Skip to content

LTX 2 and Z Image Base Full Tutorial Audio to Video Lip Sync ComfyUI SwarmUI Windows Cloud

Jump to bottom
FurkanGozukara edited this page Jan 31, 2026 · 1 revision

LTX 2 & Z Image Base Full Tutorial + Audio to Video Lip Sync + ComfyUI + SwarmUI + Windows + Cloud

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

LTX 2 & Z Image Base Full Tutorial + Audio to Video Lip Sync + ComfyUI + SwarmUI + Windows + Cloud

image Hits Patreon BuyMeACoffee Furkan Gözükara Medium Codio Furkan Gözükara Medium

YouTube Channel Furkan Gözükara LinkedIn Udemy Twitter Follow Furkan Gözükara

LTX 2 is the newest state of the art (SOTA) Open Source video generation model and tutorial will show you how to use it with very best and most performant way in ComfyUI and also in SwarmUI. Moreover, Z Image Base model published and I will show how to use Z Image Base with most amazing preset and workflow as well. Furthermore, this tutorial will show you how to install, update, setup, download ComfyUI and SwarmUI and models and presets and workflows both on Windows and on RunPod, Massed Compute and SimplePod. Linux users can use Massed Compute scripts and installers directly. This is a masterpiece entire lecture level complete tutorial. This video will kickstart your AI journey 100x. Both local Windows and Cloud.

📂 Resources & Links:

🤖 ComfyUI Installer and Presets Zip File With CUDA 13: [ https://www.patreon.com/posts/ComfyUI-Installers-105023709 ]

💻 SwarmUI Installer and Presets Zip File: [ https://www.patreon.com/posts/SwarmUI-Install-Presets-114517862 ]

🚀 Model Downloader Zip File: [ https://www.patreon.com/posts/Model-Downloader-114517862 ]

🚆 SECourses Musubi Trainer (Model Quantize and Train App): [ https://www.patreon.com/posts/SECourses-Musubi-Trainer-137551634 ]

🛠️ Image Comparison Slider Tool: [ https://www.patreon.com/posts/image-video-comparison-slider-app-133935178 ]

👋 SECourses Discord Channel for 7/24 Support: [ https://bit.ly/SECoursesDiscord ]

ℹ️ SimplePod Register : https://simplepod.ai/ref?user=secourses

ℹ️ SimplePod Template : https://dash.simplepod.ai/account/explore/100/ref-secourses/

ℹ️ RunPod Register : https://get.runpod.io/955rkuppqv4h

ℹ️ RunPod Template : https://get.runpod.io/SECourses_CU13

ℹ️ Massed Compute Register : https://bit.ly/SECoursesMassedCompute

⏱️ TIMESTAMPS

00:00:00 Intro: ComfyUI + SwarmUI presets, Z-Image, model downloader, cloud installs

00:00:28 Free prompt enhancement with Google AI Studio (prepared prompt file)

00:01:26 Demo: 45s audio-driven lip-sync image→video (LTX 2)

00:02:25 Quick demos: ComfyUI image→video, text→video, Z-Image base

00:03:43 Quick demos: SwarmUI LTX 2 image→video + Z-Image base

00:04:46 Install/update presets zip (v78+): extract & overwrite everything

00:05:11 Upgrade to CUDA 13 safely: delete venv then run ComfyUI update/install

00:06:00 Windows prereqs + the always-updated setup guide referenced in description

00:06:42 Install required node bundles: (1) SwarmUI extra nodes + (100) LTX audio

00:07:29 VRAM-optimized launcher: no-VRAM / cache / smart-memory / precision choices

00:09:36 Share one model library: configure extra_model_paths.yaml (no duplicates)

00:10:25 Model Downloader overview: set base path + one-click bundles for SwarmUI/ComfyUI

00:11:43 Download LTX 2 core bundle: multi-connection download + merge + hash verify

00:12:12 Low-VRAM path: GGUF distilled models vs recommended FP8-scaled defaults

00:14:16 URL Downloader: Civitai/HF links, folder targeting, and optional API keys

00:14:54 ComfyUI preset pack tour: LTX2 (I2V/T2V/audio) + Z-Image (base/2× upscale)

00:15:16 ComfyUI audio lip-sync preset: image + resolution + audio setup

00:16:58 Frames & prompting: 24fps math, run/stop frame count, lyrics/subtitles prompts

00:17:59 Quality/perf knobs: CRF, VRAM monitoring, and low-VRAM args recap

00:20:22 Review result + move to LTX 2 image→video preset workflow

00:21:50 Prompt enhancer workflow: drag prompt file into AI Studio (optionally add image)

00:24:11 Z-Image troubleshooting: disable Sage Attention and restart

00:24:32 Z-Image base + 2× upscale preset: when to use it and what to expect

00:26:47 Outputs & reproducibility: where renders save + drag PNG to reload metadata

00:27:57 Update SwarmUI via zip: get latest presets + utilities

00:28:27 SwarmUI setup: ComfyUI backend, passing args, and pointing to model folders

00:30:21 SwarmUI image→video: direct apply + init image workflow

00:31:17 Fix model load/caching issues: add --use-cache-none when needed

00:32:51 SwarmUI text→video + upscale: duration/frames, half-res then upscale rules

00:33:52 SwarmUI outputs: output_local/raw + metadata saved with generations

00:35:05 SwarmUI Z-Image: base vs 2× upscale comparison + speed notes

00:36:58 Image comparison slider tool: quick before/after inspection

00:37:51 RunPod start: template choice, CUDA/driver constraints, optional storage volume

00:43:30 RunPod Jupyter: upload ComfyUI zip, extract, install bundles (1+100)

00:46:28 RunPod: run Model Downloader, start ComfyUI, connect via exposed port

00:52:05 RunPod: switch to SwarmUI, map folders (case-sensitive), import presets

00:55:43 RunPod: download outputs as archive + stop vs terminate cost control

00:57:55 SimplePod: cheaper/faster alternative + persistent volume setup highlights

01:03:52 Massed Compute: deploy with coupon + connect via ThinLinc (shared folder setup)

01:08:17 Massed Compute: install bundles + download models (disk speed advantage)

01:10:48 Massed Compute: start ComfyUI, connect from PC, run lip-sync preset

01:14:26 Massed Compute: hook SwarmUI to ComfyUI backend + text→video demo

01:17:02 Wrap-up: recap providers + next steps

#comfyui #comfyuitutorial #runpod

Video Transcription

  • 00:00:00 Greetings everyone. Today I am going to show  you LTX 2 workflows, presets in ComfyUI. I

  • 00:00:07 have prepared all the configurations, one  click to install and use. Moreover, I will

  • 00:00:12 show Z-Image base model in ComfyUI. Moreover, I  will show you how to use the LTX 2 in our famous

  • 00:00:21 SwarmUI with presets, Z-Image base in SwarmUI  with presets. Furthermore, I will show you how

  • 00:00:28 to improve your prompts in Google AI Studio for  free with our prepared commands file. Moreover,

  • 00:00:36 downloading these preset models will be  one click with our interface updated model

  • 00:00:42 downloader. This is independent from both ComfyUI  and SwarmUI. It has both Z-Image and LTX 2 models.

  • 00:00:51 I also have prepared Z-Image Quant FP8 scaled  model for you and also working on NVFP4 models

  • 00:00:59 and NVFP4 generator. This application will have  everything SECourses Musubi Trainer. It will

  • 00:01:06 have all the options to quantize your models.  Also, I will make fresh installation on RunPod,

  • 00:01:13 Massed Compute, and SimplePod so that we will  refresh our memory to how to install and use

  • 00:01:20 ComfyUI and SwarmUI on cloud platforms in  addition to your local Windows computer.

  • 00:01:26 I am going to show you an amazing workflow  in ComfyUI which you can provide your audio,

  • 00:01:32 image, prompt and generate amazing audio driven

  • 00:01:35 video. It has lip sync and it has  animation. Let me demonstrate you.

  • 00:01:40 SECourses. Let's go. Late night screen glow. Ideas  on fire. Turning what if into something higher.

  • 00:01:48 From zero to flow. We learn it fast. Build it  ship it make the future last. New tools new

  • 00:01:53 rules we break the doubt. Take that prompt watch  the AI route. If you want the shortcuts done the

  • 00:01:57 right way. Hit play let's level up today. Hands  on the keyboard mind on the dream. Watch the

  • 00:02:02 model run like a laser beam. No fear no fluff just  clarity. One more step to your next victory. SECourses

  • 00:02:08 AI on the rise. Lighting up the dark  with electric skies. From prompts to pixels

  • 00:02:13 code to sound. We make it real we run this ground.  That's courses learn build repeat. Turn big ideas

  • 00:02:19 into something elite. If you want the best place  to start. That's courses straight to the heart.

  • 00:02:25 As you have seen it was amazing video. It was  45 seconds. After like 30 seconds 35 seconds

  • 00:02:32 it starts to degrade but this is what LTX 2  can do. This will be literally one click to

  • 00:02:39 install and use. I will show that. So here  another example in ComfyUI. This is image

  • 00:02:43 to video workflow. It is also one click to use  and I will show in SwarmUI as well. Let's see.

