00:04Greetings everyone
00:09Today I am going to show you the latest method, the most accurate and the most performant way of generating images and videos on your Windows local computer with one-click installation and setup
00:14I am going to demonstrate how to use SwarmUI and how to download models and start using them immediately on your Windows computer
00:26This will be an epic tutorial that will make your local image and video generation many times easier and with maximum optimization
00:36Maybe you've heard of COMFYUI
00:38COMFYUI is quite difficult to use, but we can still leverage its performance and features with SwarmUI
00:48My installation even supports out-of-the-box support for Flash Attention
00:51Maybe you've also heard of it
00:54It is the fastest
00:57Attention that makes your generations way faster for both image and video creation
01:02And I will introduce you to the application I have developed for downloading models with maximum speed and ease
01:10So far, we have two model bundles: the Flux model bundle and the Hydream L1 model bundle
01:18With just a click of these buttons, it will download all the necessary files into your Swarm UI installation
01:26Moreover, we have all kinds of models available for one-click download, so you won't spend time searching for models or downloading them into specific folders
01:34Additionally, I will show you how to use links
01:39When you go to this link, you will learn how to use this application and these models within the Swarm UI
01:42You see, we support..
01:44Hydream
01:47We support Stable Diffusion 1.5 models
01:50We support Stable Diffusion XL models
01:52We support Stable Diffusion 3.5 large models
01:58We support Inpainting models
02:03We support Text Encoder models, T5XL, UMT, T5XL clip models, LLM text encoder
02:08We support video generation models 12.1, Huion FastGAN models
02:11We support SkyASL and, as you can see, we support GGUF and also BF16 and FP16, FP8 versions
02:24We support Gamma Motion, Litrix LTX, Video VAE models, and this application will get better and better as you make requests to me
02:33So, you can tell me to add this model, add that model, and hopefully I will update it
02:39At the very top of the application, you can also search and filter by model name, so it is so
02:45So easy to use and so easy to install
02:48Using Swarm UI with a manually installed ComfyUI backend is working better than the pre-compiled ComfyUI backend
02:57So, today I will also show you how to use a properly installed ComfyUI backend in Swarm UI for maximum performance
03:05By the way, these images that you are seeing right now are being generated with Hydram-L1 Dev GGUF Q8 model
03:13GGUF Q8 is almost the same quality as the highest-quality BF16 model itself
03:22Currently, I am generating with my dual GPUs
03:24I have an RTX 5090 and an RTX 3090 Ti, so you see both GPUs are running almost at full power
03:35Also, I am recording a video right now, so you see my PC is under heavy load
03:40However, Swarm UI with ComfyUI backend is still able to handle it
03:44Generations are really, really fast
03:47They are being generated right now, live, as you are seeing
03:51So let's begin the tutorial on how to install and use these applications
03:56We have the SwarmUI one-click installer with the SwarmUI Unified Model Downloader, but this Unified Model Downloader can also be used for ComfyUI or other applications that you use as well
04:09It supports custom folder paths and also ComfyUI installers
04:14The links to these installers will be in the description of the video, so check them out
04:18First of all, let's start by downloading the latest SwarmUI Model Downloader ZIP file
04:24The latest ZIP file is SwarmUI Model Downloader version 29
04:27I just updated it to the latest version
04:31It has been downloaded
04:33Let's also download the manual installer of ComfyUI
04:37By manual, I mean that we are not using a pre-compiled version
04:39We are going to install it with our installers
04:41Move the downloaded files into any of your drives; I will use my E drive
04:47First of all, let's extract the ComfyUI installation
04:51This will be used as a backend
04:55And why not use the pre-compiled backend? Because this is way better
04:59When you use the pre-compiled backend, it is not very up to date, which can cause problems when installing new libraries
05:06So I will use the Windows install.bat file
05:08This by default supports XFormers, Flash Attention, Sagan Attention, Triton, DeepSpeed, whatever you need, and it works on all GPUs, including RTX 5000 series
05:20I recommend using Python 3.10
05:24You need to have..
