Skip to content

MultiTalk Full Tutorial With 1 Click Installer Make Talking and Singing Videos From Static Images

Jump to bottom
FurkanGozukara edited this page Oct 16, 2025 · 1 revision

MultiTalk Full Tutorial With 1-Click Installer - Make Talking and Singing Videos From Static Images

Full tutorial link > https://www.youtube.com/watch?v=8cMIwS9qo4M

MultiTalk Full Tutorial With 1-Click Installer - Make Talking and Singing Videos From Static Images

image Hits Patreon BuyMeACoffee Furkan Gözükara Medium Codio Furkan Gözükara Medium

YouTube Channel Furkan Gözükara LinkedIn Udemy Twitter Follow Furkan Gözükara

By using MeiGen MultiTalk you can generate amazing fully animated real-like videos from given audio input. Not only talking but also animating the body movements is possible. In this video I will show you how to install ComfyUI on Windows and MultiTalk bundle and workflows we prepared with 1-click. Then I will show how to very easily generated amazing videos from these installed workflows. Moreover, I will show our favorite cloud private GPU provider Massed Compute. How to install same there and use it properly. Finally I will show everything on RunPod as well. So whether you are GPU poor or have good GPU, this tutorial covers everything.

🔗Follow below link to download the zip file that contains MultiTalk bundle downloader Gradio App - the one used in the tutorial ⤵️

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

🔗Follow below link to download the zip file that contains ComfyUI 1-click installer and the WORKFLOW shown in tutorial that has all the Flash Attention, Sage Attention, xFormers, Triton, DeepSpeed, RTX 5000 series support ⤵️

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

🔗 Python, Git, CUDA, C++, FFMPEG, MSVC installation tutorial - needed for ComfyUI ⤵️

▶️ https://youtu.be/DrhUHnYfwC0

🔗 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 Intro & MultiTalk Showcase

00:00:28 Singing Animation Showcase

00:00:57 Tutorial Structure Overview (Windows, Massed Compute, RunPod)

00:01:10 Windows - Step 1: Download & Extract the Main ZIP File

00:01:43 Windows - Prerequisites (Python, Git, CUDA, FFmpeg)

00:02:12 Windows - How to Perform a Fresh Installation (Deleting venv & custom_nodes)

00:02:42 Windows - Step 2: Running the Main ComfyUI Installer Script

00:04:24 Windows - Step 3: Installing MultiTalk Nodes & Dependencies

00:05:05 Windows - Step 4: Downloading Models with the Unified Downloader

00:06:18 Windows - Tip: Setting Custom Model Paths in ComfyUI

00:07:18 Windows - Step 5: Updating ComfyUI to the Latest Version

00:07:39 Windows - Step 6: Launching ComfyUI

00:07:53 Workflow Usage - Using the 480p 10-Second Workflow

00:08:07 Workflow Usage - Configuring Basic Parameters (Image, Audio, Resolution)

00:08:55 Workflow Usage - Optimizing Performance: 'Blocks to Swap' & GPU Monitoring

00:09:49 Workflow Usage - Crucial Step: Calculating & Setting the Number of Frames

00:10:48 Workflow Usage - First Generation: Running the 480p Workflow

00:12:01 Workflow Usage - Troubleshooting: How to Fix 'Out of VRAM' Errors

00:13:51 Workflow Usage - Introducing the High-Quality Long Context Workflow (720p)

00:14:09 Workflow Usage - Configuring the 720p 10-Step High-Quality Workflow

00:16:18 Workflow Usage - Selecting the Correct Model (GGUF) & Attention Mechanism

00:17:58 Workflow Usage - Improving Results by Changing the Seed

00:18:36 Workflow Usage - Side-by-Side Comparison: 480p vs 720p High-Quality

00:20:26 Workflow Usage - Behind the Scenes: How the Intro Videos Were Made

00:21:32 Part 2: Massed Compute Cloud GPU Tutorial

00:22:03 Massed Compute - Deploying a GPU Instance (H100)

00:23:40 Massed Compute - Setting Up the ThinLinc Client & Shared Folder

00:25:07 Massed Compute - Connecting to the Remote Machine via ThinLinc

00:26:06 Massed Compute - Transferring Files to the Instance

00:27:04 Massed Compute - Step 1: Installing ComfyUI

00:27:39 Massed Compute - Step 2: Installing MultiTalk Nodes

00:28:11 Massed Compute - Step 3: Downloading Models with Ultra-Fast Speed

00:30:22 Massed Compute - Step 4: Launching ComfyUI & First Generation

00:32:45 Massed Compute - Accessing the Remote ComfyUI from Your Local Browser

00:35:07 Massed Compute - Downloading Generated Videos to Your Local Computer

00:36:08 Massed Compute - Advanced: Integrating with the Pre-installed SwarmUI

00:38:06 Massed Compute - Crucial: How to Stop Billing by Deleting the Instance

00:38:33 Part 3: RunPod Cloud GPU Tutorial

00:39:29 RunPod - Deploying a Pod (Template, Disk Size, Ports)

00:40:39 RunPod - Connecting via JupyterLab & Uploading Files

00:41:11 RunPod - Step 1: Installing ComfyUI

00:42:32 RunPod - Step 2: Downloading Models

00:45:26 RunPod - Step 3: Installing MultiTalk Nodes

00:45:52 RunPod - Step 4: Launching ComfyUI & Connecting via Browser

00:47:50 RunPod - Running the High-Quality Workflow on the GPU

00:51:11 RunPod - Understanding the Generation Process on a High-VRAM GPU

00:52:34 RunPod - Downloading the Final Video to Your Local Machine

00:53:04 RunPod - How to Stop & Restart a Pod to Save Costs

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

Video Transcription

  • 00:00:00 Greetings, everyone. Welcome to the Wan 2.1  based MultiTalk tutorial. In this tutorial,

  • 00:00:06 I will show you how to literally one-click  to install ComfyUI and MultiTalk and right

  • 00:00:12 away start generating amazing animations from  static images. MultiTalk generates not only

  • 00:00:18 speaking videos but also impressive  singing performances. Moreover,

  • 00:00:23 MultiTalk can even generate 30-second  videos like this upcoming one.

  • 00:00:28 SECourses shows the way. Generate, compose, and play.

  • 00:00:42 Images and videos flow, Animation's 3D glow.

  • 00:00:57 Learning AI every day, SECourses shows the way. So as usual, I have prepared an amazing  post. I really recommend you to read this

  • 00:01:03 post fully. I will begin with Windows, then  Massed Compute, and then finally RunPod.

  • 00:01:10 First of all, we need to download the zip file.  The link will be in the description of the video.

  • 00:01:14 So you see the latest zip file here. Move this  zip file wherever you are going to install or

  • 00:01:20 your previous installation, like here. Replace.  Then right-click and you need to extract all.

  • 00:01:26 I'm going to use here, extract files, okay.  You see everything extracted here. Overwrite

  • 00:01:32 the previous files. You can follow these steps for  a fresh installation as well, but you need to be

  • 00:01:38 careful is that do not have space character  in your folder path or special characters.