  • 00:02:50 This is the tutorial that you  have been waiting for so long.

  • 00:02:53 I mean amazing. You see it did a very  nice emotion at the end. The LTX video

  • 00:02:59 can generate up to 20 seconds very  good. It can generate up to 1 minute

  • 00:03:03 but after 20 seconds it starts to  degrade. Let's see text to video.

  • 00:03:13 You took a wrong turn partner.

  • 00:03:18 By the way these were quick demos that I  did. They are not perfect. They are not

  • 00:03:22 cherry picked. And this is Z-Image workflow.  This is an amazing quality. So many people

  • 00:03:29 is not using accurately and as you see this is  what kind of quality Z-Image base can generate

  • 00:03:35 in ComfyUI. All is preset I will show. Now let  me show you the examples on the SwarmUI. This

  • 00:03:43 is image to video in SwarmUI. It is so easy to  use with our preset. I will show how to use.

  • 00:03:49 This is the tutorial that you  have been waiting for so long.

  • 00:03:58 As you have seen it is amazing.  These are quick demos so you can

  • 00:04:02 generate even much better videos. So here  another example in SwarmUI. This is video.

  • 00:04:12 Faith falls! Witness me!

  • 00:04:17 With video upscalers they would become much better  or with more tries or improving the text prompt.

  • 00:04:24 These are quick demos. And we have the Z-Image  base preset in SwarmUI as well. It is working

  • 00:04:30 amazing. Amazing quality, amazing realism and  it is following the prompts very well very good.

  • 00:04:36 So you should begin with upgrading your ComfyUI  if you haven't yet or download the latest zip

  • 00:04:41 file to obtain the all the presets that we  have. So the link will be in the description

  • 00:04:46 of the video. Download the latest zip file  version 78. It can be bigger version when

  • 00:04:51 you are watching this tutorial. Then copy  the downloaded zip file. Go to your ComfyUI

  • 00:04:56 installation or into a fresh folder. Both works.  Then you need to extract archive and overwrite all

  • 00:05:04 the files. This is really important. Overwrite  all. Then if you didn't upgrade to CUDA 13 yet,

  • 00:05:11 enter inside ComfyUI, delete your virtual  environment folder like this. This is proper

  • 00:05:17 way of upgrading CUDA 13 if you haven't yet.  Since my ComfyUI is open it is preventing me

  • 00:05:23 to delete the virtual environment folder. So  let me delete all of them. Try again. Deleting

  • 00:05:29 virtual environment will not cause you to lose  any data. Then install or update ComfyUI run.

  • 00:05:36 This is not mandatory every time. This is  mandatory when you are upgrading from CUDA

  • 00:05:40 12.9. So I am using Python 3.10. You need to have  this installed in your computer. And this is the

  • 00:05:47 version that I recommend. This is same for fresh  installation and upgrade. We use the same file.

  • 00:05:54 If you are first time into generative AI you  really should follow Windows requirements

  • 00:06:00 tutorial. When you scroll down a little bit  you will see this. We have a up to date video

  • 00:06:06 here so watch it and we have an up to date  post which you will follow while you are

  • 00:06:12 watching this tutorial video. It has all the  links instructions. It is fully up to date.

  • 00:06:17 Last time updated is 8 January 2026. I  am using exactly as in here so this is

  • 00:06:26 identical to my setup so therefore when  you watch this tutorial all of the AI

  • 00:06:31 applications will work on your PC perfectly  fine on your computer perfectly fine.

  • 00:06:36 So the ComfyUI installation and update has been  completed. To use the newest presets you need to

  • 00:06:42 use the Windows custom nodes bundles installer.  Double click it. This is a unified bundles

  • 00:06:49 installer that I have made for you. So you need  to install option 1 SwarmUI extra nodes to use

  • 00:06:55 the presets. Plus you need to install bundle 100  LTX audio to video bundle. So when you install

  • 00:07:02 these two you will be ready. So 1 and comma and  100 and hit enter. It will ask you to proceed.

  • 00:07:09 You can hit Y or enter and it will install the  necessary bundles and update them. So this is for

  • 00:07:16 both update and installation. This covers both  of them. You see it is updating my bundles as

  • 00:07:22 well and it is all getting updated. Okay all the  nodes have been installed or updated properly.

  • 00:07:29 Then I also have added  windows_run_vram_optimized.bat

  • 00:07:34 file. When you run this it will ask you to start  your installation with specific configuration.

  • 00:07:40 So if you are getting out of VRAM options  you can do this. You can start with no VRAM

  • 00:07:46 mode. When you do this you can even generate 1  minute videos on LTX 2. This will use very very

  • 00:07:52 minimal amount of VRAM memory. Actually let me  demonstrate you. So you see currently I am using

  • 00:07:57 2 megabytes of VRAM. So I will choose the option  5. And if you are also low on RAM memory you can

  • 00:08:04 start with no cache. So it is option 3. You can  also disable smart memory. So let's disable as

  • 00:08:12 well. And it will be GPU default or you can  also choose VAE on CPU. I will make it GPU.

  • 00:08:19 I am going to use Sage Attention. You shouldn't  use Sage Attention for Z-Image base model at the

  • 00:08:24 moment but for all other models you should  use it. Currently Z-Image base model is not

  • 00:08:29 working with Sage Attention. And the precision  I am going to just make it auto. And for this

  • 00:08:34 option just hit yes. Additional optimizations I  am just going to leave it auto. Disable pinned

  • 00:08:41 memory no. And it is all done. So you see it  shows you the configuration options. You can

  • 00:08:48 do this if you don't want to pick these every  time. Copy this windows run GPU and paste. So

  • 00:08:55 you see I have copied version. Right click edit in  Notepad++. So you can append these options to here

  • 00:09:04 like this. You see or like this. So this way you  can customize your starting bat file and use it

  • 00:09:13 to start directly with all these optimizations  you want. This is also how you remove the sage

  • 00:09:21 attention from run GPU here. So you can also  edit this file and remove the Sage Attention

  • 00:09:30 from here to start without Sage Attention. This  is how you customize your starting with ComfyUI.

  • 00:09:36 Moreover to use the SwarmUI downloaded models you  need to do this. Copy extra model paths into your

  • 00:09:43 ComfyUI folder. Then right click and edit it.  Then select the models base path. This is where

  • 00:09:51 I have downloaded all of my models so therefore  when I give this it sees all the models inside

  • 00:09:57 my SwarmUI installation folder or wherever I  want. You can add multiple base paths to here.

  • 00:10:04 Okay so the ComfyUI started with the minimal  amount of VRAM memory usage. So let me show you

  • 00:10:11 45 seconds audio to video generation. I have the  preset here. How you are going to use them? You

  • 00:10:18 need to first download the models. To download the  models we are going to use the model downloader

  • 00:10:25 which is included in SwarmUI model downloader  zip file. This is also how we are going to

  • 00:10:29 install SwarmUI updated and get the presets.  So let's download this. The link will be in

  • 00:10:34 the description of the video. So move the SwarmUI  model downloader into your SwarmUI installation

  • 00:10:40 or into any drive make a fresh installation.  Right click. Then you need to extract here and

  • 00:10:46 overwrite all the files. This is super important.  Overwrite all the files so that it will update the

  • 00:10:53 utilities and everything. Then double click  Windows start download models application

  • 00:10:57 file. Independent from SwarmUI. You don't need to  install SwarmUI for this. This supports ComfyUI

  • 00:11:05 as well. So this is our new interface. This is  where you give the base model path. So you see

  • 00:11:11 currently it is seeing automatically my SwarmUI  models. You can change this into your ComfyUI

  • 00:11:17 models as well. So enter inside here. Go to  the models and copy the path. Copy paste it

  • 00:11:24 here. You can also remember settings. And ComfyUI  folder structure. Then you need to download the

  • 00:11:30 LTX 2 video core bundle or Z-Image core bundle  if you want to use them. But I am using SwarmUI

  • 00:11:38 mainly therefore I am downloading models there.  So refresh and it will return back to original

  • 00:11:43 since I didn't remember settings. Then I need  to click download all models of the LTX 2 core

  • 00:11:51 bundle. Then it will start downloading with 16  connections. You can also drag and drop this

  • 00:11:56 like this to expand or make it small or click  here. And it will start the downloads. It will

  • 00:12:01 do hash verification. It will download all the  models that I have prepared for you. You can

  • 00:12:06 also see their names here. If you have a very low  VRAM GPU alternatively go to the video generation

  • 00:12:12 models. Go to the LTX 2 video models and you  can download the LTX 2 distilled GGUF model.

