00:02Welcome to another tutorial

00:05Stable Diffusion 3 has just been released, and today I’ll show you how to install it and start using it immediately on two interfaces

00:07The first interface is Stable Diffusion WebUI, and the second one is ComfyUI, so you can choose the one you prefer

00:12Let’s begin with the Stable Diffusion WebUI

00:16I’ll provide the links in the description

00:21It’s made by Stability AI, and you can find it on GitHub

00:24Currently, it’s still in beta

00:28You can install it on various operating systems, but since I have Windows, I’ll demonstrate how to install it on that

00:32If you’re using Windows 10, you’ll need to manually install Git and .NET 8 first

00:41However, if you’re using Windows 11, this process is done automatically

00:45Click on the provided link to download

00:47A batch file, after downloading, create a new folder

00:50Remember not to place it in the Program Files folder; instead, choose a location like your D drive, create a new folder there, and name it whatever you prefer

00:57Let's go with "swarm" for now

01:00Once the folder is created, move the downloaded batch file into it

01:04Now, navigate to the folder you created earlier and double-click to open it

01:12Windows might prevent it from running, but you can click on "More info" and then select "Run anyway." If you're concerned about security, you can scan the file for viruses online, but rest assured, this file is safe

01:25Now, it's downloading .NET version 8, which is around 220 megabytes in size

01:31I'll fast-forward the video during this download process since it might take some time

01:35Remember not to close any Windows until the download is complete

01:39An installer window will appear and will begin installing what was downloaded earlier

01:44Once the installation is complete, it will proceed to download or update Git

01:49After that, Git will be installed on your computer, followed by some configuration steps

01:53Once everything is done, you'll finally get the Swarm UI installer

01:59At this point, you'll encounter a legal notice

02:01The Swarm UI is under the MIT license, but the models like Stable Diffusion 3 are only for personal use

02:09For commercial use, you'll need to purchase a license

02:11Click on the Agree button to proceed

02:13At this stage, you have the option to simply click on Install, but I'll choose to customize

02:19Settings to explore the available options, and after that, I'll click Next to proceed with the installation

02:25Here, you can choose a theme that, um, suits your preference

02:27I think I'll go with, uh, Cyber, then I'll click Next

02:33Here, you'll find more options for advanced users, but I'll stick with the default settings and click Next

02:38Moving on, you'll encounter options for the ComfyUI

02:44I'll also go with the default ComfyUI

02:46For those with an API key, there's a second option available, and the last option is for custom backend configurations

02:56Now, you'll reach the part where you can choose which models to download

03:00Since all the available models are base models that aren't very good, I won't select any of them

03:04Instead, I'll download my own model

03:07The final step is to confirm by clicking on "Yes, I am sure" and then click "Install Now." Once clicked, you'll need to wait for the installation to complete, and in the command window, you can monitor the progress

03:19At the end of the process, you'll be greeted with the interface

03:22Note that the top bar will disappear after loading is complete

03:30Let's check the installation folder to see what we have there

03:32If you want to update the StableSwarmUI, you can click on "Update" or the appropriate option for your operating system

03:42Similarly, if you want to launch the interface, you would select "Launch," again choosing the appropriate option for your operating system

03:50To close the existing StableSwarmUI, simply close the command window

03:54Now, if you click on the launch.bat file, it should successfully launch the interface

04:03What I usually do is right-click on the batch file and create a shortcut

04:05Then I rename that shortcut to Stable Diffusion UI

04:08After that, I right-click on the shortcut, select Properties, and choose Change Icon

04:15If it says there are no icons available, I click OK, and then it presents me with some system icons

04:24From there, I can choose any icon I like, such as a star, for example

04:29Click OK, then Apply, and finally OK again

04:34You can then copy and paste that shortcut onto your desktop for easy access

04:40If you double-click on the shortcut, you'll see that it runs the interface smoothly

04:45Now, let's get those Stable Diffusion 3 models

04:48You'll need to go to the..

04:51Hugging Face website, links provided in the description

04:54Once there, you'll need to log in

04:56If you don't have an account, simply click on register, just like I'm doing now

05:01You'll be prompted to add your email and choose a password, then click next

05:07Next, enter your username, full name, check the "I agree" box, and finally create your account

05:16It asked me to confirm that I am human; these CAPTCHAs are getting more interesting

05:22Then you check your email for the confirmation email and click on the link you received

05:26Once you do that, you'll get a message confirming that you are verified

05:30Now, on the download page, you need to enter your name, email, country, and company

05:36You can also choose whether you want to receive email updates

05:40To check the box that says, "I acknowledge that the model is for non-commercial use," and then hit the "Agree" button on the model card tab

05:52You can read more about the models; however, you need to click on the "Files and Versions" tab to actually get the models