  • 00:01:43 If you are first time following my  videos, you need to have Windows

  • 00:01:48 requirements tutorial followed and installed.  You need to have Python, Git, CUDA, FFmpeg,

  • 00:01:54 and other stuff. So follow this tutorial. The link  will be also in the description of the video. This

  • 00:01:59 workflow has been tested on a fresh installation.  So if you don't want to make a fresh installation,

  • 00:02:06 still it should work, but if it doesn't work,  you need to make a fresh installation. Moreover,

  • 00:02:12 you can make fresh installation with your existing  installation like this, which I'm going to show.

  • 00:02:18 So first of all, let's delete this empty  folder. Then I'm going to enter inside my

  • 00:02:22 ComfyUI folder. What I need to do is I  need to delete this virtual environment

  • 00:02:26 folder like this. It is going to delete.  Okay, it is deleted. Then I also need to

  • 00:02:32 delete the custom nodes. You need to do  this if it doesn't work. After doing that,

  • 00:02:37 first of all, I will do installation. The fresh  installation is also same, what you need to do.

  • 00:02:42 Double-click windows_install.bat file. More  info, run anyway. Do not run as administrator.

  • 00:02:50 Then you will get to this screen. I really  recommend you to have Python 3.10, but you can

  • 00:02:55 also use Python 3.11 or 12. However, do not use  Python 13. It will not work. So I will choose the

  • 00:03:03 option one. I'm going to use both Flash Attention  and Sage Attention, three. I know that some of you

  • 00:03:08 are wondering whether we should use Sage Attention  2.2 plus. I tested it and it is same. It doesn't

  • 00:03:15 bring any speed benefits. So our Sage Attention  is currently working with the best performance.

  • 00:03:22 Many of the libraries are automatically installed  which you may need. Moreover, it is supporting

  • 00:03:27 RTX 5000 series GPUs as well, as well as RTX 4000  series GPUs or 3000 series GPUs. If you have older

  • 00:03:35 GPUs like 2000 or 1000, some of the workflows  may not work or they may work slower. However,

  • 00:03:42 starting from 3000 series GPUs, it is working just  perfect. If you are making a fresh installation,

  • 00:03:47 it is exactly same. You just double-click and  start it. Do not run anything as administrator

  • 00:03:53 unless I state it explicitly because if you run  that way, it will break your installation. Just

  • 00:04:00 wait for installation to be completed. You need  this step only if your existing ComfyUI doesn't

  • 00:04:05 work. And ComfyUI may get broken very easily,  therefore I am showing you all the cases. And

  • 00:04:12 don't forget to follow this Windows requirements  tutorial. This is mandatory only one time. All of

  • 00:04:17 my applications, AI applications, scripts will  work if you follow this video just for once.

  • 00:04:24 Okay, so the installation has been completed. Then  enter inside workflows folder, enter inside Kijai

  • 00:04:31 MultiTalk, and double-click install_MultiTalk.bat  file. More info, run anyway. This will download

  • 00:04:39 and install necessary nodes that you need  to run this workflow. So it is all automatic

  • 00:04:46 for you. You will not be wasting time with the  ComfyUI manager. It will just handle everything

  • 00:04:52 automatically for you. Okay, everything has been  installed. You see we have the workflows here,

  • 00:04:58 but we are not ready yet. What we need is go to  the top of the page and you will see that we have

  • 00:05:05 SwarmUI Installer and Unified Massive Models  Downloader. Go to here. This link will be also

  • 00:05:11 in the description of the video and download  the latest version, SwarmUI model downloader.

  • 00:05:16 Extract this into anywhere. I'm just going to use  my existing installation, so you also will see how

  • 00:05:24 I update, which is here. Then we need to extract  it, extract files, overwrite your older files,

  • 00:05:31 so they will get updated. Then double-click  windows_start_download_models_app.bat file.

  • 00:05:37 More info, run anyway. This will start the  model downloader. Why we need this? Because

  • 00:05:42 with this, we will avoid manually  downloading the necessary models.

  • 00:05:46 So what we need is, we need to give a custom path.  Currently, it is using the SwarmUI path, but we

  • 00:05:52 need to give the ComfyUI model path. So my ComfyUI  model path is here, models. So I will copy this,

  • 00:05:59 paste it here, then I will check this ComfyUI  folder structure. Then you see there is ComfyUI

  • 00:06:05 bundles, download ComfyUI MultiTalk bundle. It  will download all these models which are necessary

  • 00:06:11 to run this workflow. Some people are asking me  how they can set custom path for ComfyUI models.

  • 00:06:18 ComfyUI has an amazing feature. You see there is  extra_model_paths.yaml.example. Open this file

  • 00:06:25 with any text editor, read here and change it if  you need custom paths. I don't use custom paths,

  • 00:06:32 but if you need, this is the way of it. So this  downloader will download everything automatically

  • 00:06:36 for you with maximum speed, absolutely maximum  speed. It will skip the existing models if you

  • 00:06:41 already have. We can see the download speed is  here. It is around 100 megabytes per second.

  • 00:06:46 It is around 800 megabits. So it will use your  entire internet speed. If you get errors here,

  • 00:06:53 it could be due to your antivirus, your VPN,  your internet service provider, or temporary

  • 00:07:00 problem with the Hugging Face. So you need to keep  trying until it is successful. You can see that it

  • 00:07:06 has queue of eight and it will download everything  automatically for us. Just follow the status here.

  • 00:07:12 Okay, so all the models have been downloaded.  As a last step, what we need is using

  • 00:07:18 windows_update_comfyui.bat file. Do not forget  this. Double-click, more info, run anyway. This

  • 00:07:25 will update ComfyUI to the latest version. It  will reinstall new versions of libraries and

  • 00:07:33 you see it is done already. Then all I need is  windows_run_gpu.bat file, more info, run anyway.

  • 00:07:39 It will start the ComfyUI on my Windows computer.  Okay, so the ComfyUI has been started. What I

  • 00:07:46 need is I need to drag and drop my workflow. So I  have prepared a lot of different workflows based

  • 00:07:53 on what you need. So let's begin with simple  480p 10-second workflow. After 10 seconds,

  • 00:08:01 this workflow degrades in quality. However, up to  10 seconds, it is really good. So which parameters

  • 00:08:07 you need to change? These are same among all of  the workflows. First of all, you need to load your

  • 00:08:12 image. I really recommend you to start with our  test image. It is inside workflows. So load it.

  • 00:08:18 Then I really recommend you to start with  test MP3 file. It is also inside the workflow

  • 00:08:24 folder. Let me show you. You see test.jpeg and  test.mp3. Then you need to set your width and

  • 00:08:30 height according to your input image. This is  480p model and this is a tall image. Therefore,

  • 00:08:37 this is the resolution that I did set. If  your image is not matching this resolution,

  • 00:08:41 it will keep proportion and crop it and center  crop. You see the parameters are all here.

  • 00:08:48 Okay, what else you need to set? Number of blocks  to swap. It is by default set to 25. If you are

  • 00:08:55 low on GPU, you can set this up to 40. How  to decide? So open a CMD, type pip install

  • 00:09:03 nvitop like this. Then type nvitop and then you  are going to pay attention to the watt usage of

  • 00:09:09 your GPU because if it is not using enough power,  that means that it is using shared VRAM. Okay, so

  • 00:09:17 25 is default and it is good. Then the next most  important parameter you have to set is the prompt.