  • 00:12:22 Which is here. So you can download GGUF Q4 model  if you are on very low RAM and VRAM. This is not

  • 00:12:30 recommended. Normally I use FP8 scaled. This is  also what the bundle downloads. And additionally

  • 00:12:35 you can also download FP4 mixed text encoder but  by default I am using this one. This is mixed FP8

  • 00:12:43 scaled. The name is like this because SwarmUI  uses this automatically and my presets are also

  • 00:12:49 set for this so don't worry the presets will  be one click. So follow the download process

  • 00:12:54 here. It will download all the original models  as you are seeing very fast. 100 megabytes per

  • 00:13:01 second on my computer. Then it will verify them  then you will be ready to start using them.

  • 00:13:06 Moreover since I am using extra model paths.yaml  file inside ComfyUI when I start my ComfyUI you

  • 00:13:12 see it adds all the models that I have given it.  So it sees all the models that I have in SwarmUI

  • 00:13:20 in ComfyUI. Therefore I am avoiding all the model  duplication. And I am using central model folder.

  • 00:13:28 You can also modify this file according to  your needs. Add more paths or more mapping.

  • 00:13:34 It works. SwarmUI also has this feature. You  can give another base model path and it see

  • 00:13:40 all of them. So you can have all of your models  inside ComfyUI and also use in SwarmUI as well.

  • 00:13:46 Since model downloader downloads with 16  connections it will merge all the downloaded

  • 00:13:52 pieces into single file at the end. So you will  see like this it is merging then it will verify

  • 00:13:59 the hash. So now it is verifying its hash value.  This is also working on cloud and Linux machines.

  • 00:14:05 This way we are ensuring that we have accurately  downloaded the models. Therefore we will never

  • 00:14:10 have any corrupted model at all. Moreover this  downloader application supports custom model

  • 00:14:16 downloads as well. So let's hide this. Go to the  URL downloader. In here you just need to give the

  • 00:14:23 model path. It can be CivitAI. It can be Hugging  Face. And you can select the folder where it will

  • 00:14:29 download so that you can download your custom  models this way on cloud machines or on your

  • 00:14:35 Windows computer very fast. You can also provide  your API keys to download even faster. This is

  • 00:14:42 very very useful especially on RunPod on Massed  Compute to download fast into the accurate folder.

  • 00:14:48 Okay so all the models have been downloaded.  Let's return back to our ComfyUI. Let's refresh.

  • 00:14:54 Then inside ComfyUI extracted zip file you will  see there are presets. So the newest presets are

  • 00:15:02 LTX 2 image to video, text to video, audio lip  sync and Z-Image base and Z-Image base with 2x

  • 00:15:09 upscale. For Z-Image base don't forget you need to  download same way Z-Image core bundle. Okay first

  • 00:15:16 I will show you LTX 2 audio to lip sync image  to video. Drag and drop it. So it is loaded. You

  • 00:15:24 shouldn't have any red or warnings here if you  have followed the steps that I have just shown

  • 00:15:29 you. This is extremely high quality and fully  optimized to work on low VRAM GPUs and systems

  • 00:15:36 as well. So first of all you need to pick your  image. Let's choose file to upload. I am going

  • 00:15:43 to use this anime image. Then you need to set  your width and height accurately. This is 9

  • 00:15:51 to 16 therefore I am going to use 720 to 1280.  If you are not sure how to set your width and

  • 00:16:00 height accurately I recommend you to use SwarmUI.  Unfortunately this preset doesn't exist in SwarmUI

  • 00:16:08 because it doesn't support it. It uses a lot  of custom nodes. Then quick tools reset params

  • 00:16:14 to default. I am going to use my LTX 2 image to  video. Then you can see that it will show you the

  • 00:16:21 resolution like this and init image. So choose  the same file in here. Resolution use closest

  • 00:16:27 aspect ratio. You see this image is 720 to 1280  pixel for this model. Okay it is set. Then you

  • 00:16:35 need to set your audio file which will be driving  audio. I am going to use this 45 seconds audio.

  • 00:16:44 SECourses. Let's go.

  • 00:16:48 Okay this way you can verify your audio is  working. So you see there is audio length in

  • 00:16:53 seconds. So first set your audio length.  You can also set audio start index here.

  • 00:16:58 Then when you click this run and stop  immediately you will see that it is

  • 00:17:04 showing you how many frames you need here. So  I need 1081 frames for this video. After that

  • 00:17:11 you need to write your prompt. When you are  using audio this is how you should prompt it.

  • 00:17:19 Describe your audio in details plus write the full  spoken or singing like this. You see this is the

  • 00:17:30 lyrics of my audio. If it was a speaking it would  be the subtitles transcription of the speaking.

  • 00:17:38 So this is the way of it. Then I have after few  tries I have come up with this prompt. You can

  • 00:17:44 pause the video and read it. And like this. So  this is my prompt like this and run. Then all

  • 00:17:53 you need to do is just wait. You don't need to  change anything else. You can also change your

  • 00:17:59 final video quality from here. You see CRF this  is the video quality. If you make this very low

  • 00:18:05 it will be very big generated MP4 file. Let's also  monitor our VRAM usage and the speed. Currently we

  • 00:18:14 are running it on no VRAM. Therefore it will use  minimal absolutely minimal amount of VRAM. So we

  • 00:18:22 can see both of the screens like this. It will  be therefore slow but it will work on every GPU.

  • 00:18:30 Okay once the generation starts. So even if you  have a 6 gigabyte GPU 4 gigabyte GPU you can use

  • 00:18:37 all these presets with the way I have shown you  in the beginning of the video how to start your

  • 00:18:44 ComfyUI. This applies to SwarmUI as well. In  the SwarmUI you set the same arguments into

  • 00:18:51 your ComfyUI backend. And it is also using  my GPU very decent. I mean look at that. It

  • 00:18:57 is using like 500 watts. This is really really  good. It is using my GPU very well. And it is

  • 00:19:02 only using 5000 megabytes VRAM right now. So this  way you can generate very long videos on very low

  • 00:19:10 VRAM GPUs. And you can use the same arguments  in your backend in your SwarmUI. It will work

  • 00:19:16 exactly same. Remember when we started the  ComfyUI it has shown what parameters what

  • 00:19:24 arguments have been used like no VRAM cache none  disable smart memory. Use Sage Attention disable

  • 00:19:31 asynchronous offload. This way I have minimized  my VRAM usage. So it is only using 5 gigabytes

  • 00:19:39 of VRAM for 1081 frames at HD resolution. And it  is generating it right now. This workflow first

  • 00:19:49 do 8 steps for initial video generation then it  does 3 steps of the upscale. The upscale will

  • 00:19:56 be of course slower. And during the upscale  it uses more VRAM as expected since it is

  • 00:20:03 now bigger resolution. So you can reduce your  length if you get out of VRAM error or on your

  • 00:20:09 computer it may do even further optimization  since I have VRAM maybe it didn't do I am not

  • 00:20:15 sure. And the time is still very decent. I mean  very acceptable time even at these settings.

  • 00:20:24 Let's see it.

  • 00:20:27 SECourses. Let's go. Late night screen glow. Ideas  on fire. Turning what if into something higher.

  • 00:20:34 From zero to flow. We learn it fast. Build it  ship it make the future last. New tools new

  • 00:20:39 rules we break the doubt. Take that prompt watch  the AI route. If you want the shortcuts done the

  • 00:20:43 right way. Hit play let's level up today. Hands  on the keyboard mind on the dream. Watch the

  • 00:20:49 model run like a laser beam. No fear no fluff just  clarity. One more step to your next victory. SECourses

  • 00:20:54 AI on the rise. Lighting up the dark  with electric skies. From prompts to pixels

  • 00:20:59 code to sound. We make it real we run this ground.  That's courses learn build repeat. Turn big ideas

  • 00:21:06 into something elite. If you want the best place  to start. That's courses straight to the heart.

  • 00:21:11 Purely amazing. This time there weren't any  noise. There weren't any text. There weren't

  • 00:21:16 any issues. This was even better than my initial  demo that I have shown you. So what about using

  • 00:21:22 the presets? They will be surely faster and they  will be using lesser VRAM since we will be doing

  • 00:21:29 lesser amount of frames. So let's begin with  the image to video. Drag and drop it. You can

  • 00:21:35 also open. So this is much more straightforward.  Choose and upload your image. Same way like this.

  • 00:21:43 Set your video height and width. I am going to  use the same like this. Type your prompt. And

  • 00:21:50 how you can make your prompt better? Now let  me show you. In the SwarmUI zip file you will

  • 00:21:56 see that there is LTX 2 enhance prompt feed for  LLMs. When you open it you will see its content

  • 00:22:04 like this. So to use this I am going to use Google  AI Studio. Then drag and drop the file like this.

  • 00:22:11 Enter your prompt and if you are also using an  image add your image as well. This will help

  • 00:22:18 it to also modify prompt according to the image.  And run. If you are not using image just text to

  • 00:22:25 video just type like this and it will enhance it  automatically for you. The logic is same. Okay it

  • 00:22:32 did enhance it. By the way using Google AI Studio  is free. Therefore just copy paste it and run.

