 and video  

• 00:27:00 steps to 50. I'm not using Teacache. So this is  the official usage of the Wan 2.1 image to video.  

• 00:27:08 Let's see if there will be much difference or not.  This will take a while, but if this is also same  

• 00:27:13 way, not very much animated, that means that this  image and this prompt with this model is not much  

• 00:27:21 compatible. You need to use Wan 2.1 VACE with a  driving video. All right, this generation has made  

• 00:27:29 a significantly different video as you are seeing  right now. So it is best to install Teacache as in  

• 00:27:36 the previous tutorials and use it in that way if  your animation is not that great. It's all about  

• 00:27:45 some experience, experimentation and testing, and  hopefully you will find your very best workflow  

• 00:27:50 for your case. If you have any questions, just  message me on Patreon, just reply to this video,  

• 00:27:56 or contact me from Discord. Hopefully,  amazing more tutorials coming soon. Thank you.