  • 00:09:24 So you see, by default, we have a man patiently  singing, a close-up shot captures his expressive

  • 00:09:31 performance. You need to change this according to  your image and which video you want to generate.

  • 00:09:36 And this is the default negative prompt. Then  you can enable VAE tiling if you are on low VRAM.

  • 00:09:43 Other than that, you need to set number of frames.  Now, this is super important. You need to set it

  • 00:09:49 according to the length of your audio or how much  you want to animate. So my audio is 9 seconds,

  • 00:09:56 but it is more than 9 seconds actually. So you  need to have precise timing. For precise timing,

  • 00:10:02 I am using MediaInfo from here. You see this  MediaInfo and it shows me that it is 9 seconds,

  • 00:10:10 960 milliseconds. You can use any application to  get the exact resolution. This is just MediaInfo

  • 00:10:15 application, you can install it. So this is 9  seconds, 960 milliseconds, which makes 9.960

  • 00:10:23 multiplied with 25, 249 frames. So I'm just  going to use default 250 frames. However,

  • 00:10:30 this workflow is not exactly as precise. As you  generate, you will see. Therefore, you may be need

  • 00:10:37 to set the frame count accordingly. And once  you set frame count, your prompt, your input

  • 00:10:44 image and its resolution, actually you are ready.  You don't need to change anything else. So let's

  • 00:10:48 just hit the run and let's see the result. So this  is from my previous generation, it remembers it.

  • 00:10:54 So let's watch our VRAM usage and the watt usage.  With this way, we can decide whether we are

  • 00:11:01 utilizing GPU with the maximum performance or  not. If it is not, we can increase the block

  • 00:11:06 count or we can reduce the block count. This is  the way of deciding block swap count. Moreover,

  • 00:11:12 you need a lot of RAM to run these models. If you  don't have sufficient RAM, I really recommend you

  • 00:11:18 to set at least 100 GB of virtual RAM. If you  don't know how to set a virtual RAM, type how to

  • 00:11:25 set virtual RAM and go to this link. It shows you  how to set your virtual RAM. It is so easy. Okay,

  • 00:11:31 so our generation started. We are using 500 watts,  which is amazing. You see it is sometimes dropping

  • 00:11:37 to 400 watts because it is doing some block  swapping, but as you are seeing, 500 watts is

  • 00:11:43 amazing because my maximum is 575. So I am with  over like 80% of performance and it is generating

  • 00:11:50 the video right now. It will be pretty quick even  though I am recording a video right now. So I am

  • 00:11:55 using some GPU already. So let's see how much  time it will take. If you still get some out of

  • 00:12:01 VRAM error, what you can do is you can enable VAE  tiling wherever you see. For example, here there

  • 00:12:08 is enable VAE tiling. Also here, tiled VAE. So  pay attention to the nodes and if you are seeing

  • 00:12:15 tiling, enable it if you get out of VRAM error,  out of memory error. How you can know whether

  • 00:12:21 you got out of VRAM error or not? Just follow this  CMD window, the terminal. And you may notice that

  • 00:12:27 it is generating with four steps to four steps.  Yes, it cannot generate entirely the video at

  • 00:12:35 once because this is 250 frames and this model  is actually made for 81 frames. Therefore, it is

  • 00:12:43 using some embedding system for this workflow and  that is how it is generating parts by parts. This

  • 00:12:50 sampler is four steps, you cannot change it. This  is generating decent quality, but I will show you

  • 00:12:56 a better quality as well. Moreover, as I said,  this workflow is up to 10 seconds. If we need

  • 00:13:02 more than 10 seconds, what you need to use is the  long context generation, which I will show next.

  • 00:13:09 Okay, so the video has been generated. It shows  that it used maximum 16 GB of VRAM. It took 258

  • 00:13:17 seconds while I am recording full HD 4K resolution  video. And we got the result here. So when I hover

  • 00:13:24 my mouse here, it will also show the audio. How  you can save it? Right-click and you can save

  • 00:13:31 as preview. It will download it. Moreover, it  is by default saved inside the output folder,

  • 00:13:38 inside ComfyUI, inside output, and you will see  that the video is there, the PNG which contains

  • 00:13:45 the workflow is there, and there is also video  itself without the audio file. This quality is

  • 00:13:51 not great. If we want a better quality, let's do  it. So for better quality, we have long context

  • 00:13:58 generation. This is better quality and even  I have 10 steps version of it. So let's see

  • 00:14:03 the 10 steps version. So drag and drop. With 10  steps version, you don't need to set anything,

  • 00:14:09 everything is set. Let me set up from beginning  what you need to change. So choose file to upload,

  • 00:14:15 let's use our test. Select our MP3 file, test.  How many frames? It is around 250 frames again.

  • 00:14:22 Then what we else need to set? Nothing else,  actually. We just need to set the prompt as usual,

  • 00:14:29 as before. So the prompt is here. A man patiently  talking. It is remembering from my previous prompt

  • 00:14:36 probably. So let's actually close everything to  open from beginning. I have been doing a lot of

  • 00:14:42 tests for you. Okay, let's open unsaved and let's  drag and drop our high quality. Yes. Okay. Okay,

  • 00:14:49 I think it's still remembering. Anyway, not very  important. So let's set up our parameters. Okay,

  • 00:14:55 250 frames. This is extremely precise with  frame count. It is working better than the

  • 00:15:00 other workflow. So a man patiently singing. Okay,  I did set the frames count, I did set the prompt,

  • 00:15:07 I did set the load audio and image. I also  need to set the resolution. So therefore,

  • 00:15:13 this will be 720 and 1280 pixel resolution since  this is HD resolution. This will take longer than

  • 00:15:23 the previous generation. Why? Because this  will be 10 steps, the quality version. It

  • 00:15:28 will be higher resolution and I'm going to use  25 block swaps again. It works with my RTX 4090,

  • 00:15:35 but depending on your GPU, you can just make it  40. Okay, you don't need to change anything else

  • 00:15:41 except if you get out of VRAM error, use enabled  VAE tiling here. Then also there are some other

  • 00:15:48 enabled VAE tile here. You see tiled VAE and also  there is VAE tiling also here. Or you can just

  • 00:15:56 use my low VRAM configuration. You see this one.  This one is 15 GB and this one is like 8 GB. You

  • 00:16:04 need to test and see which one is working best for  you and run. So let's see the result of this one

  • 00:16:11 and this one will use definitely more VRAM. It is  going to load the new 720p model. Oh, by the way,

  • 00:16:18 if you already have your models, you need to  select them from here. For this workflow, I am

  • 00:16:22 using 1 2.1 14 billion parameters image to video  720p GGUF Q8. This is excellent quality model. You

  • 00:16:32 can select this. For the previous workflow, I am  using 480p. So it is up to you, select your model.

  • 00:16:39 If you are not using GGUF, you can also select  quantization, but with GGUF, it is disabled.

  • 00:16:44 And we are using Sage Attention. If you want  other attentions like Flash Attention, you can

  • 00:16:49 also change them from here. By the way, my auto  installer doesn't have Flash Attention 3 yet, so

  • 00:16:55 it wouldn't work. We don't have flex attention as  well. We have Flash Attention 2, Sage Attention,

  • 00:17:00 and SDPA. However, I can tell you that the Sage  Attention is the best working one with the best

  • 00:17:06 speed at the moment. Okay, I think we are going to  start the steps. So this will take a lot of time.

