00:00Welcome to episode two of our Comfy UI tutorial series

00:02In this video, I will show you the basics of nodes and workflows, helping you get started with building your own projects

00:08Even if you don't plan to create your own workflows, maybe you download workflows from others and need to make changes, so it can be useful to have an idea of how things work

00:20In episode one, I demonstrated how to install and generate your first image using Stable Diffusion

00:26If you haven't installed Comfy UI or downloaded the model yet, please check out that episode first

00:33This is a basic Stable Diffusion workflow that can generate an image from text, often referred to as text-to-image

00:38Uh, to help you understand how it works more easily, I'll provide a simple

00:46analogy that will give you an idea of why it requires numerous nodes and steps to function properly

00:50I added some notes around the nodes to make it easier

00:53Now, imagine you're designing a custom cake at a bakery

00:57First, you choose a skilled baker, the model, who specializes in the type of cake you want

01:03In this case, our baker is the Juggernaut model

01:06Then, you give the baker detailed instructions on what you want in the cake, such as vanilla flavor, strawberries, and whipped cream

01:16This is called the positive prompt, and what to avoid, like, uh, nuts, chocolate

01:20This is called the negative prompt

01:24The baker starts preparing the cake by mixing the ingredients and baking it

01:28This is what the K sampler node does

01:31Then, you decide on the size of the cake

01:35Perhaps an 8-inch round cake

01:39This is where we use the empty latent image node, where we specify the width and height using pixels

01:43Once the cake is baked, the baker decorates it and prepares it for serving

01:48In this workflow, we use the VAE decode node for that step

01:51Finally, you take the completed cake home and display it on your dining table

01:55We use the save image node for that step

01:58Did this analogy help you understand the workflow better? Let's review it quickly

02:02We select the baker

02:05You give the baker detailed instructions on what you want in the cake and what to avoid

02:08The baker starts preparing the cake by mixing the ingredients and baking it

02:12You decide on the size of the cake

02:15Um, once the cake is baked, the baker decorates it and prepares it

02:17for serving

02:19Finally, you take the completed cake home and display it on your dining table

02:24Let's talk a little bit about, um, the interface

02:27Uh, the Q prompt button adds the current workflow to the queue, which prepares your setup for image generation

02:34Um, it's like hitting start on a task list

02:37By checking extra options, it reveals more advanced settings for generating images, such as the number of images to create at once, the batch count, or automatic queuing when parameters change

02:50The save button saves the current workflow so you can load it later

02:53It's like saving a document on your computer

02:57The load button loads a previously saved workflow, restoring it to your current session

03:01This is useful for reusing workflows

03:04The refresh button updates or reloads the current view or settings, ensuring you see the most recent changes

03:12The clear button clears the current workflow, resetting it to a blank state

03:16The load default button loads the default settings, returning the interface to its original state

03:21The reset view button resets the layout or arrangement of the interface elements to their default positions

03:27The manager button opens the ComfyUI manager, which is a management interface for, um, detail control over your projects, settings, and custom nodes

03:38The share button allows you to share the current project or workflow with others on different websites

03:42So, save your current workflow by clicking save, give it a name, and store it somewhere on your

03:50drive, then click the clear button to remove everything from the screen and start fresh

03:54The area where you add nodes in ComfyUI is commonly referred to as the canvas or node canvas

04:01This is the main workspace where you can create, connect, and manage various nodes to build your workflows

04:10You can add a new node by either double-clicking

04:12For example, you can search for "load" and select "load checkpoint," which will add that node to the canvas

04:17Uh, you can left-click on it and move it around

04:22And the second option is to right-click on the canvas, then go to "add node," and from the list, select it from the loaders category

04:33I can select the same node, "load checkpoint." To select different nodes, simply click on them

04:38If you press the Alt button and then Left-click and drag, it will drag a copy of that node

04:45Alternatively, you can use Control+C for copy and Control+V for paste

04:50If you right-click on a node, you have access to more options and settings

04:55You can also remove the node using the remove option, or simply press the delete key while the node is selected

05:04Sometimes you need to resize a node to make it look better

05:06You can do this by dragging from the bottom right corner

05:11Resizing from other corners won't work, only that corner will allow you to resize

05:16Here you have the node name, and in front of it, there is a small circle

05:22When you click on it, it will make the node smaller

05:23This action is often referred to as collapsing the node

05:29Collapsing a node reduces it to a more compact form, which helps in managing and organizing the workspace more efficiently, especially when dealing with complex workflows

05:40Here are also some shortcuts you can use to select nodes

05:43Remember these three control keys: Alt, Control, and Shift

05:50Knowing these three keys is usually sufficient for most software out there

05:54If one key doesn't work, try another

05:56That is what I do since I cannot remember all the shortcuts

05:59Many programs have hidden features accessible through these keys

06:05To select a single node, simply click on it

06:07If you want to move the node to a different position, click on the node and drag it to the desired location

06:14To select multiple nodes, hold down the Shift key and click on each node you want to select individually

06:21You can also select multiple nodes by holding down the control key and then dragging to create a selection box that encompasses all the nodes you want to select

06:30Release the mouse button to select all nodes within the box

06:34Once multiple nodes are selected, hold down the shift key, then click and drag any of the selected nodes to move all selected nodes together

06:42Let's clear the canvas and, uh, let me show you a few more things about the nodes

06:47I will add a few different nodes so I can explain what several key elements are composed of

06:54Uh, we have the node name, um, that identifies the function or operation of the node

06:59Uh, for most nodes, you can guess what they do from their names

07:02For example, if "load" is in the name, um, it probably loads

07:07something like a checkpoint model

07:08In this case, K sampler may not be obvious from its name, but you can think of it as a family of samplers