06:00The smallest one available is the SD3 Medium

06:03If you're using other interfaces, you can download the second one, but I will get the 4.34 GB version

06:07After downloading, you'll need to place it in the "models" folder and then in the "Stable Diffusion" folder

06:16Click "Save" and wait for it to download

06:19Next, we navigate to the AUTOGENESIS Swarm UI interface and click on the "Models" tab

06:27Here, you'll find a refresh button, which you can use to ensure the interface shows the model you just downloaded

06:32Select the model from the dropdown menu, and in this case it should say SD3 Medium

06:38If you choose None, it will deselect that model

06:43Another method to load the model is by clicking on the model thumbnail, specifically on the image preview, not on the text

06:52Once clicked, you'll see that it's now loaded

06:57Let's see if it works, and you can choose from resolution options, a ratio, or input a custom ratio if you prefer

07:03Then type a prompt of what you want to see, such as a cute cartoon bunny in a forest

07:07After that, hit Generate

07:11Please note that only the first time you have to wait

07:16You'll notice that it downloads a file called CLIP G and another one called CLIP L

07:24Once it's ready, it will generate the image based on your prompt

07:30Okay, for the first try, you know these CLIP models are, uh, text encoders and can be found in the Text Encoders folder

07:37You can see that, uh, both models are here, and it's nice that StableSwarmUI downloads them for you

07:43You know what it did was place them in the Models folder, inside the CLIP folder

07:51You can change the seed from here; like, if I click on 'Reuse,' I get the seed of the prompt I just used

07:56If I click on 'Random,' it's random

08:00The nice thing is that every setting has a question mark, so you can see what it does

08:05Let's see, uh, what settings are recommended for this model

08:09I messed up some settings in the recording, but these are the settings you should use: for steps, use 28 steps; for CFG scale, use 4.5

08:19Then go to Sampling and enable the sampler

08:23Choose DPM Plus+ 2M for the sampler and for the scheduler, select SGM Uniform

08:28Hit generate to test it out

08:33Now you can enable this option with text encodings

08:38Here, you can choose from different options depending on how good your video card is

08:41Keep in mind that, um, this can use more RAM

08:45From what I understand, I'll choose one option

08:48When I hit generate, you'll see that the first time it downloads another model called T5 in the CLIP folder, or you can choose both CLIP and T5

09:04I'll have to experiment more with those options in the next few days

09:07If you check the licenses on Stability AI, you'll see that it's free for non-commercial use; however, it costs $20 for commercial use

09:16So, see which license works best for you every time you hit

09:22The generated image is saved in the output folder, labeled with the date of creation

09:31You know, you can create a shortcut to that folder for easy access

09:36You can download more models from the Civit AI website; click on Models, then navigate to the Filters section, where you can sort them by popularity, such as the most popular in the last month

09:48I personally like to use SafeTensor models, among which the most popular ones are the SDXL and v1.5 models

09:56One that I frequently use is JuggerNAUT XL, and in this case, version 1.0

10:02The Hyper version is faster for those with lower VRAM, but I prefer the results of the X version, so I click download and put it in the same folder under Models and Stable Diffusion while it downloads

10:19You can check the recommended settings for each model so you know what settings to use when you utilize this model

10:25The one I downloaded is the standard model

10:32For aspect ratio, you can use a custom ratio and input the exact width and height if you want, but I will go with the default size in this case

10:38For samplers, I use DPM Plus+ and the scheduler is Karras

10:44I'll also turn off the SD3 encoding

10:47For steps, it says 30-40, so I'll use 30 since it's a bit faster

10:53For the CFG scale, it's recommended between 7 and 12, so I'll use 5

10:56You can check the info on how the CFG influences the generation

11:01If your video card isn't so powerful, you can try using a smaller model, like an older version of the ReVived Juggernaut, such as the Reborn 1.5 model

11:07I'll download it and..

11:13Place it in the same folder with the rest of the Stable Diffusion models

11:19While it's still downloading, let me show you some other features

11:21You have a Comfy Workflow tab where you can perform more advanced tasks with ComfyUI right here, but in a few minutes, I can show you how to install only ComfyUI if you prefer that

11:34From the Generate tab, the default window is where you generate images, and on the Simple tab, you'll find some simple workflows

11:45Additionally, you have the Utility tab and a User tab

11:47On the Server tab, you can access server info, where you can see CPU and graphics card usage

11:56In the Server Configuration, you can select your root folders

11:59Right now, it's set to the Model route in the models folder, which means this folder from here

12:07You'll see all those folders that you have inside with the name for each

12:14The models are now downloaded, and we can test them

12:16If your interface is open, you'll need to click refresh to be able to see them

12:20You can see information about the type and what resolution they are trained on

12:26Click on the model thumbnail to load it

12:29I will test all three models using a fixed seed with the settings already set for the Juggernaut XL model

12:33I'll add a prompt generated with ChatGPT and then hit generate, and the result is quite nice

12:41Next, let's test the smaller one, version 1.5, which can handle smaller images

12:47The result looks like this

12:50Finally, for SD 3, I'll input the recommended settings for that model

12:53To get the best quality, I'll enable..