  • 00:17:12 However, the quality will be significantly better  than the previous generation. Currently, I am

  • 00:17:17 using 28 GB of VRAM with all the things running.  Therefore, actually, I could reduce the number of

  • 00:17:24 block swapping to speed up my generation. So it is  up to you to test yourself. For example, I could

  • 00:17:30 make this 20. It would use more of my VRAM, more  of my GPU, and it would be faster in that case.

  • 00:17:37 This is how it works. Also on the CMD, if you pay  attention, you will see that it is showing how

  • 00:17:43 much of the model is on the CPU, offloaded block  swapping, and how much of it on the GPU right now.

  • 00:17:51 Okay, so our video has been completed and I will  make a comparison, but before making a comparison,

  • 00:17:58 if you don't like the results, what you can  do is trying different seeds. So currently,

  • 00:18:04 this is set as seed 8. When you change this  seed, you may get a perfect output. So therefore,

  • 00:18:11 you can try changing seeds and see if you are  getting a better result or not. It is totally

  • 00:18:16 up to you. Moreover, with ComfyUI, you can just  change this and click run and it will queue,

  • 00:18:22 change it, click run, it will queue. You can  also make multiple generations and so on. But

  • 00:18:28 I will just terminate everything. So how am I  going to compare these two generated videos? We

  • 00:18:36 have an upcoming amazing application,  Ultimate Video Upscaler. Hopefully,

  • 00:18:42 it will be published this week or maybe next week.  It is almost completed. This application has been

  • 00:18:49 developed with the mindset of professional  level upscaling. So it has so many features.

  • 00:18:55 It is almost done. I am going to use the output  comparison tool we have here. So we have up to

  • 00:19:03 four videos comparison. Let's select the  two outputs and make a comparison video.

  • 00:19:09 I am going to use the first generation, which is  here, and the second generation, which is here.

  • 00:19:17 So the left one will be the 480p default workflow  and the right one will be the high quality 720p.

  • 00:19:25 So let's make a side-by-side comparison. It  is really fast, the comparison video. Okay,

  • 00:19:30 it is made. Let's download it. Now I will  play it. So let's see the difference.

  • 00:19:35 SECourses lights the way, through  the maze where the data play.

  • 00:19:45 As we can see, there is a huge, huge difference  in quality. I can try different prompts,

  • 00:19:52 different seeds, and get better quality, but  this is how it works. So I can tell you that

  • 00:19:58 the long context has a better quality than  the default 10-second workflow in many cases,

  • 00:20:04 not in all cases, but in many cases. I  cannot say that it will work every case.

  • 00:20:09 So it is up to you to test, unfortunately.  The testing is the big part of AI generation,

  • 00:20:15 finding the very best workflow for your usage  case. But I did a lot of testing and I have

  • 00:20:21 prepared all these presets, the workflows  for you, so you can use them right away.

  • 00:20:26 If you are wondering how I made the videos at  the beginning of the tutorial, at the intro,

  • 00:20:33 so basically, I have used the free credits of  the 11Labs, generated audios with text-to-speech,

  • 00:20:40 downloaded them, gave my audio file, image, and  I just changed the prompt as a man talking or a

  • 00:20:49 character talking, such as that. But I didn't  like the 11Labs audio quality. Definitely,

  • 00:20:55 it is not that great. And I am using the  11Labs version 3 alpha, the very best

  • 00:21:00 version they have. Therefore, hopefully, I will  make the greatest local text-to-speech for you,

  • 00:21:07 hopefully in near future. But this is how I made.  So basically, uploading the files as I have shown

  • 00:21:13 and generating the videos until I get the result  I want. Even right now, I am generating some of

  • 00:21:20 the intro videos that you are going to see them  at the beginning of the tutorial, hopefully.

  • 00:21:25 Now, I am going to begin showing you how to use  this amazing workflow on Massed Compute. If you

  • 00:21:32 don't have a powerful GPU, you can use Massed  Compute cloud service. So I really recommend

  • 00:21:37 you to watch the first part of the tutorial,  then you can return back here and continue.

  • 00:21:44 So first of all, we need to download the zip  file. The link will be in the description of

  • 00:21:47 the video. Put this zip file into any folder  like this. So let's extract it. Why extract?

  • 00:21:53 Because we need to get the workflows from here  and also we need the instructions TXT file. You

  • 00:21:58 see Massed Compute instructions. I will show  from beginning like you have never watched

  • 00:22:03 my previous Massed Compute tutorials. First of  all, read these instructions file. This includes

  • 00:22:09 everything that you need. I really recommend  you to follow this link to register Massed

  • 00:22:14 Compute. So just double-click. So register, log  in, then add some billing, set up your balance,

  • 00:22:21 then go to deploy. Massed Compute interface  slightly changed. Now our coupon works on

  • 00:22:27 every GPU and you can find that RTX 6000 Ada has  a best, as a best price-performance GPU. But if

  • 00:22:36 you want speed, you can use like A100 or H100.  Let's go with H100. So GPU H100, quantity one.

  • 00:22:44 This is super important. First of all, let's  select the creator as a category and image as

  • 00:22:48 an SECourses because my installers, my scripts  are made for this image. This image already has

  • 00:22:55 SwarmUI and other features, but we are going to  use manual installation of ComfyUI. So currently,

  • 00:23:00 this is 2.35 dollars per hour. Let's apply our  coupon, SECourses, verify. And now you see H100,

  • 00:23:09 which is one of the very best GPUs that you can  get is 1.76 dollars per hour. And this machine

  • 00:23:16 has over 1 terabyte disk space. Disk space is  also important when you download different models

  • 00:23:22 and install. So let's deploy. The slowest part of  the Massed Compute is the initial initialization,

  • 00:23:28 but after that, it is like 10 times, 20 times  faster than RunPod because their disks are very,

  • 00:23:35 very faster. So we just need to wait our  machine to be initialized. Meanwhile, if you

  • 00:23:40 don't have the ThinLinc client, if you didn't set  it previously, so click here. From here, download

  • 00:23:47 according to your machine. Whether you are using  MacBook, Apple, whether you are using Linux, you

  • 00:23:52 see it is supporting a lot of different operating  systems. I am Windows, so let's download this. So

  • 00:23:58 open it with clicking. It will ask you permission,  click yes, next, accept, next, default, install,

  • 00:24:06 run ThinLinc client. Then first of all, you  need to set your shared folder. We will use

  • 00:24:11 shared folder to transfer small files. So I will  go to local devices, clipboard synchronization,

  • 00:24:18 drives. Okay. So I will click this details and in  here, remove everything and click add, click this

  • 00:24:25 three dots icon, select a folder where you want  to synchronize. So I have a folder here named as

  • 00:24:32 Massed Compute shared folder. Okay, make it read  and write so I can both upload and download and

  • 00:24:38 okay. And these others are unchecked. You can also  make some speed optimizations. So everything is

  • 00:24:44 good. Okay. Now all I need is just waiting this  machine to be initialized. If you're wondering

  • 00:24:49 what is this end existing session does, it will  close all of the applications on the cloud machine

  • 00:24:56 and then connect. Use this only when you cannot  access the machine because it will terminate all

  • 00:25:02 the running applications. You will have the files,  but the running applications will get terminated.