  • 00:22:39 You can also change the duration from frame  count. This is 24 FPS model therefore when

  • 00:22:45 you divide this 24 it makes 5 second. So 5  multiplied with 24 plus 1 equal to our frame

  • 00:22:54 count. If I want it 10 seconds then I need to make  this 241 frames. So let's see the VRAM usage. It

  • 00:23:03 will use very minimal amount of VRAM since we  are only generating 121 frames. Okay let's see

  • 00:23:11 how much VRAM it will use. It is also using  very minimal amount of RAM memory too. During

  • 00:23:16 the upscale it will use more GPU obviously.  And the speed is also amazing even though I

  • 00:23:22 am running it with no VRAM command argument.  Okay with the upscale it is using 2.7. It is

  • 00:23:32 using 3 gigabytes of VRAM during the upscale.  This is amazing. I mean you will be able to

  • 00:23:37 generate 5 second HD videos with only 3 gigabytes  of VRAM memory. Can you believe that? Okay at the

  • 00:23:45 end it used some more for the VAE tiling but it  is fine. And we have the video. So let's see.

  • 00:23:53 Hello everyone welcome.

  • 00:23:55 It did some animation maybe changing the seed  will help or I need to change the prompt but it

  • 00:24:02 is working great so play with it to see. Then  I will restart my ComfyUI with proper starting

  • 00:24:11 and I will show Z-Image but for Z-Image I need  to remove the Sage Attention because currently

  • 00:24:18 Z-Image base is not working. The text to video  is also same I am not going to show it. So let's

  • 00:24:24 run the application with full optimizations with  full speed and without Sage Attention. Okay the

  • 00:24:32 application started. For Z-Image we are going to  use the Z-Image with 2x upscale. I recommend this.

  • 00:24:39 This is working really amazing. Amazing quality.  Then type your prompt as you wish you know this

  • 00:24:46 is just regular prompting. Let's use this prompt  to see what we are going to get. Okay run. I am

  • 00:24:52 also working on Z-Image NVFP4 model a properly  made NVFP4 model so it will be 2x faster with

  • 00:25:01 almost same quality. Our SECourses Musubi Trainer  application is going to have all these convert to

  • 00:25:09 quant options FP8 int8 and FP4 so that you  will be also able to convert any model you

  • 00:25:16 want into FP8 scaled quantized FP8 it will be  amazing quality like original BF16 or NVFP4. It

  • 00:25:24 will be 100 percentage faster on Blackwell GPUs.  All of them. 5060 5070 5080 whatever 50 or newer

  • 00:25:34 versions you get in future. So this is going to  generate 4 megapixel resolution image. This has

  • 00:25:41 the best upscale. If you don't want to upscale  you just need to use the Z-Image base preset.

  • 00:25:48 Okay let's see our generated image. By default  I am doing 40 steps but you can reduce the step

  • 00:25:55 count to 20 as well. With LTX 2 the step  count is not changing. It is 8 and in the

  • 00:26:02 upscale it is 3. But with the Z-Image base  or with majority of the other models you can

  • 00:26:08 change the steps count according to your need  the speed. Currently it is using 18 gigabytes

  • 00:26:14 of VRAM but with optimizations you can run this  as low as 6 gigabyte GPUs. Since we are doing 4

  • 00:26:21 megapixel image it is taking some time but we  will get an amazing quality image. By the way

  • 00:26:26 I don't know how will this video prompt work but  you know the drill you can change the prompt as

  • 00:26:32 you wish. Okay it generated image let's open in  the new tab. So open image like this. Yeah for

  • 00:26:41 this prompt this is a very decent output. All  the generated files will be saved inside your

  • 00:26:47 ComfyUI into your output folder. You see they  are all saved. Even the videos are saved. To

  • 00:26:54 remember the videos you need to do this. You see  there is PNG file drag and drop it and it will

  • 00:26:59 load everything from that PNG with the sets. So  with the videos it saves a PNG file to save the

  • 00:27:06 metadata. And the images are also saved like this  with the ComfyUI extension. If you want to save

  • 00:27:12 images with another prefix you can change the  file name prefix from here. Or with the Z-Image

  • 00:27:18 you can change the save prefix from here. You see  ComfyUI you can change it to anything you want.

  • 00:27:24 So this is all for ComfyUI for today. All other  presets are working exactly same logic. You can

  • 00:27:30 use all of our amazing presets that is coming  with our zip file. You see all of them is working

  • 00:27:37 tested working perfect. You just need to download  the bundles. Like Z-Image, Flux 2, Wan, Qwen. We

  • 00:27:45 have bundles for every major model out there. And  for SwarmUI I am going to use the same ComfyUI

  • 00:27:52 installation the backend. So let's just close our  ComfyUI. So as usual download the latest SwarmUI

  • 00:27:57 model downloader zip file. Overwrite your previous  files. Then you should update your SwarmUI Windows

  • 00:28:04 update SwarmUI. This will install necessary newer  stuff and also update your SwarmUI to the latest

  • 00:28:12 version. I have developed an extension for SwarmUI  to support LTX 2 image to video upscale. Normally

  • 00:28:20 SwarmUI is not supporting it but we are now  supporting it. You can see that my extension is

  • 00:28:27 compiled like this when you start it. Then go to  the server backends and enter your backend as in

  • 00:28:34 the same as other videos. You see ComfyUI backend.  How you give the ComfyUI backend it is so easy. Go

  • 00:28:41 to your ComfyUI installation ComfyUI folder copy  it. Edit the here. Actually let me add a new one

  • 00:28:50 to show you. So let's refresh. And ComfyUI self  starting and edit it. Copy paste main.py. Extra

  • 00:28:59 arguments. This is where you define your ComfyUI  starting arguments like --use-sage-attention like

  • 00:29:06 --cache-none like --no-vram. This way you can  customize and optimize your ComfyUI backend.

  • 00:29:14 Normally I just use Sage Attention. The only  case that I don't use at the moment is Z-Image

  • 00:29:20 base model because it degrades the quality  significantly. It is broken at the moment. So

  • 00:29:25 wait for your backend to start and it is started.  Quick tools reset params to default. You need to

  • 00:29:31 get the latest presets. So either import from  here or my recommended way is Windows preset

  • 00:29:38 delete import. So run it and hit yes and it will  delete all of your previous presets and refresh

  • 00:29:44 them with the new one. So refresh sort by name.  All the newer presets are arrived like Z-Image

  • 00:29:50 base or LTX 2 text to video LTX 2 upscale. When  you use upscale you need to reduce your resolution

  • 00:29:57 to half. So let's make a demo of image to video.  Quick tools reset params to default. Click here

  • 00:30:02 and direct apply. Since I have downloaded  the LTX 2 video core bundle all the models

  • 00:30:09 are automatically downloaded and selected. I  have spent huge time to prepare these presets.

  • 00:30:14 Then to use the image to video you need to choose  your init image like this one. Then you need to

  • 00:30:21 go to the image to video tab and set your video  frame count. It is 121 frames right now. So it

  • 00:30:28 will be 5 second video. Then enter your prompt  same as in ComfyUI and you can generate as many

  • 00:30:36 as videos you want. Let's generate one. So this  will generate an image to video. You can always

  • 00:30:41 follow the status from server logs debug menu to  see what is happening. Let's also write another

  • 00:30:48 prompt with an enhancement. So I am just going to  use the Google AI Studio Gemini 3 Pro preview and

  • 00:30:56 from SwarmUI we have LTX 2 enhance type my prompt.  So this is my prompt. Let's see the enhancement.

  • 00:31:05 Okay you see it had an error. Sometimes you may  get this because of how the ComfyUI working just

  • 00:31:11 hit generate again. This is happening because  when you first time run it for some reason it

  • 00:31:17 fails to properly load model. To prevent it you  need to add --use-cache-none. So when you first

  • 00:31:25 time click generate if you get error just wait  and click generate again. This happens with the

  • 00:31:32 LTX 2 video they didn't fix it yet. I am expecting  fix hopefully soon. Don't worry just hit generate

  • 00:31:39 again. You can also add the optimizations like  disable smart memory or like cache none or low

  • 00:31:49 VRAM. So one of them will work for your case.  Once the first generation is completed it will

  • 00:31:56 keep working. Okay now it is generating image to  video. It should be also pretty fast. Yes. Okay

  • 00:32:02 you see the image is changed like this  because of my prompt so I need to modify

  • 00:32:09 my prompt. For example like this. A video of a  talking man saying hello everyone. So if your

  • 00:32:17 image changes dramatically like this because  of your prompt. You need to make it in a way

  • 00:32:24 that it will not change the initial image.  Let's see the new prompt impact. You see the

  • 00:32:30 first video was done in 55 seconds. Second  one is done in 35 seconds. Let's see it.

  • 00:32:37 Hello everyone welcome.

  • 00:32:39 Yes perfect quality. Perfectly matching the  initial image. There is no changes so it is

  • 00:32:44 all about your prompt. Prompt enhance may break  image to video however it works really good at

  • 00:32:51 text to video. Now let's see the text to video.  So quick tools reset params to default and we

  • 00:32:57 have LTX 2 text to video direct apply. Let's see  our prompt like this. So this is enhanced prompt.