12:55All I can say is, the result is this, and here is a side-by-side comparison

13:03Let's move on to the second part, where I show you how to install ComfyUI

13:07You'll find the link in the description

13:09Scroll down until you find the section on installing

13:12Since I have Windows, I'll click on the direct download link

13:14Then, in my D drive, I'll create a folder named ComfyUI—you can give it any name you want—and save that ZIP file inside that folder

13:23Go to that folder and wait for the download to finish, then extract the contents of that archive

13:28It might take a while

13:32Once extracted, navigate inside the new folder

13:37You can read the README file for any additional instructions

13:39Assuming you have an NVIDIA GPU, click on the Run Nvidia GPU Docker file

13:46A command window will open where you can see some..

13:48Settings and information about your video card, then ComfyUI will be ready

13:52This is how the interface looks

13:57I also recommend getting the ComfyUI Manager from this page

14:00You can scroll down to the installation section, and you'll see it has three steps

14:04I'll copy the text from Step number two, and then I'll start with the first step

14:09Go to the ComfyUI folder and navigate to the 'custom nodes' folder

14:14Then, in the address bar, type 'CMD' and press Enter

14:18This will open a command window in that exact folder

14:22Here, you paste the text and press Enter

14:25This will use Git to clone that manager into your folder

14:29Once it's ready, you can close that window

14:31Let's go back to the folder where we installed ComfyUI

14:36I'll create a shortcut of the 'Run' script

14:38Nvidia, at file, then I'll rename it to ComfyUI with a right click

14:41I'll change the icon, maybe to a butterfly, and then click OK and Apply

14:47Now we have this shortcut that we can paste on the desktop so we don't have to enter the folder each time we want to launch ComfyUI

14:57Close any ComfyUI window you have open, and now let's use the shortcut to run it

15:03You'll notice there's an extra button called Manager

15:06Now click on that button, and you will see more options

15:10From here, you can install custom nodes if you import a workflow from the internet, and if some nodes are missing, you can click on Install Missing Custom Nodes

15:20You can also install models, update all custom nodes and ComfyUI, or you can choose to update only ComfyUI

15:27Additionally, you have a..

15:30Restart button here

15:32Now that we have the interface, let's get the Stable Diffusion 3 models

15:34Go back to minute 4:47 of this video to see how I logged in and what information it asked me for to access the models

15:42On the Files and Versions tab, you can find the models

15:45I recommend getting the SD3 Medium model and the second one, which includes CLIP

15:50Download them and place them in the COMFYUI folder

15:56Then go to the Models folder, then Checkpoints

15:59Here, you'll place all the Stable Diffusion models you get from the internet

16:05I'll also add that second model

16:10You'll also need some text encoders, so go to the Text Encoders folder, download the CLIP G model, and place it into the Models and CLIP folder

16:25Do the same for the second one, the CLIP L model

16:28Then, repeat the process for the last one with FP8 in the name

16:34If we go back, you can also find a folder with ComfyUI example workflows

16:38I'll download the first one, the basic one, and I'll create a folder named Workflow on my drive just to be more organized, but you can place it anywhere you like

16:49While the models are still downloading, you know, you can explore the SD3 demo prompts to see examples of prompts you can use

17:01And now that the models are downloaded, I'll go back to the manager and restart ComfyUI

17:08Click on the Load button and then navigate to the folder where you saved the workflow

17:14Open it, and that will load that specific workflow with all the nodes

17:17If there are no red nodes, then it means no nodes are missing

17:22If one was red, we can go

17:24To the manager, click on Install Missing Nodes

17:27Let's see if it works

17:30Click on the Q prompt button

17:33We got an error: value not in the list

17:36Checkpoint SD3 Medium

17:40Uh, but we downloaded that model, so let's select it manually from the list

17:44As you can see, the node with the problem has a red contour

17:47You can use the mouse wheel to zoom in and left-click on the workspace to pan around

17:54Here, on the checkpoint name, it appears to be in a different location

18:02I'll just select the SD3 Medium from the list and try to generate again using the Q prompt button

18:09As you can see, now the nodes have a green contour, and you can follow to see how it goes from node to node, each one turning green when it's active

18:18It's like water flowing through pipes until it hits… A wall, and then it will make that node red, and now everything worked perfectly, and we got our first Stable Diffusion 3 generation