  • 00:25:07 Okay, so the machine has been started. Let's open  the ThinLinc client again. Copy this IP address,

  • 00:25:14 paste here like this. Then copy username, paste  here. Then copy password and paste here. Then

  • 00:25:21 connect. Click continue. Wait for connection  to start. Click start. Do not wait here. It

  • 00:25:27 is starting the session and the Massed Compute  machine has been started. This is running on

  • 00:25:33 a remote server. Whatever I do here will be  made on a remote server, not on my computer.

  • 00:25:40 We already have SwarmUI, OneTrainer, Kohya, a  lot of stuff installed here. So you can right

  • 00:25:48 away start using SwarmUI. First of all, you should  update it with run stable SwarmUI update, then you

  • 00:25:54 can use Cloudflare stable SwarmUI. However, today  I'm going to show you how to install ComfyUI and

  • 00:25:59 use it. Remember, we had the zip file here, so  copy it, put it into synchronization folder that

  • 00:26:06 you did set, which is here, Massed Compute shared  folder in my case. You can set anything you want.

  • 00:26:13 Once you get this, just cancel, you don't need  it. Then enter home folder from here, scroll

  • 00:26:19 down and you will see Thin drives. This is your  synchronization folder with your computer. You see

  • 00:26:24 the folder appeared here in my computer. When I  enter inside it, I will see the files that I have

  • 00:26:31 on my computer. So do not install anything here.  Drag and drop it into downloads folder. Wait for

  • 00:26:38 copy operation to be completed. ThinLinc client is  extremely slow to copy big files, but you can copy

  • 00:26:45 small files like this one. Even this 12 megabyte  file took a lot of time. If you want to transfer

  • 00:26:51 big files, you can use Google Drive, you can use  OneDrive, Hugging Face. I also have a tutorial for

  • 00:26:57 that. Just reply me and I will link that to you.  So right-click and extract here. Then enter inside

  • 00:27:04 it. First of all, we will install the ComfyUI.  So you see Massed Compute instructions, read

  • 00:27:10 TXT file, double-click it. Then copy this part.  This is the installer. Return back to the folder,

  • 00:27:17 click this three dots icon here, open in terminal.  So pay attention that you are in the same folder

  • 00:27:22 where you have extracted. You see ComfyUI version  35, ComfyUI version 35. Right-click and paste and

  • 00:27:28 hit enter. It will start the installation of  ComfyUI. Just wait it until it is completed.

  • 00:27:35 Okay, so the ComfyUI installation  has been completed. As a next step,

  • 00:27:39 we will install the workflow nodes. So enter  inside Kijai MultiTalk and you will see that

  • 00:27:46 there is RunPod Massed Compute instructions  TXT file. Let's copy the command and start

  • 00:27:53 the terminal in the same folder here. This  is super important. Paste it and hit enter

  • 00:27:59 and it will install all of the necessary  nodes into the accurate folder. Okay,

  • 00:28:04 it is done. Now, final step, which is downloading  the necessary models. You see under this zip file,

  • 00:28:11 we have SwarmUI Installer and Unified Massive  Models Downloader. So open this post. The link

  • 00:28:17 will be also in the description of the video.  Download the latest version of the zip file. Then

  • 00:28:22 move this zip file into your shared folder. You  can also directly log into your Patreon account

  • 00:28:28 in this machine and also download directly  there, but it is a choice. So go to home,

  • 00:28:34 Thin drives, shared folder. Whatever I put here  will be also appearing on my computer. So drag

  • 00:28:42 and drop this into downloads folder. Wait for  copy to be completed. Okay, it is completed.

  • 00:28:47 Right-click and extract here. Then enter inside  the folder and you will see Massed Compute model

  • 00:28:53 download instructions TXT file. Double-click it.  All you need to do is just copy this entirely,

  • 00:28:59 open a terminal in the same folder, right-click  and paste and hit enter. It will install and start

  • 00:29:05 the application in a moment to download models.  Okay, the application has been started. First of

  • 00:29:11 all, we need to give the base download path. This  is super important. Where is the path? Go back

  • 00:29:17 to your downloads and ComfyUI installation.  Enter inside ComfyUI and enter inside models

  • 00:29:23 folder. Then control L, it will select the path,  control C, return back and paste it here. You see,

  • 00:29:31 home/ubuntu/downloads/comfyui_version_35/comfyui/models.  Wherever you have installed. Then delete

  • 00:29:38 this last part and select this ComfyUI folder  structure. Then click this ComfyUI bundles and

  • 00:29:44 drop down and go to the ComfyUI MultiTalk bundle  and download ComfyUI MultiTalk bundle. Then follow

  • 00:29:50 the terminal. It will download all the models  ultra fast. One of the big advantages of the

  • 00:29:57 Massed Compute is it is really, really fast.  Let's see the download speed. So we see 500

  • 00:30:02 megabytes per second, 1 gigabytes per second.  It is like 8 gigabits per second. So you see,

  • 00:30:10 10 gigabytes of file is downloaded in 20 seconds  only. It will be super fast. Don't worry.

  • 00:30:16 All right, so the models have been downloaded. Now  we are ready to begin. The rest is exactly same

  • 00:30:22 as in the Windows tutorial part. All we need to is  just start the ComfyUI. So return back to ComfyUI,

  • 00:30:30 Massed Compute instructions, read TXT file, and  we are going to start with this command. So open

  • 00:30:37 a terminal in this folder, pay attention to the  folders, open terminal, right-click and paste,

  • 00:30:43 and it will start the ComfyUI locally in the  Massed Compute. So you can use this inside

  • 00:30:49 Massed Compute as you wish, 100% private.  If you want to access it from your computer,

  • 00:30:56 we also have options like connecting by the IP or  with ngrok. I will show the connecting with IP. So

  • 00:31:05 let's see the starting. If you get some such error  messages like this one, it is not important. Okay,

  • 00:31:12 we are getting started. Yes, it started. So  click this link to open it in the browser and

  • 00:31:19 the ComfyUI started. Now I can drag and drop my  workflow to start using it. So enter inside the

  • 00:31:27 workflow and let's do a quick test with this one.  So drag and drop it and I need to load the file.