  • 00:33:05 Change the resolution or aspect ratio. You can  also change the duration from text to video. The

  • 00:33:12 image to video is here image to video you change  the frame count there. For text to video it is

  • 00:33:18 here. So let's make this 241 frames and that's  it generate. You can also upscale with text to

  • 00:33:27 video how for text to video upscale you need to  custom and half resolution like 640 to 360 and

  • 00:33:36 go to presets and just apply in additionally this  LTX 2 upscale like direct apply and let's generate

  • 00:33:44 another one. The first one will be original HD  the second one will use the upscaler model of

  • 00:33:52 the LTX 2. All the generations of the SwarmUI  will be saved inside SwarmUI output local raw

  • 00:34:02 and they are categorized by the days that you  generated them. You see all my images videos

  • 00:34:08 are all saved here with their metadata as well.  Okay the first video generated let's watch it.

  • 00:34:15 Faithfulls! Witness me!

  • 00:34:24 Yeah for this prompt this is what it has  generated. Now we have the upscaled version

  • 00:34:29 let's watch it as well. So this way you  can tune your prompt try different prompts

  • 00:34:44 and see what you are getting you can also  compare. This was a 10 second video and it

  • 00:34:49 was only generated in 40 seconds. 40 seconds  for 10 second video. Amazing. I mean this is

  • 00:34:56 241 frames. This would take forever on Wan 2.2  model. It is only taking 40 seconds on LTX 2.

  • 00:35:05 So what about Z-Image? For Z-Image I am going to  go to backends and edit my backend and remove the

  • 00:35:12 Sage Attention. Currently this is mandatory.  Probably it will be fixed soon. Then let's

  • 00:35:18 open another tab reset params to default and  in the preset our base preset is this one. So

  • 00:35:25 let's use the same prompt as in this case I don't  know what it will generate. Then I am also going

  • 00:35:33 to apply the upscale. Then with the same seed  let's also use the upscale. So for upscale we

  • 00:35:40 are going to use image upscale direct apply  and generate another one. The first one will

  • 00:35:45 be 1024 to 1024 the second image will be 2048 to  2048 with our image upscale preset. This image

  • 00:35:54 upscale preset works on every image generation.  Okay the first image is done it is here it is

  • 00:36:01 like this. Let's see its upscaled version too.  So this is what it has generated as a base. I

  • 00:36:08 am doing 40 steps for highest quality but you  can reduce this to 20 steps as well. There will

  • 00:36:15 be minimal impact on quality and it will be twice  faster. As I also said I am also working on NVFP4

  • 00:36:24 version of the Z-Image base model therefore  it will be 100 percentage faster and almost

  • 00:36:30 same quality. Moreover we have FP8 scaled of  the Z-Image base model so you can download

  • 00:36:37 it with our downloader and also use it if you  are low on VRAM like 6 gigabyte GPUs. However

  • 00:36:44 this model is not very big. The FP8 scaled  is only 5.6 gigabyte the BF16 is like only

  • 00:36:51 12 gigabytes. Okay the upscaled image is here.  So this is the base image this is the upscaled

  • 00:36:58 image. Let's compare them with our comparison  application video comparison slider. The link

  • 00:37:04 will be also description of the video to install  this. Start image comparison. Let's put both the

  • 00:37:11 files and full screen. So on the left we see  the original normal image and on the right

  • 00:37:17 we see the upscaled version. I mean look at  the quality difference. The upscaled version

  • 00:37:22 is much better. Many times better. It has added  amazing amount of details quality so I recommend

  • 00:37:30 you to use upscaled preset of the Z-Image  base model to get really good images. The

  • 00:37:38 realism of the Z-Image is already amazing  so you can type a realistic prompt to get

  • 00:37:44 realistic image like in this case. So it is all  about your prompting to get whatever you want.

  • 00:37:51 Now I will show you how to use ComfyUI and SwarmUI  on RunPod. It is exactly actually same as before.

  • 00:37:58 So once you downloaded the zip file you will  see that we have RunPod SimplePod ComfyUI

  • 00:38:03 instructions. RunPod and SimplePod are working  exactly same way. You just need to follow this

  • 00:38:09 file. You see we have RunPod section. So please  use this link to register I appreciate that. Once

  • 00:38:16 you registered login if it doesn't auto login  you. Then go to billing add some credits. Then

  • 00:38:25 return back to our instructions file always  use the template mentioned here. So currently

  • 00:38:32 we need to use this template for CUDA 13.  This is based on official template don't

  • 00:38:37 worry it just have some tweaks for you. Now  this is super important. If you are going to

  • 00:38:44 use your permanent storage select your permanent  storage from here. So everything you do will be

  • 00:38:49 saved 100 percentage. If you are not going to use  permanent storage so just don't select anything

  • 00:38:55 here. What you need to select here is additional  filters and CUDA version 12.9. Unfortunately they

  • 00:39:04 still didn't add CUDA 13 here. They need to add  and upgrade their Nvidia drivers but they still

  • 00:39:11 didn't do. Therefore we are selecting 12.9. This  should bring us machines that has Nvidia drivers

  • 00:39:19 that can run CUDA 13. Then select NVME and 100  gigabytes. So when we apply these two presets

  • 00:39:27 unfortunately there aren't many machines we can  use yet with CUDA 13 ComfyUI. So these are the

  • 00:39:34 machines available. I am going to use RTX Pro  6000. This is one of the best as you know. Then

  • 00:39:39 you see it has selected our template. I am going  to contact to RunPod team to upgrade their drivers

  • 00:39:45 so that CUDA 13 will arrive and they will upgrade  their Nvidia drivers so we will be able to use

  • 00:39:52 more machines. And in the template you can edit  and change your volume disk if you are using your

  • 00:40:00 permanent network storage. It will use whatever  the space you have there but this will make a new

  • 00:40:07 storage. So making permanent storage is so easy  go to storage tab here click new network volume

  • 00:40:13 select the server according to your server  you will be limited to that region only so

  • 00:40:19 this will limit your GPU availability. You can  also click from here to see which regions have

  • 00:40:25 more GPUs of this for example this one. Type  your storage name 600 gigabytes type like 600

  • 00:40:32 gigabyte and create network volume. Then you  will be able to use it but I am not going to

  • 00:40:37 use right now so I am going to delete it to  not waste any money. Okay let's delete it.

  • 00:40:43 So then deploy on demand. Now you need to wait  a little to start the machine. It will take a

  • 00:40:51 while like 5 minutes depending on the region on  the server maybe 2 minutes. Unfortunately RunPod

  • 00:40:58 machines are slow. Therefore Massed Compute  is much faster. Now we are also supporting

  • 00:41:04 SimplePod so easy to use same way actually. It  is also faster and more GPUs available there.

  • 00:41:10 I started 3 identical machines to see which  one is starting fastest. Once I verify the

  • 00:41:17 starting fastest one I will delete the other  ones. Unfortunately with RunPod this is like

  • 00:41:23 this you don't have many options you can also  pick another GPU to make it start faster for

  • 00:41:31 example like H200. This is a beast machine.  It will work even faster than RTX Pro 6000.

  • 00:41:38 So if you don't want to wait too long you can  pick your machine accordingly or you can use

  • 00:41:43 permanent storage to avoid this. So all of them is  failing to start right now. Okay this one started.

  • 00:41:51 Yeah this is probably H200 let's see yes. You  see the H200 has a much better connection and

  • 00:41:58 disk speed but I just emailed the RunPod team  I think they will upgrade their Nvidia drivers

  • 00:42:05 therefore we will have more machines available.  Of course always you can use SimplePod and Massed

  • 00:42:13 Compute. They need to upgrade their Nvidia driver  so that we can use any machine available. Yeah I

  • 00:42:19 will continue with H200 for this tutorial  because it is already starting so I will

  • 00:42:24 disable other machines stop not waste any time  any money start. This is a shame the size of

  • 00:42:32 the template is 9 gigabytes therefore it should  be instant but you see the RunPod is very very

  • 00:42:38 slow nowadays. They are degrading in quality  unfortunately. I hope they become better. So

  • 00:42:44 we are giving them feedback you should also give  them feedback email them contact their support

  • 00:42:50 tell them to upgrade their Nvidia driver add  more GPUs. Okay this should start fairly fast.

  • 00:42:56 You can see the download process. It doesn't  matter which GPU you select my installers and

  • 00:43:02 template supports all of them. What matters is  the Nvidia driver which we are going to see in

  • 00:43:08 telemetry when it starts it will show the  Nvidia driver here. It should be 575 plus.

  • 00:43:15 I just emailed them to upgrade all into 580 plus  so that we can use any machine hopefully they do.

  • 00:43:23 Okay the machine started when I go to connect you  see it shows connect from 3001 this is ComfyUI

  • 00:43:30 port we didn't start it yet and JupyterLab. So  click the JupyterLab interface. You see the driver

  • 00:43:36 is when we go to telemetry we see the driver is  575. You can also try CUDA 12.8 but it may not

  • 00:43:44 work. This is guaranteed to work with all cases  575 driver. Hopefully they will upgrade and we

  • 00:43:52 will have it. Okay so the JupyterLab interface  started I will just drag and drop my ComfyUI

  • 00:43:58 zip file into workspace wait for upload to be  completed. So you see it is uploading right now.