05:27Python is installed on your computer
05:30So, if you are using another version, you can select it
05:32I really recommend Python 3.10, so I will select option one
05:35You can choose Flash Attention, Saga Attention, or both attentions
05:39Let's choose both attentions, and it will also install XFormers as well
05:47This is going to install with the latest official Torch 2.7
05:49Don't worry, it will work on all GPUs, and also pre-compiled Flash Attention, Saga Attention, and XFormers
06:01I have compiled XFormers, Flash Attention, and Saga Attention myself, based on our legendary WT0RD Ho
06:07Maybe you've heard of this guy; he is the sole developer of Saga Attention for Windows
06:15Meanwhile, this is getting installed, we can also install SwarmUI
06:17We don't need to wait
06:20So let's just extract the Swarm UI model downloader
06:23Extract all and here
06:25So let's begin with installing Swarm UI
06:28You see, we have the Windows installer swarm UI.bot file
06:30This is from the official repository
06:32Nothing is special here
06:35It will just start the installation like this
06:37The options that you are going to choose are important
06:39We are going to choose Agree, customize settings
06:41Yes, modern dark or whatever you wish, just yourself
06:48And this is the most crucial part
06:51We are not going to use the ConfUI local
06:53This is pre-installed, pre-compiled ConfUI
06:55I do not recommend this
06:58So let's go with None, and next
07:01You can download, but I don't recommend it
07:02Our downloader is faster
07:03Next
07:05And yes, I am sure
07:06To install and set it up, that's it
07:08So, the SwarmUI has been installed
07:10Let's look at the configuration
07:12It is still installing, as you are seeing right now
07:15Alright, the configuration has been installed
07:18You see, press any key to continue, and the SwarmUI has been started
07:20Now, what we need to do is add the backend
07:23So, click on Server, then click Backends, and go to this page, and click Configure Startup
07:29Click OK
07:31It shows two backends
07:33This is a bug
07:35Let's refresh, and it is fixed
07:36So, let's check this
07:38Click the green icon for editing
07:40We need to give the path of the configuration here
07:43So, let's return back to our configuration installation
07:46It was located here
07:48Enter the folder of the configuration, and you see there is a main.py file here
07:53So, let's copy the path like this
07:55/install.py file
07:58This is the back end
08:01If you want to use scaled attention, use Sage Attention like this
08:04Sage Attention works only with RTX 3000 and newer GPU models, such as the 3060, 4070, or 5090
08:07So, the 3000, 4000, and 5000 series are supported
08:16But don't worry if your GPU is not supporting it; it will use either XFormers or the torch attention itself
08:24So, wait for the back ends to be loaded
08:27When we go to the logs and debug, it will show us what is happening right now
08:32It is installing the missing libraries into our manually installed config, and our config has been set up with a virtual environment
08:41This installation will not break any other applications or installations that I have
08:48This is 100% isolated to here, so it is completely safe
08:53Secure
08:58With this setup, you can have as many different configuration installations as you like, and we will use the model folder of the SwarmUI itself, or you can specify any model folders
09:03How? When you go to the server configuration, you can set your model root and other model paths from here
09:06Okay
09:10So, the installation has been completed
09:13I have my backend ready
09:15Since I also have another GPU, let me show you
09:18So, open a command prompt, type `pip install torch`, then type `nvidia-smi`, and you will see the details
09:21By the way
09:23Yes
09:26So, let's start again
09:28`nvidia-smi`, and here
09:31I have an RTX 4090 and an RTX 3090 Ti, so I can also utilize my second GPU
09:39To do that, let's copy this, click configure to start
09:41Okay
09:43And paste it here
09:45So, copy this, paste here, and give the..