  • 00:31:34 This file has to be inside the Massed Compute  right now because it is running in the cloud

  • 00:31:39 machine. So I will just use the test image here  like this one and I will use this example file

  • 00:31:48 we have as well in the workflows in here, test  MP3. And that's it. Now I can generate it. All

  • 00:31:55 we need to do is just click run and it should  start running. Let's follow the terminal. Okay,

  • 00:32:01 it is downloading some of the files first,  that fee files it need to download from a

  • 00:32:06 remote repository. It is fine. Then it will  start processing. Okay, so the generation has

  • 00:32:11 been started. This workflow was optimized  for Windows low VRAM GPUs. Therefore,

  • 00:32:18 it is not working with the maximum efficiency  right now. To make it maximum efficiency, you

  • 00:32:25 need to set the number of block swap to zero for  this workflow. I have explained it all of this in

  • 00:32:32 the Windows tutorial part. So if you have watched  that part, it is fine. After setting that as zero,

  • 00:32:38 at your next generation, it will be faster. So  the video is getting generated. How you can access

  • 00:32:45 this from your computer, not inside from Massed  Compute? It is also easy. So you need to get

  • 00:32:51 your IP, then use this port, which is the default  starting port. How you're going to do that? Return

  • 00:32:58 back to your Massed Compute interface and copy  this URL. Then type the port to the end of it,

  • 00:33:06 which is 8188 and hit enter. Then you will be able  to access the ComfyUI from your computer as I am

  • 00:33:16 accessing right now. You see, it started. So this  is the way of it. Now I can use it as it is in my

  • 00:33:22 computer, but it will run in the remote machine  completely. However, from my browser, I will be

  • 00:33:28 able to download the generated video, I will be  able to upload images. Let me demonstrate you.

  • 00:33:34 So let's drag and drop the workflow, then choose  file to upload, upload whichever the file you want

  • 00:33:41 from your computer. This is in my computer. So  you see, from directly my computer, I can upload

  • 00:33:47 and I can generate. Let's set the block swap to  zero. By the way, let's see if the previous one

  • 00:33:53 was generated because we shouldn't run twice.  Yes, previous one was already generated. Then I

  • 00:33:59 will just run and it will generate the video. You  see, I am accessing it from my computer right now,

  • 00:34:06 not from Massed Compute, and I will be able  to download the generated video from here.

  • 00:34:12 To monitor the progress when you are running on  your computer, you need to look at the interface.

  • 00:34:19 You need to look at the CMD window, the terminal.  By the way, this long context workflow is way, way

  • 00:34:27 slower than the other workflow for some reason. So  therefore, when you need maximum quality, use the

  • 00:34:35 long context workflow, but if you want something  faster, use the other workflow. From that, what I

  • 00:34:41 mean is 10-second workflow is way faster compared  to the long context workflow. Okay, so the video

  • 00:34:48 has been generated. This was generated in Massed  Compute, but it is available in my browser. So

  • 00:34:54 how can I save it? I can right-click and save  preview and it will download into my computer

  • 00:35:00 as you are seeing right now. Or you can transfer  all the files from the Massed Compute. How? Go to

  • 00:35:07 the ComfyUI installation like this, go to ComfyUI,  go to outputs here, then you see the entire files

  • 00:35:15 are here. Copy the whichever one you want. So  let's copy the entire folder like copy, home,

  • 00:35:22 Thin drives, and enter inside shared folder and  paste it. This pasting may take time because it

  • 00:35:30 will synchronize it with your computer. Depending  on how many files you have, it may take time. So

  • 00:35:36 what you can do is like you can log into your  OneDrive, Google Drive, upload it there, upload

  • 00:35:42 to any platform and download from your computer.  However, for small files, it is fine. When I go

  • 00:35:48 back to my computer, you see the files already  arrived into my computer. There is one another

  • 00:35:53 thing that I want to show before I end the Massed  Compute part. So right-click and quit all windows.

  • 00:36:00 So I will close everything. Then control alt D, it  will minimize all the open windows. First of all,

  • 00:36:08 I will update the SwarmUI because I will show you  how you can use this manually installed ComfyUI

  • 00:36:14 with the SwarmUI we have in the Massed Compute.  This is pre-installed. It is coming with our SE

  • 00:36:20 courses image. So it updated the SwarmUI to the  latest version and started. Then go to server,

  • 00:36:26 go to backends, and you see this is the backend  it is using. Now I am going to give my manually

  • 00:36:33 installed ComfyUI backend. So it is inside  downloads, inside the ComfyUI, ComfyUI here,

  • 00:36:40 control L, it will select the path, control C,  and click this green, paste it like this. You see,

  • 00:36:47 starting from home, okay, like this.  /home/ubuntu/downloads/comfyui/comfyui

  • 00:36:52 and main.py, then save. So instead of the  backend it comes up with, it will use my

  • 00:36:59 installed backend. So now you need to just wait  it to restart the backend and load everything. It

  • 00:37:05 will install the missing necessary libraries  at the first time. So let's just wait. Okay,

  • 00:37:10 it is loaded. Now it is using the backend that  I have pre-installed. So you can use it locally

  • 00:37:16 right now. If you want to access it from your  computer, it is so easy. Close this window,

  • 00:37:21 control alt D, and use run Cloudflare stable  SwarmUI. This will generate a secure URL. So it is

  • 00:37:27 here. Open it with control clicking, and now I can  access this link from my computer, from my phone,

  • 00:37:35 from my tablet, wherever I want. We also have some  pre-downloaded models like SDXL base model. So

  • 00:37:41 let's just type a fast car and see the generation.  To see the logs with the SwarmUI, always server,

  • 00:37:48 logs, and debug menu. It is right now using the  ComfyUI backend that we have installed. You can

  • 00:37:55 see here. It is the latest version with the  latest libraries. So this is how you can use

  • 00:38:00 ComfyUI on Massed Compute, also with SwarmUI, and  it is generated. It is working perfectly fine.

  • 00:38:06 One more thing is that if you stop your machine,  it will not stop billing. So after I stopped it,

  • 00:38:13 it doesn't matter. I have to delete this instance  on Massed Compute so that it will not use my

  • 00:38:19 credits anymore. So let's end this to delete this  Massed Compute part and I will begin RunPod. Okay,

  • 00:38:27 it is gone. Everything is gone. So make sure that  you did back up your data before you delete it.

  • 00:38:33 Okay, now we are going to see how to use  this model, MultiTalk on RunPod. You need

  • 00:38:40 to watch Windows tutorial part at the very  least. So download the latest zip file. You

  • 00:38:46 see that now it is version 36. It was 35 in  the previous two parts. I made an update. I

  • 00:38:53 added --use-sage-attention to the prompts and  fixed an error in the one particular workflow.

  • 00:39:01 So move this zip file into any folder and extract  it. So extract all like this. Then double-click

  • 00:39:08 and open RunPod instructions read TXT file. This  is important. First of all, you need to register

  • 00:39:15 RunPod. Please use this link. I appreciate that.  After registration and sign in, put some credits

  • 00:39:22 into your account from billing like from here.  Then go to the pods and we are going to deploy.

  • 00:39:29 This is the newest interface. So click deploy. You  can use pretty much any GPU. However, if you want

  • 00:39:36 some speed, you can use RTX Pro 6000. This is a  really good GPU with huge amount of VRAM. First

  • 00:39:43 of all, we need to select our template. So select  the template Torch 2.2.0. So type Torch here,

  • 00:39:51 then select the Torch 2.2.0. Then you need to  edit template from here. Make the volume disk

  • 00:39:59 200 GB if you are going to download more models,  make it bigger, and you need to add a port here,

  • 00:40:06 which is 3001. We will use this port to connect  ComfyUI. Set overrides, deploy on demand.

  • 00:40:14 RunPod is extremely slow at the model loadings,  installation. It shows you some huge speed here,

  • 00:40:22 but you never get it because it is shared with  so many other people. So it is totally up to your

  • 00:40:28 chances. The machines will start initially quickly  because we are using official PyTorch templates.