  • 00:44:04 Then I will extract archive refresh. Yes all done.  Now open the RunPod SimplePod ComfyUI instructions

  • 00:44:13 copy this command open a new terminal paste and  hit enter. Then it will ask you which options to

  • 00:44:20 install same as Windows 1 and 100 we are going  to install these two bundles and hit enter then

  • 00:44:26 hit enter. Then it will install the ComfyUI fully  properly for us. Now you need to download models.

  • 00:44:33 For downloading models we are going to use SwarmUI  model downloader so it is here just drag and drop

  • 00:44:38 it here refresh. It is still uploading wait  for upload to be completed. This installation

  • 00:44:45 is extremely optimized and fast but it depends  on your machine as always. By the way you should

  • 00:44:52 also verify your GPU open a new terminal pip  install nvitop then type nvitop. And verify

  • 00:45:01 that GPU is there and it is empty. Sometimes  GPUs are being stuck or frozen or something.

  • 00:45:08 Okay the SwarmUI installer also downloaded.  So I will also install SwarmUI exactly same as

  • 00:45:16 in Windows tutorial part. Okay so for SwarmUI  installation we have RunPod SimplePod SwarmUI

  • 00:45:24 install instructions. Just copy this entire  command open a new terminal and paste it.

  • 00:45:32 This will install SwarmUI very fast. Once you see  the SwarmUI folder here you can start downloading

  • 00:45:38 models or you can not install SwarmUI and  just download models into ComfyUI as well.

  • 00:45:44 This machine is a little bit slow at installation  unfortunately. Okay SwarmUI folder appeared so we

  • 00:45:50 can begin downloading models. RunPod SimplePod  model download instructions copy this open a new

  • 00:45:57 terminal paste it. Meanwhile both SwarmUI and  ComfyUI are getting installed. You can follow

  • 00:46:04 each of them. Unfortunately this machine is  very slow. So for SwarmUI installation you

  • 00:46:09 see it has generated Cloudflared link agree  customize okay next just yourself none this

  • 00:46:17 is important because we are always going to  use our custom ComfyUI installation next I

  • 00:46:22 am not going to download anything next yes I  am sure to install and SwarmUI installed. So

  • 00:46:28 our model downloader started let's open it from  this Gradio link. It will automatically see my

  • 00:46:34 SwarmUI installation and its accurate folder.  So you see it sees the models folder here.

  • 00:46:41 If you don't want to install SwarmUI and use  it go to ComfyUI right click models copy path.

  • 00:46:48 Then put a backslash and type like this. So  it starts with backslash then enable ComfyUI

  • 00:46:55 folder structure and download the models. So let's  download LTX 2 for ComfyUI installation. I will

  • 00:47:04 also use same folder for SwarmUI. Okay let's  wait for ComfyUI installation to be completed

  • 00:47:11 meanwhile the models are getting downloaded  we can see them here it is downloading them.

  • 00:47:17 Okay so our ComfyUI installation has been  completed. Now let's start our ComfyUI. To do

  • 00:47:24 that return back to your RunPod SimplePod ComfyUI  instructions txt file and in the below you will

  • 00:47:31 see the run command. This is important to use it.  Open a new terminal and copy paste and hit enter.

  • 00:47:39 You see we are starting with use Sage Attention.  You can remove this add here other additional

  • 00:47:45 arguments. This is how we start the ComfyUI  instance. Initial starting may be a little bit

  • 00:47:51 slow but you should see the appropriate messages  in here. We can see that how much VRAM and RAM

  • 00:47:58 memory our machine has. We have 143 gigabytes of  VRAM and 1.8 terabytes of RAM memory yes this is

  • 00:48:07 an massive RAM memory. So it is using CUDA  13 torch 2.9.1 as I said my installers are

  • 00:48:15 properly made for every GPU out there so you  can use my installer with every GPU on cloud

  • 00:48:22 or on Windows. So the ComfyUI has been started.  When you see this it means it has been started.

  • 00:48:28 How to connect? Return back to your MyPod instance  and you will see that there is port 3001. When you

  • 00:48:35 open it now you will see your ComfyUI interface  which is running on RunPod and the rest is exactly

  • 00:48:44 same as using on your Windows computer. So my  ComfyUI is started. All of the models also have

  • 00:48:51 been downloaded. Now I am ready to use presets. So  let's drag and drop our LTX 2 audio lip sync image

  • 00:48:57 to video preset. It is loaded. You see there  is no red no error messages it is perfect. So

  • 00:49:06 choose file to upload. I will do the same as on  the Windows part. So let's see how much it will

  • 00:49:13 take. This is 45 seconds. Let's also copy paste  our prompt. I also need to set my duration the

  • 00:49:21 rest is exactly same as on the Windows tutorial  part. You can also calculate 45 seconds equal to

  • 00:49:29 45 multiplied 24 plus 1 1081 frames. And we can  hit run here. Now you can follow the CMD window

  • 00:49:41 on here. Let's also see the nvitop. pip install  nvitop. nvitop. So we can follow the VRAM usage

  • 00:49:51 from here. The initial loading of the model may  be slow but we will see. You will also get these

  • 00:49:58 messages you can ignore them it is still working  perfect. This is ComfyUI mistake actually on some

  • 00:50:05 models it shows these as unexpected or some other  messages but it is working perfect right now. This

  • 00:50:11 is also ComfyUI mistake which they have to fix but  it is not fixed yet. By the way on this machine we

  • 00:50:18 can also use BF16 model. Our unified downloader  is downloading the FP8. How you can download it

  • 00:50:24 on this machine? In the video generation models  you can go to LTX 2 video models and you can

  • 00:50:30 download the BF16 version directly. Because  this machine has a huge amount of VRAM memory.

  • 00:50:37 So the initial loading of the model is taking  forever on RunPod disks because they are slow.

  • 00:50:43 Okay model loaded and generation started. You  see it is almost instant even for 45 seconds.

  • 00:50:52 The first 8 steps took only 13 seconds. And now  it is upscaling. It is using 48 gigabytes of

  • 00:50:59 VRAM memory. So if you have RTX Pro 6000 on your  Windows machine you can run this model very fast.

  • 00:51:06 I mean shame on Nvidia. They are still selling us  RTX 5090 from only 32 gigabytes of VRAM memory.

  • 00:51:15 Okay this was also done very fast so it is like  30 seconds to generate 45 seconds video. We are

  • 00:51:22 almost done. It is doing the now VAE decoding.  It is using 57 gigabytes of VRAM memory right

  • 00:51:29 now. Yeah it is already okay it is now combining  video after that it will be completed. So it takes

  • 00:51:36 like 1 minute to generate 45 seconds video on  H200 and it is already done yes. Let me show.

  • 00:51:58 Okay it is same as in the Windows tutorial part  it is working perfect no issues directly working.

  • 00:52:05 So how you can also use the SwarmUI in here?  Since everything is ready all I need to do is

  • 00:52:12 I need to turn off the ComfyUI. So that SwarmUI  can use the entire GPU. To turn off the ComfyUI

  • 00:52:21 normally you wouldn't need this if you don't  run the ComfyUI. When you return back to the

  • 00:52:26 prompts you will see that it has this command.  So copy and open a new terminal if you want to

  • 00:52:33 terminate your ComfyUI. This will terminate  the ComfyUI properly and you will see that

  • 00:52:39 it is using zero amount of VRAM. Now my SwarmUI  already had started. I am going to go to server

  • 00:52:47 backends ComfyUI self starting. I need to copy  the path of the ComfyUI main.py so right click

  • 00:52:56 copy path. Go to your SwarmUI put a backslash like  this and --use-sage-attention and save. Now the

  • 00:53:07 SwarmUI will be ready. Normally we won't have any  models here because we didn't download them into

  • 00:53:14 here we downloaded them into ComfyUI folder. So  I can go to server configuration and I can go to

  • 00:53:21 models folder here copy path put a backslash like  this save after that we may be needed to restart

  • 00:53:30 but let's see. Okay let's refresh. Yes the model  arrived but it doesn't see the other model because

  • 00:53:38 I need to change this into checkpoints. This is  where it is downloaded in the ComfyUI. Yes. Now

  • 00:53:46 I am ready to use everything. The VAE is here  let's see. Okay it doesn't show but let's see

  • 00:53:54 the Loras. Okay Lora is also different it must  be something like this. By the way you don't

  • 00:54:00 need this if you just download them into your  SwarmUI folder. So let's see the Loras arrived.

  • 00:54:07 Okay server configuration checkpoints Loras VAE.  Oh I know why it is not working because it is not

  • 00:54:15 Loras but Lora in ComfyUI. Yes like this. And the  VAE folder is we need to check inside models it

  • 00:54:24 is yeah accurate VAE. So you can look for the  folder names. The Lora is Lora here let's see.