18:30The first node lets you load a checkpoint, which can be SD3, SDXL, and so on, depending on the workflow

18:35SD3 uses these CLIP models

18:39Then you have the prompt area where you write your prompts, and under that is the negative prompt, where you write what you don't want to appear

18:49From here, you can also set the width and height and how many images you want to generate at once

18:56In this workflow, the C is fixed, which means you get the same image for that prompt each time

19:01So I want to make it random, so I'll use the "Randomize" option from the dropdown menu

19:08Then, for the K sampler, the settings depend on the model each time

19:13Works better with certain settings for the SD3 model; these are the recommended settings

19:21Lastly, this is the node for previewing the image

19:24You can also use a Save Image node if you want

19:27Additionally, you can right-click on the image and save it

19:32Let's generate another image now with the random seed

19:35As you can see, it's working, and this is the result—quite cool

19:39Let's do it again and check the command window to see how many seconds it takes

19:43It seems that on my RTX 4090, it takes around 5 to 6 seconds

19:48You can select nodes and move them around so you don't have to scroll so much and can fit more in your window, or you can put the nodes you use more often closer together

20:00Now it fits a little better

20:02Let's test another prompt from their..

20:05Like this one with the tiger, I'll put it in the prompt area and generate again

20:10Nice, it almost got the open source text on the flag right

20:12Let's see if it still works if I replace the name with words that are not actually in the dictionary, like my channel name, Pixaroma

20:21As you can see, it struggles with that name, but after a few tries, it can get one right

20:27If you're happy with the workflow, click the save button and save it so you can load it later

20:34If you want to get more models, you can go to the CivitAI website, click on the Models tab, and filter them by model

20:40The most popular ones are the older v1.5 models, then SDXL models, and now the latest that appeared is SD3

20:50I will just search for one that I use more often, like Juggernaut XL

20:59You can choose either the Hyper version, which is faster, or the v1.0 version that has better quality

21:05For me, you should always check here for the base number, click on the download button, then go to the Stable Diffusion UI folder, then models and checkpoints — just where we put the other SD3 models

21:24I forgot to test the second SD3 model; we downloaded the bigger one with CLIP included

21:32I'll select it from the list and then generate again

21:37Nice, this time it got the text right

21:40I'm not sure if it was the seed or just the model is better

21:43Now let's test that Juggernaut SDXL model

21:45If you click on Manager, then where it says Workflow Gallery, click on the arrow and select one like OpenArt AI

21:54That will open their website, some basic ones that you..

21:59You might want to try, which can be found in the workflow templates here

22:00You can find one that works with SDXL, for example, this one with an owl

22:07It says "Basic SDXL Workflow." It has only 13 nodes

22:14Click on it, then download it to your computer

22:17I'll put it next to the other one I downloaded, just to stay organized

22:22Make sure you save the existing workflow using 'Save' before importing a new one

22:28Click on 'Load' and select the Basic SDXL Workflow you just downloaded

22:34As you can see, this one has different nodes

22:36I will add the SDXL one here, but it doesn't appear in the list since the interface was open when I downloaded it

22:46So I have to click 'Refresh' first to see the new models

22:49Now I select the Juggernaut X version, and then, in the recommended..

22:56Settings on Civit AI

22:59I look at the model and add those settings here

23:02First, the seed was fixed, so I changed it to randomize

23:07For steps, it says 30 to 40, so I chose 30, and so on for the rest of the settings until I completed them all

23:16I click generate to see if it works

23:20I watch the nodes activate with green, and yes, no errors

23:24And we got our first SDXL generation

23:27This one has the last node as "Save Image," so if we go to the ComfyUI folder and the output folder, we can find all the generations there

23:40You can also right-click on the image and save it wherever you want

23:43If you open a workflow and the nodes are filled with red color and give you an error that something is missing, you can go to the manager and click on "Install Missing Nodes." You will see a..

23:55Here, I list the steps for installing on the right, and you can install them one by one until you fix the node error

24:00If you go back to the workflow gallery and click on all workflows, you can find many workflows, but some have over 100 nodes, making it hard to figure out all the settings and errors

24:15While missing nodes are most of the time fixed, some settings and models are missing from these workflows unless they have a step-by-step tutorial

24:26It's better to play around with simple workflows first and try to build your own workflow

24:32Another problem is that having too many custom nodes installed can create conflicts with other nodes

24:38I have a Discord server now; you can join using the link on my YouTube channel

24:42There, I post..

24:45News resources and discussions about different AI tools

24:49You are welcome to join

24:52You can also share your AI-generated images, and maybe we can learn from each other

24:57I also have another YouTube channel where I mainly post short videos with AI experiments

25:09The channel is called AI Toley

25:11If you find something useful, please leave a like

25:13Thank you

25:20[Music]