  • 00:40:34 So once you see it is initialized like this,  click connect and click the Jupyter Lab. Even if

  • 00:40:39 it is not ready, click it and it will open in most  cases like this one. Then click this arrow icon,

  • 00:40:47 go back where you have downloaded the zip file and  upload it. In the bottom of screen, you see it is

  • 00:40:52 uploading. So wait for upload to be completed. You  can also use RunPod CTL. I explain it in another

  • 00:40:58 tutorial video. RunPod CTL is amazing to transfer  big files between RunPod and your computer. Okay,

  • 00:41:05 then right-click and extract archive. After that,  click this refresh icon and everything extracted.

  • 00:41:11 You see RunPod instructions read TXT file. Just  copy this command. This is for installation. Plus

  • 00:41:18 icon, new terminal, paste it and hit enter.  Now you need to wait for installation to be

  • 00:41:23 completed. The installation may take 10 minutes,  20 minutes, 30 minutes, depending on the pod you

  • 00:41:29 got. The pod could be broken. So there are  so many cases which can happen. Therefore,

  • 00:41:34 I am recommending Massed Compute. The installation  is not even started yet. We are still waiting for

  • 00:41:40 some reason. The pods are terrible in most cases,  unfortunately. The GPUs are good, but RunPod is

  • 00:41:47 losing from the disks and the network speed they  have. Okay, the installation started. If you see

  • 00:41:54 the installation is too slow, just terminate the  machine and get a new machine. I'm just waiting

  • 00:41:59 for installation to see the speed, but so far it  is very slow. We can also deploy another machine

  • 00:42:05 and we can make some filters from here like  NVMe disk and we can make the region like US

  • 00:42:13 like this one. There is no Pro 6000 here. So we  can make it like Texas and there is no. These are

  • 00:42:21 usually better, but none of them are available.  Okay, the installation speed is looking decent,

  • 00:42:26 so we don't need a new machine. Meanwhile, it is  installing, we can start downloading models since

  • 00:42:32 it will take time on RunPod. For downloading  models, we are going to use SwarmUI Installer

  • 00:42:38 and Unified Massive Downloads here. So click  this link, download the latest zip file from

  • 00:42:43 here. I really recommend you to read these posts  because they include a lot of information. Then

  • 00:42:50 we will click this upload files, upload the  zip file, wait for it to be uploaded, and the

  • 00:42:55 zip file has been uploaded here. You see the zip  file is here. Right-click, let's fix the name,

  • 00:43:00 and you see the zip file has been uploaded. It is  here. Right-click and extract archive. Then click

  • 00:43:06 refresh and we can see that there is RunPod model  download instructions TXT file. Double-click it,

  • 00:43:12 copy this, and open a new terminal and paste it.  We are still installing the ComfyUI and meanwhile,

  • 00:43:19 we will download the necessary MultiTalk models,  1 2.1, some other models that it needs. Everything

  • 00:43:26 will be automatically downloaded. Wait for Gradio  live share here. Okay, it is here. Click this,

  • 00:43:32 open the Gradio live link. This is all  running on RunPod, on a cloud machine,

  • 00:43:37 not on my PC. Okay, interface started. First of  all, we need to get the models folder path. It is

  • 00:43:43 inside ComfyUI and it is models. So right-click  and copy path. Then go to path here, delete it,

  • 00:43:50 paste. You see it is missing a backslash at  the beginning, so put it like this. This is

  • 00:43:56 super important. Then select this checkbox,  go to ComfyUI bundles and download ComfyUI

  • 00:44:02 MultiTalk bundle. So it will download everything.  Let's see the download speed. It is looking

  • 00:44:07 decent right now. Okay, slow, getting faster.  Yeah, decent speed. And it's also installing

  • 00:44:14 the ComfyUI. All I need to do is right now  just wait for installation to be completed.

  • 00:44:19 Okay, the model downloads have been completed.  Now all we need to do is just wait for ComfyUI

  • 00:44:26 installation to be completed. One advantage of  the RunPod is that you can turn off the machine,

  • 00:44:33 it will use a very little amount of money  from your account, then when you restart it,

  • 00:44:40 when you start it again, all you need to do  is just starting command and it will right

  • 00:44:44 away work. So you will not reinstall it.  This is the advantage of RunPod. You can

  • 00:44:50 also use RunPod permanent storage. I also have a  RunPod permanent storage tutorial in the readme

  • 00:44:56 file. You can see here, it is so easy. So that  these are the advantages of the RunPod compared

  • 00:45:03 to Massed Compute. The permanent storage or  turning off the machine and starting again.

  • 00:45:09 When you turn off the machine, it uses very  little amount of money. Then after starting it,

  • 00:45:14 you just run the starting command. Okay, we are  still waiting for installation to be completed.

  • 00:45:20 Okay, so the installation has been completed. As  a next step, we need to install the workflow. As

  • 00:45:26 in the other videos, so go to the workflows  and go to the Kijai MultiTalk and you will

  • 00:45:32 see RunPod Massed Compute instructions. Just  copy this and open a new terminal inside here,

  • 00:45:39 new terminal. You see the path is here. Copy,  paste it, and it will install the necessary

  • 00:45:46 nodes we need to run the MultiTalk. This part is  the fastest part. Okay, it is done. Now we are

  • 00:45:52 ready to start. Go back to ComfyUI from here.  You see I'm using this navigation, workspace,

  • 00:45:59 then open the RunPod instructions read TXT file.  This has command to start the ComfyUI. As I said,

  • 00:46:07 if you restart the machine, if you stop and  start, you will run this. Terminal, I am inside

  • 00:46:13 this path and paste it. Then this will start the  ComfyUI with the necessary models and the nodes

  • 00:46:20 and the workflow. We have everything ready.  And you will notice that it is using the Sage

  • 00:46:26 Attention. So my installer is extremely advanced.  It is installing the pre-compiled Sage Attention,

  • 00:46:33 Flash Attention, xformers, the very best libraries  that you need to obtain the maximum performance

  • 00:46:39 and the speed. It is still starting. The first  time start may be slow. You will also get this

  • 00:46:44 error. This is not important. You can just  ignore it and it also fixed it itself. The

  • 00:46:50 model loadings, the start times on the RunPod  is just terrible due to their disk speed. Okay,

  • 00:46:56 we are seeing the Blackbell RTX Pro 6000  GPU with 96 GB of VRAM. You can see the

  • 00:47:04 total VRAM and the RAM of the machine here.  We can see that it is using Sage Attention,

  • 00:47:09 so it will be the fastest attention. Okay, we are  still waiting for start. And when you see this, it

  • 00:47:15 means that it is started on the 3000 port. How we  are going to connect? Go back to your My Pods and

  • 00:47:21 click this arrow icon and you see there is connect  to port. Connect to port and connect by 3001 port.