  • 00:54:32 Okay this is empty. The Lora empty. Loras empty.  And it is Loras. So it is case sensitive on Linux

  • 00:54:42 machine therefore it is this. Yes. Save. And  yes now I can see the Loras. Now I can import

  • 00:54:51 the presets. SwarmUI model downloader the preset  45 import and it is ready. So now I can generate

  • 00:55:02 anything I want. Let's generate text to video.  So it is let's sort by the name. You can also

  • 00:55:09 search from here like LTX LTX 2 text to video  direct apply. So models are selected. The Loras

  • 00:55:18 are also selected. All I need to do is just type  my prompt. A man talking to camera and speaking

  • 00:55:27 this is amazing. This is a just simple prompt you  know the drill from the Windows tutorial part. You

  • 00:55:34 can follow the status from logs debug menu. Now  the SwarmUI will work perfectly fine on RunPod.

  • 00:55:43 So how you can download the generations? Go to  your output folder. So in here I can download

  • 00:55:50 entire output with right click and download  as an archive. It will start to download like

  • 00:55:56 this. So you see it is downloaded the entire  output. Since our downloader also downloads

  • 00:56:02 the clip models text encoder models VAE models  you won't see any of them downloaded from here.

  • 00:56:08 We do everything fully automatically for you  with the fastest and most accurate way both

  • 00:56:15 in Windows and also in Linux. My Massed Compute  instructions Massed Compute install files are

  • 00:56:22 all for Linux so if you are a Linux user you  can use them right away. It is still loading

  • 00:56:27 model for SwarmUI to generate. Okay so the  generation has been completed let's see it.

  • 00:56:33 This is amazing.

  • 00:56:35 Yeah this was our simple prompt and we got it.  So how to download SwarmUI generations? Go to

  • 00:56:41 SwarmUI go to output local. You can download  entire raw folder download as an archive and

  • 00:56:48 it will start download and save it and you  have that. So how to terminate your machine

  • 00:56:53 from stop pod. This will reduce your cost  per hour but it will not make it zero. To

  • 00:57:00 make it zero just terminate pod and everything  is deleted. If we were using storage we would

  • 00:57:06 still have it in here. But we are not  using permanent storage so all is gone.

  • 00:57:11 And one more thing before we proceed the RunPod  developer team had replied my emails and they

  • 00:57:18 are working on improving the Nvidia driver  of the pods. You can see their answer here.

  • 00:57:27 So hopefully in future all of the machines will  work with our ComfyUI CUDA 13 installation. So

  • 00:57:35 if your machine doesn't work currently  you can select the CUDA 12.9 probably

  • 00:57:42 they will add 13 too or CUDA 12.8 but  this will probably won't be necessary

  • 00:57:50 once they upgrade the Nvidia driver  of all the machines. This is for now.

  • 00:57:55 And as a next step I will show how to use  SimplePod. For SimplePod we are going to use

  • 00:58:00 RunPod SimplePod ComfyUI instructions file. You  see the SimplePod is here. What is the difference?

  • 00:58:06 Let's register from this link I appreciate that.  Then go to your dashboard add some billing like

  • 00:58:13 this. Then we are going to use this SimplePod  template. This is faster than the RunPod. You

  • 00:58:20 see it is showing this use template like this.  This template already has all the features that

  • 00:58:29 we need. Now you need to rent a machine. So  we have RTX 5090 very cheap on SimplePod you

  • 00:58:38 see it is only 47 cents. This is double price  on RunPod. RTX Pro 6000 is only 72 cents. This

  • 00:58:47 is almost triple price on RunPod. So let's go  with the RTX Pro 6000 Blackwell GPU. This GPU

  • 00:58:54 has 96 gigabyte of VRAM memory and it is same  fast as the RTX 5090 even faster than that. And

  • 00:59:04 then run. So this is going to have 51 gigabytes of  disk space therefore you can edit it and increase

  • 00:59:12 your disk space according to your need like 200  gigabytes. Then it will re-create the disk. When

  • 00:59:21 you are first time creating the disk you can also  select from there. So let's return back and click

  • 00:59:27 here. The template is selected. You can edit  and use. You see there is use template and edit

  • 00:59:32 and use. Let's use the edit and use and there  is also add persistent volume. SimplePod also

  • 00:59:38 has persistent volume let me show you that. So  go to storage add a new persistent volume like

  • 00:59:44 200 gigabytes and the name is 200. It is like 3  times cheaper than the RunPod. Save. Let's return

  • 00:59:51 back to our template selection double click.  And I am going to select add persistent volume

  • 00:59:57 and edit and use. Now it will let me select from  here 200 make this workspace this is important.

  • 01:00:06 And the system is ready. So this is going to  use my persistent volume. Save and use. And

  • 01:00:12 when I go to my servers you will see that it is  going to start. Okay we made a mistake let's do

  • 01:00:20 again. Add persistent volume edit and use. So  200 workspace. Okay save and use. Oh by the way

  • 01:00:29 it changed the template it still didn't selected a  machine so we are going to select this machine for

  • 01:00:35 example and run. So this one is going to use my  persistent volume. The other one is also started

  • 01:00:42 here. They don't show the persistent volume  but you will understand it actually there is

  • 01:00:48 one hint of it. This shows 200 gigabytes and this  shows disk volume and volume. When you have disk

  • 01:00:54 volume and volume it means that you are using your  persistent volume. So the rest is same how same?

  • 01:01:01 You see that there is Jupiter interface direct  link go to there. If it asks you to continue

  • 01:01:08 continue because direct means that it is using  direct connection. Then you see the RunPod like

  • 01:01:14 interface started. Upload your ComfyUI SwarmUI  whatever you want. By the way everything is much

  • 01:01:21 faster on SimplePod that is the main difference  and everything is much cheaper on SimplePod. So

  • 01:01:27 you can use both SimplePod and RunPod then extract  archive. And the rest is same let me just show you

  • 01:01:35 the start of the terminal like this 1 and 100.  Let's see the installation speed. Everything

  • 01:01:43 is ready in this template. Okay the installation  is just amazing amazingly fast as you are seeing

  • 01:01:48 right now. Moreover you can download files much  faster on SimplePod or upload. You see there is

  • 01:01:55 file browser when you click this it will let  you to download and upload files directly. Go

  • 01:02:00 to workspace and every file will be here you  can just select and download like this. Or you

  • 01:02:07 can just upload from here. The upload allows you  to upload file and folder. It is very very fast

  • 01:02:14 compared to the RunPod. This way you can upload  and download much faster than RunPod. So how to

  • 01:02:22 terminate machines once you are done go back to  your my servers there is delete options there

  • 01:02:29 is no stop option here. So let's delete and let's  delete. But all the data in your permanent storage

  • 01:02:37 will remain there even if we deleted exactly as on  RunPod. By the way on SimplePod we have forgotten

  • 01:02:44 to show one more thing which was connecting the  ComfyUI interface. For connecting the ComfyUI

  • 01:02:51 interface we are going to use the 3000 port. My  template already exposes that 3000 port so that

  • 01:03:00 you will be able to connect from 3000 port on  SimplePod to your ComfyUI instance. For SwarmUI

  • 01:03:08 we are also using Cloudflared link as in the  RunPod. So it is same as the RunPod actually.

  • 01:03:14 So this is an alternative to RunPod much  faster much cheaper but the disadvantage

  • 01:03:19 is that it has limited amount of different  GPUs. So when I go to rent GPU docker you

  • 01:03:25 see 5090 and RTX Pro 6000 are the main GPUs  that you can use. These two are the very best

  • 01:03:31 GPUs actually that you can use for cheap  but it is missing like A100 H200 B200 very

  • 01:03:38 specific GPUs but it has the most commonly  used amazing GPUs. So you were asking me to

  • 01:03:43 make Vast AI scripts but instead of Vast AI I am  providing you SimplePod. Cheaper better faster.

  • 01:03:52 So as a next step let's do on the Massed Compute.  To follow the Massed Compute open the Massed

  • 01:03:58 Compute instructions read txt file. Please use  this link to register I appreciate that. Once

  • 01:04:05 you registered and logged in go to billing and  add some credits as usual then go to deploy.

  • 01:04:15 And in here it will show you the available GPUs.  They have amazing prices so you can use H200 if

  • 01:04:23 it is available and apply our coupon. Let's see  if there is any availability. Oh actually there

  • 01:04:29 is available yes nice. So you see currently it is  2.6 dollars per hour with our coupon SECourses it

  • 01:04:36 becomes 1 dollar 95 cents. This is an amazing it  is like 4 dollars on RunPod it is 2 dollars here.

  • 01:04:45 So we can rent this or you can rent RTX Pro 6000  if it is available. Oh they have sold out all

  • 01:04:52 the RTX Pro 6000s. They are adding new GPUs but  they are very popular. So let's go with this one.