09:47GPU ID one and save
09:49So, this will work on the second GPU, and this will work on the first GPU
09:54We are ready; our SwarmUI has been started
09:56However, currently, we do not have any models
09:58To download the models, let's return to our SwarmUI folder where we have extracted the files
10:06You see here, click Windows Start, then download models app.bot file for more info
10:10Anyway, it will generate a VENV (virtual environment) folder here, where it will install the necessary libraries, which are not much—just GRADUATE and some other libraries
10:21Then, it will start the application
10:23Let's just wait for this to work
10:24You also need to have Python installed, so if you don't know how to install Python and the other libraries, I have an excellent tutorial video
10:33The link will be in the config installation post, as you are seeing right now
10:38Windows requirements, so you can watch this tutorial and learn it with full details and step-by-step guidance
10:44So, the model downloader has started
10:47This is the model downloader interface
10:50Whichever model you want to use, you can search here and download
10:53This is an extremely advanced UI
10:55And also, if a model is missing here, just message me on Patreon, and hopefully I will update it as soon as possible
11:02 For demonstration, we can download the Lux models bundle
11:05It will download all these models
11:07So, let's click "Download Lux models bundle," and you can watch the status here
11:12It will utilize my entire network right now
11:14Why? Because we are… Utilizing enable HV transfer
11:19This is a special download method, and it will speed up your downloads to your maximum network speed
11:24Let me show you from Task Manager performance, and you see I am downloading at 500 Mbps, like 600 Mbps
11:32I am also uploading models into more optimized Hugging Face repositories, so that it will be even faster, hopefully, in the future, without any errors
11:48Moreover, this downloader has a queue system, and if a file has been downloaded previously, it will not redownload it
11:53So, this is an amazing feature
11:59For example, I can also queue video generation models for download, like the 1.2.1 models
12:04So I can click this, and it will queue that
12:06Let's see
12:08So, you see, the queue size is now seven
12:09And it shows that cute 12.1 Text-to-Video model
12:12Moreover, if you don't know how to use these models inside SwarmUI, just click these links and read them
12:20For example, this is one 2.1 parameters in Swarm UI
12:24You can read here and learn it
12:28Hopefully, I will make different tutorial videos, so stay subscribed to our channel to not miss any of them
12:33This application works on Rampot and also on multiple compute setups as well
12:38The instructions are all in the zip file: MassCompute instructions, Rampot instructions
12:43Hopefully, I will make a separate video for them so as not to make this video very long
12:49So stay subscribed, and this will download all of the models with maximum speed
12:54Then you will be able to use models inside your SwarmUI
12:57The models will be automatically downloaded into the correct folders, so you don't need to spend time setting them up
13:06It will show you everything—every piece of information on the screen and also in the terminal window on the CMD
13:13If you have any questions, just ask me, and I will hopefully answer them
13:15Using these models is very simple
13:18Let me start my other installation, which I have been showing you at the beginning of the video
13:25To start SwarmUI after installation and the models have been downloaded, just click the Windows Start menu, then select SwarmUI.bot
13:32SwarmUI now updates automatically, so there is no need for an update button anymore
13:41Once the models have been downloaded, go to the Models section, click Refresh, and select the..
13:48model that you want
13:55One of the amazing features of our downloader is that we download the highest quality T5XXL models, and we give you the option to download the highest quality models as well, like the FluxDev FP16 model, or you can also download the FluxDev GGUF model
14:02When you first try to run a GGUF model, it will ask you to install the GGUF extension
14:09It is important, and also, in some models, it may require you to set the model type
14:15For example, this is automatically recognized Hydream L1 Dev GGUF Q8 model
14:21So, when I click this hamburger menu and select "Edit Metadata," you can set it from here, like Hydream architecture
14:31It usually recognizes all of the models
14:34But if it doesn't recognize a model, select the architecture from here and save it, then it will recognize the model correctly
14:41Let me also show you the gguf installation so you will not have any issues with that
14:46To do that, I will just copy my previous model to avoid waiting for it to download
14:52The models are usually saved inside the diffusion models folder
14:56So this model is saved inside the diffusion models
14:59Let's copy this and go back to our new installation
15:03Just copy and paste it
15:06You see, our download is continuing
15:09It downloaded the Flux Dev model, now downloading the Flux Field Dev model, and returning to our new installation
15:12It is here, and the models are here
15:14So, let's refresh