  • 00:47:30 And the ComfyUI is starting. This is running in  the RunPod machine. So open a new terminal. We

  • 00:47:37 want to watch the what is happening. pip install  nvitop, then just type nvitop like this. Okay,

  • 00:47:44 the machine is ready and set and the ComfyUI  also started. So I'm going to drag and drop the

  • 00:47:50 workflow from my extracted folder. Let's use the  best performance workflow, which is here. You see

  • 00:47:59 720p long context high quality, drag and drop it.  The rest of the usage is exactly same as in the

  • 00:48:06 Windows tutorial part. So let's upload our demo  images. This is the demo image. Let's upload our

  • 00:48:13 test file. This is around 250 frames. So after  uploading your image and audio, first of all,

  • 00:48:21 set your width and height. Since this is a tall  image, this is 720p and 1280p. It will auto

  • 00:48:29 crop and resize it. This is 250 frames. So the  parameters that you need to change is input image,

  • 00:48:37 image size, input audio, number of frames. This  is depending on your input audio length. 25 frames

  • 00:48:45 means that it is one second of audio. Then the  prompt. Nothing else you just need to change it.

  • 00:48:52 So this is a man patiently singing like this. So  I did set the prompt as well and I am pretty much

  • 00:49:00 ready. All I need to also need to set is the  blocks to swap. Currently, it is set to zero

  • 00:49:07 because this machine has huge amount of VRAM. If  the VRAM was low, you could set it as 40 and it

  • 00:49:15 would run it on the CPU RAM. However, I am going  to use as zero. This is a big machine. Moreover,

  • 00:49:22 if you get out of VRAM error, you can enable all  the tiling, tiled VAE, enable it if you get out of

  • 00:49:29 VRAM. Then you will see enable VAE tiling, enable  it if you get out of VRAM, and you will see enable

  • 00:49:36 VAE tiling also here. Okay, I don't need to change  them on this machine and I did set the block swap

  • 00:49:43 to zero and then I need to hit run here. You  see it is at the bottom, this one. Then let's

  • 00:49:49 return back to machine and see what is happening.  This is the GPU, it is loading the model. We can

  • 00:49:55 also watch the status on the terminal where we  started the ComfyUI. So between these two, you

  • 00:50:00 need to check it out. It will show us everything.  So you see it shows the total route duration of

  • 00:50:05 the audio is 9.96 seconds. You need to get the  audio second yourself. You need to calculate it,

  • 00:50:12 unfortunately, to set the number of frames here.  It is not automatically set. If you know how to

  • 00:50:19 make it automatically set according to the number  of seconds from here, I can update the workflows

  • 00:50:25 later, but currently, we need to manually  calculate the duration and therefore, the

  • 00:50:31 calculate the number of frames we need. Okay, it  is loading the models. Let's see what will happen.

  • 00:50:36 The initial loading of the models may take  huge time on RunPod. So you just need to wait

  • 00:50:41 patiently. Moreover, you can follow always NVI  top screen to see what is happening. You see,

  • 00:50:48 I can see that the memory is increasing the RAM  of the machine. Therefore, I can say that it is

  • 00:50:55 loading model right now. Moreover, it is using  CPU, therefore it is doing some operations.

  • 00:50:59 Probably setting up the precision of the model  while loading. And I can see that now it is

  • 00:51:05 loading the transformer parameters to CPU, then it  will move them into the GPU. Why it is happening?

  • 00:51:11 Because we are using block swapping. You can  right-click and bypass this if you want, if

  • 00:51:16 you want to avoid that on a big VRAM machine GPU  like this one. It also shows you other parameters.

  • 00:51:22 You can always read here to understand and the  processing is starting. Let's see how much VRAM

  • 00:51:28 it will use. Okay, it is using 35 GB of VRAM right  now. So this machine was an overkill for this,

  • 00:51:37 but we will get maximum speed. Okay, it is using  entire 600 watts. This is amazing. This means

  • 00:51:44 that we are fully utilizing the GPU and it is  processing as chunks. Meanwhile, it is changing

  • 00:51:51 between chunks, you will see drop of the watt  usage. So it processed the zero between 81 chunks,

  • 00:51:58 then it is processing 52 between 181 because this  model is normally able to process 81 chunks at

  • 00:52:05 once. Therefore, it is splitting and it's also  have some inter-crossing of the chunk so that

  • 00:52:12 it will be consistent. And this is the first  step. It will do 10 steps total because this

  • 00:52:17 is high quality processing and we did set 250  frames. You can also see this as the frames.

  • 00:52:24 Okay, so it has been generated  and let's see the result.

  • 00:52:29 Through the maze where the data play. Amazing result as in the previous part

  • 00:52:34 of the tutorial. So how you can download  this? Right-click and save preview and it

  • 00:52:39 will download it into your computer. Moreover,  you can go to your ComfyUI in workspace, output,

  • 00:52:45 and you can right-click and download from here  like this, download, download, or if you want to

  • 00:52:51 download entire generations, go to the ComfyUI and  right-click output and download as an archive. It

  • 00:52:58 will zip the entire folder and it will let you  download all the files like this. So what else

  • 00:53:04 left? Let me show you how to stop and start this  machine another time. So currently, I am using

  • 00:53:12 2.79 dollars per hour. Let's stop this. This will  only cost me 0.56 dollar per hour. So it will cost

  • 00:53:21 me only 1.33 dollar per day if I stop this pod.  Then I can resume it whenever I want. So I will

  • 00:53:29 just click start and there must be a GPU, so start  it. Let's wait for it to start. So this is the

  • 00:53:36 biggest advantage of RunPod. As I said, you don't  need to reinstall everything from scratch. We will

  • 00:53:42 be able to just start and use the application.  However, if there were no GPU at the time when

  • 00:53:49 I am starting again, I wouldn't be able to use.  And to prevent that, you can use the storage

  • 00:53:55 and network volume, permanent volume storage  from here. Okay, the machine started. Let's

  • 00:54:01 connect from the Jupyter Lab. I will just open the  RunPod instructions here and I will just copy this

  • 00:54:09 starting command from here. So new terminal, wait  for terminal to start. Sometimes the terminal may

  • 00:54:15 not open. This is because of the browser. You can  restart your browser to fix. Okay, copy, paste it.

  • 00:54:21 Now it will start the application and I will be  directly use it right away. Okay, it is starting.

  • 00:54:28 You see the restart is pretty fast compared to the  of course installation. Moreover, pay attention to

  • 00:54:34 how much volume you are using. We are using like  50% of our volume. If we get out of disk space,

  • 00:54:40 we have to stop and expand the storage. Okay, we  need to see the local host URL started here. It

  • 00:54:47 is still not yet. Almost there. It is starting.  If you want to update your ComfyUI, then you just

  • 00:54:54 need to run the installation command again. It  will update it very quickly without installing

  • 00:55:01 entirely. Okay, the application almost started,  still not. We are waiting for URL to appear here.

  • 00:55:08 Okay, here. So it started. Now I will go back to  my pods, connect, HTTP servers 3001. It's still

  • 00:55:16 not ready yet, probably. Yes, now it's ready and  the ComfyUI will start like before. So this is

  • 00:55:23 the way of using RunPod. Yes, it started. It is  keeping everything as it is. Now I can stop this

  • 00:55:30 so it will not charge me a lot of money. If  you want to prevent any money usage, then you

  • 00:55:37 need to click terminate and it will terminate the  machine entirely. So you will be not spending any

  • 00:55:44 money on it. That's it. I hope you have enjoyed.  Please reply this video, share it, comment. You

  • 00:55:50 can ask me any questions you want. Hopefully,  see you in future amazing tutorial video.

Clone this wiki locally