  • 01:05:00 This is even better at training than inference.  So if you are going to do training since this GPU

  • 01:05:06 has HBM memory high bandwidth unified memory it  is faster than the other GPUs like RTX Pro 6000

  • 01:05:15 or 5090. By the way when you are deploying don't  forget that select category as a creator select

  • 01:05:22 image as an SECourses. This is our image. This  is mandatory to select then apply our coupon and

  • 01:05:28 pick any GPU you want. You need to pick any  GPU spots are not working but all other GPUs

  • 01:05:34 are working like this. You see it is 41 cents.  For example L40S it is 61 cents. Like H100 it

  • 01:05:44 is okay this is spot instance yes H100 it is 1.76  cents but I recommend H200 this is an amazing GPU.

  • 01:05:54 So wait for GPU to be initialized it will take  a little while. To connect we are going to use

  • 01:06:00 ThinLinc client as usual so download the ThinLinc  client from here for Windows Linux Mac. So just

  • 01:06:07 download it. Then double click and start and  next next install it will install and start.

  • 01:06:14 From this ThinLinc client you need to go  to options local devices just clipboard

  • 01:06:19 synchronization and drives and add a shared  folder with read and write permission I am

  • 01:06:24 using this folder. Then move the downloaded  ComfyUI and SwarmUI files into your shared

  • 01:06:32 folder so that you will be able to access  and use them. You can also put your other

  • 01:06:37 files but you cannot use this for big files.  For big files you have to use other storage

  • 01:06:45 system like Hugging Face or Google Drive.  Okay Massed Compute is getting initialized.

  • 01:06:50 Machine initialization is now faster than  before it should be done in 5 to 10 minutes.

  • 01:06:56 Okay so the machine has been initialized let's  connect it copy this login URL. Okay this was

  • 01:07:02 done. Paste here copy password paste here copy  username paste here connect. Continue. Okay it

  • 01:07:11 is connecting then start. Okay it is starting  and it is started. This is running on cloud on

  • 01:07:19 Massed Compute. Go to home go to thin drives.  You can also login your Patreon from here and

  • 01:07:25 download links. Go to Massed Compute you see  it is like your Windows computer there is

  • 01:07:31 Chrome you can login download upload you can do  whatever you want. Wait it to list the files in

  • 01:07:38 your shared folder in here. So it will list all  the files I have in here. This is synchronizing

  • 01:07:45 with my computer but it is slow. So I am going to  copy paste all the files version 78 and SwarmUI

  • 01:07:54 latest version into downloads folder. Do not run  anything in this shared folder. Always copy them

  • 01:08:01 into downloads folder then we are going to run  them from here. So wait for copy to be completed

  • 01:08:08 it shows the progress here. The ComfyUI is done so  extract here and Massed Compute instructions read

  • 01:08:17 txt file. So to install we are going to use this  command copy open a new terminal and paste it.

  • 01:08:27 Then it will ask us to which ones to install 1 and  100 and hit enter. So the installation on Massed

  • 01:08:36 Compute is like 10 times to 100 times faster than  on RunPod the disk speed is like 100 times faster

  • 01:08:42 than RunPod on Massed Compute. Their disadvantage  is that they don't have permanent storage system

  • 01:08:48 yet so every time you have to start from beginning  but the speed is just mind blowingly fast so it

  • 01:08:56 takes like 2 minutes to setup and start the Massed  Compute. So installation is almost completed you

  • 01:09:02 see. Let's also download the files to download the  files we are going to use SwarmUI model downloader

  • 01:09:07 extract here. Massed Compute already has SwarmUI  so that you don't need to install but if you want

  • 01:09:14 you can install. So Massed Compute model downloads  here I will copy copy paste and open terminal and

  • 01:09:23 paste. Okay it is going to install and start  the model downloader this is really fast.

  • 01:09:29 I will download files into the SwarmUI folder but  I will make them to be used by the ComfyUI as well

  • 01:09:39 like in the Windows. So I will have the access to  the files from both applications. You see it is

  • 01:09:45 currently using this as a base path this is where  the SwarmUI installed in Massed Compute. So let's

  • 01:09:54 download LTX 2 bundle download all. The download  is also really fast so it started. Before I start

  • 01:10:03 the ComfyUI instance I am going to do a change.  Okay it is already installed. It took like 1

  • 01:10:08 minutes. Go back to ComfyUI and you will see that  there is extra model paths.yaml file open it and

  • 01:10:18 this is the base path I am going to change this  into this. Okay. Let me zoom in. Okay can I zoom

  • 01:10:25 in? No. So this is the path where I want ComfyUI  to read model files save it. Okay it is not saved

  • 01:10:34 yet it is saving right now. Okay it is saved.  Then copy this and paste into ComfyUI folder.

  • 01:10:42 Then return back to Massed Compute instructions.  I am first going to show ComfyUI usage then we

  • 01:10:48 will see in SwarmUI. Okay then we are going to  use this command to start this one so copy this

  • 01:10:58 open a new terminal inside this folder ComfyUI  version 78. This is our main installation folder

  • 01:11:06 paste. And it will start the ComfyUI. You can  see the messages in here. Always look for the

  • 01:11:14 CMD to see the accurate messages. You see it is  loading everything properly with torch 2.9.1 and

  • 01:11:22 CUDA 13. It is loaded. How to access this from our  computer? For accessing it I am going to use this

  • 01:11:31 URL copy link and in here type 8888 like this and  hit enter. Because it has started in this port as

  • 01:11:41 you can see oh actually this is the port not 8888  so let's make it like this yes. So now I am able

  • 01:11:49 to access the ComfyUI running in Massed Compute  from my Windows computer. The rest is same as in

  • 01:11:58 the Windows tutorial part I am going to just  load my preset for example audio to lip sync.

  • 01:12:05 Yes perfectly loaded this is running on the Massed  Compute. I am going to choose my demo files like

  • 01:12:13 this one. Let's choose our audio file this one.  Then set this as in the Windows 45 seconds 1081

  • 01:12:25 frames and I need to type my prompt as usual  this is my prompt and run generate. Now let's

  • 01:12:36 see the speed of beginning the model loading  and everything. Okay it is already loaded.

  • 01:12:43 It is going to start processing in a moment.  The speed of Massed Compute is unchallenged.

  • 01:12:49 Not SimplePod not RunPod nothing can match the  disk speed of the Massed Compute it already

  • 01:12:55 started generating so the whole process takes  like 5 minutes once you done it. You will get

  • 01:13:03 used to it you can use the Massed Compute all the  time if you don't have a powerful GPU this is one

  • 01:13:08 of my recommended cloud server provider. So we  are almost done to get 45 seconds HD video. It

  • 01:13:18 is taking like 1 minutes lesser than 1 minutes to  generate. We can see the whole process live while

  • 01:13:25 I am recording. Everything is so transparent.  So it is now VAE decoding. Let's see the output

  • 01:13:33 in here so it is going to do video combine this  is running on cloud not on my machine. But I am

  • 01:13:39 using it from my computer. And it is almost  ready. Yes. Yes it is ready let's see it.

  • 01:13:55 So you can right click and save as  the video like right click and save

  • 01:14:00 preview or it is also saved in the outputs  folder inside ComfyUI inside output so the

  • 01:14:09 video is here I can copy this and move it  into ThinLinc client you see it is here

  • 01:14:16 this is with audio. So save preview will also  download the video into your computer as well.

  • 01:14:26 So how to use SwarmUI with this ComfyUI  installation in Massed Compute? I am going to

  • 01:14:32 close my ComfyUI and Ctrl Alt D it will minimize  everything. You see we have RunPod stable SwarmUI

  • 01:14:41 update this is actually latest version of SwarmUI  and it will update it alternatively you can also

  • 01:14:47 use the new installer from here but this is  working pretty good so it will update SwarmUI

  • 01:14:56 then go to server backends and change this.  You should use new installed ComfyUI always

  • 01:15:02 so it is here ComfyUI Ctrl L it will select the  path Ctrl C and edit this and paste it and then

  • 01:15:12 main.py then in here this is like this main.py  then in here --use-sage-attention and save. We are

  • 01:15:23 ready but we need to check the models are already  here since I have downloaded them already into the

  • 01:15:54 SwarmUI folder I need to check the Lora folder  where it is. So it is Lora here. Choose file.

  • 01:16:00 The presets are here amazing SwarmUI  presets import refresh sort by name.

  • 01:16:07 Now I can use the preset that I want for  example LTX 2 text to video direct apply

  • 01:16:16 always quick tools reset params to default then  direct apply. Then type your prompt a spaceship

  • 01:16:23 going in space with cinematic background music.  Okay let's select our aspect ratio and generate.

  • 01:16:33 So the usage is exactly same as in the Windows  I am not repeating in details please do not

  • 01:16:39 skip those parts. But this is running on the  cloud machine you can see that it will start

  • 01:16:46 the processing immediately on Massed Compute.  This is fast and I can generate very long videos

  • 01:16:53 this way advanced options text to video. I can  make this 481 frames let's play this one.

  • 01:17:02 So you see it is working amazing.  Please like and subscribe. Join

  • 01:17:07 our Patreon to support us and get  these amazing presets. See you later.

Clone this wiki locally