15:16Yes, they have arrived
15:18When I click "Load Model," it will ask me to install gguf support
15:19You have to install this to be able to run gguf
15:25So, click "Install." Okay, you need to wait for the backends to load
15:29Then it says, install it
15:32Please wait while the backend is restarting
15:34If it doesn't work, you may need to restart SwarmUI, but it usually works
15:38Just follow the logs from the debug menu, and you will see that it is installing necessary libraries and starting
15:45After it has been started, it will load the model into the VRAM of my GPU
15:48You may also see this error: "test CLI link fatal error." This is a Linux-related error since we are running on Windows
16:00You can just ignore this
16:02It happens in my older applications, but you can safely ignore it, and the backend has been loaded and is ready to use right now with GGUF support as well
16:12You can also install Tcash and other extras, but hopefully I will make different separate videos
16:17For them
16:19So, this is how you can begin using the very best models on your local computer after installing and downloading them
16:27For example, this is a DreamBooth model that I have trained myself on
16:31And let's generate an image as a demonstration before we end the tutorial
16:33So, we have an amazing Coloia configuration
16:36It is also available on Patreon
16:38Let's go to the test prompts
16:40Let's select a realism prompt
16:42For example, this man
16:44This is a model that has been generated based on me
16:47Since this is the Flux model, let's set the 40 steps sampling IP and DM
16:51My favorite sampling CFG is 1
16:54This is for Flux, and DreamShaper is also the same
16:56Flux-guided scale 4
16:58And that's it
17:00Let's generate an image first
17:03It will load into my RTX 4090
17:05Then
17:07It will generate the image
17:08Okay
17:10Oh, we did set it to 100 images
17:11Okay
17:13Let's see
17:15And we have the two backends here, so it will use both of them
17:17All right
17:18It should start loading in a moment
17:20You can always watch the logs from here
17:22Let's go to debug, and let's see what is happening
17:25Okay
17:27It says that it is waiting for the model to be loaded
17:29Yes, now it is loading
17:30We can watch the status from here
17:32Yes, it is loading now, first into the RAM
17:36Then it will move them into the GPUs
17:39Okay, the first GPU is warming up, loading the model
17:43Yes, the generation has been started
17:44We can see that
17:46Okay, let's see
17:47The image is getting generated, then it will inpaint the face
17:49I have been getting asked if the inpainting model is not working
17:54Because your installation was not accurate, install it as I have done in this tutorial, then download the face models
18:01For example, you can, let's refresh this page
18:03It shouldn't impact the download
18:06You see, it is returning it back
18:08Go to the other models, and you see we have AutoYolo masking models
18:13Then click here and just download
18:16It will download the best face automatic recognition segmentation models, and you see it has generated the image
18:21Oh, there is one more thing I need to show you
18:26Let's cancel all sessions, which is upscaling
18:29This was also something I was asked about: what is the best upscaling? The best upscaling is like this
18:34Let me show you, so go to the Refine upscale
18:37Enable it first
18:40If you don't see advanced options, Options, click here to see the advanced options
18:44Set the refiner percentage to 60%
18:46Let's click reuse parameters to upscale this image
18:49Enable upscale refiner to 60% and refiner upscale 2x
18:52And this is super important
18:55I use the 4x RealESRGAN version 4
18:58This is automatically downloaded with our downloader, where you see we have other models here
19:06When you click the other models and go to the image upscaling models, you can download all of the image upscaling models by clicking here, and it will automatically download them into the appropriate folder
19:18Now, first, it will generate this image, then it will upscale it
19:20Upscaling is really heavy because we are upscaling from 1 megapixel to 4 megapixels
19:29With Refiner, upscale 2x, but this is really, really next level
19:33It really improves the quality
19:35I will show you in a moment
19:38Okay, the upscale has been completed
19:39Okay, now watch carefully
19:40Original image
19:42Upscale image
19:43Original image
19:45Upscale image
19:46You see the difference is huge
19:47And probably I don't even need to do face inpainting this
19:52You see, the face inpainting has some noticeable areas here because it is black
19:57So, to remove these noticeable areas, you need to adjust the parameters here, like Segmentation Mask Blur, Segmentation Mask Grow, and other settings to remove it
20:06I don't know the exact parameters right now; I don't remember, but you can pretty much remove this as well
20:10It is barely noticeable, but and also, it can be really..
20:14Easily fix it, but from this quality into this quality, and this is amazing
20:17This is next level
20:19Thank you so much, guys
20:21You can always message me on Patreon
20:23You can also reply to this video, hopefully
20:25See you later
20:26The links will be in the description of the video and also in the pinned comment of the.