07:14Um, to make it easier to remember, consider K sampler as a key component

07:18Save image does exactly what it says, it saves the image to your computer automatically

07:23Then you have the inputs, connection points where data or signals are received from other nodes

07:29For example, K sampler has many inputs

07:31Uh, load checkpoint doesn't have any inputs, and save image has one input called images

07:39Next, there are the outputs, connection points where data or signals are sent to other nodes

07:45K sampler has only one output named latent

07:48Load checkpoint has three outputs, and save image doesn't have any outputs, so you cannot add more nodes after the image is saved and then have parameters or settings

08:00These are configurable options within the node that control its behavior and processing

08:07For example, K sampler has settings like seed, steps, and CFG that you can adjust

08:13For save image, you can change the prefix value of the image name so that it starts with whatever value you add there

08:22By the way, you can move this menu from the top left corner where those tiny dots are located

08:30The connections between nodes and ComfyUI are called links, but you can call them whatever you like, cables, spaghetti

08:38These links enable data or signals to flow from the output of one node to the input of another, allowing for the creation of a functional

08:43workflow

08:49So, if I want to create a link between these two nodes, I look at the connection points that usually have the same name and color

08:55For example, here, uh, both nodes have, um, connection points named model, and the color is purple

09:02This allows me to drag from purple to purple

09:04However, if I try to drag to a different color, it doesn't let me create the connection

09:11So, uh, I will drag it to the, um, purple connection point, and when I release the mouse, uh, the connection is made

09:18Now you can see the link has the same color as the connection points

09:22Now, if you look at these two points, they may look similar, but in fact, one is yellow and the other is orange

09:29Additionally, their names are clip and positive

09:32Therefore, you cannot make a connection between them

09:35These two different types of points

09:37If you want an idea of what node you can place there, you can drag and release the mouse somewhere on the canvas

09:44A menu will appear where you can add or search for a node

09:48Uh, the menu also suggests nodes that you can select

09:51For example, if I select the, the clip text and code node, it creates that node and connects it in one move

09:59When you delete that node, um, the connection is also deleted

10:01So the easiest way is to drag and release to get one of the suggested nodes or search for a node

10:09With a little practice, you'll learn which nodes connect with each other

10:13For example, with latent, I can get a VAE decode node, then from the VAE decode node, I can connect the pink output with the VAE from load checkpoint, and so on

10:24As You can see, um, you cannot connect nodes of other colors

10:26To unlink or disconnect a link from a node, you simply click on the connection point and drag it away somewhere on the canvas

10:36When you release the mouse, the link will be disconnected

10:38You can also delete, uh, a link by left-clicking on the little circle that appears on the link, then choosing delete

10:44You can also right-click on a node and go to Colors, uh, where you can pick different colors to differentiate nodes that are more important

10:55Additionally, when you right-click, you can change the title of the node, for example, you can rename it, um, to exactly describe what it does

11:03You can also change, uh, the shape of the node to box, round, or card, um, a combination of both, or you can leave it

11:13at the default shape

11:16Additionally, you can use the resize option to, um, automatically adjust the node to a smaller form

11:21Let's clear the workflow and start our text-to-image workflow for Stable Diffusion from scratch

11:28First, let's start by loading our model

11:30We'll use the load checkpoint node for this task, and this node is straightforward

11:34It loads the pre-trained model, uh, the checkpoint that we'll be using

11:39In our case, we're using a model called Juggernaut X

11:41You'll notice that this node has three outputs: model, clip, and VAE

11:46These outputs will be useful shortly

11:48Next, I'll add the K sampler node

11:51This node begins with random noise and gradually turns it into, um, a clear image by following specific instructions and settings

11:59as the number of steps and how strongly the instructions guide the process

12:04This helps create a final, high-quality image

12:07I will create a link between these two nodes, connecting the output from the model, um, output of the low checkpoint node to the model input of the K sampler node

12:17You see here, we have a clip output

12:20Clip stands for Contrastive Language-Image Pre-training, uh, which is a machine learning model that can interpret both images and their descriptions in natural language

12:29It effectively bridges the gap between visual and textual data

12:34To make an analogy for what clip does, it's like having a bilingual person who can understand both English and images

12:41When they see an image, they form a mental summary

12:45When they read a Description, they also form a summary, they can then compare these summaries to see if they match

12:53So if I drag and release, you can see in the suggested nodes, the clip text encode option

12:58If I select that, I now get the clip text encode prompt where we can input our text prompt

13:03Now, we need both a positive prompt and a negative prompt

13:07I can either hold down Alt and drag a copy of the clip text encode node, or I can search for the node labeled prompt and add it

13:14So now that we have both nodes, I will connect them to the load checkpoint node as well

13:19You might wonder how the system distinguishes between, um, the positive and negative prompts when both nodes look the same

13:27In the case sampler node, we have two inputs, one for positive prompt and another for negative prompt by connecting each respective node to these inputs, the system understands which prompt is positive and which is negative

13:40Um, to quickly distinguish between nodes, many people assign colors to them

13:44For example, uh, we can color the, uh, positive prompt node in green and the negative prompt node in red

13:51I will quickly add a positive prompt and a negative prompt so we can test them later when the workflow is ready

13:58Now, each checkpoint has some recommended settings that we can use in the K sampler, as you saw in the first episode when we got our settings from Civit AI, where I downloaded this model

14:09I will add those settings here

14:13So we have another input here called latent image

14:16If we drag and release, we can see options like VAE encode or empty latent image

14:22It took me a while to understand this, but everything that happens in the K sampler node revolves around the latent space

14:30Latent space represents hidden features of data in a simplified, abstract form, making it easier for models to process

14:37Everything that enters this space must be latent

14:40If it's pixel data, we need to encode it first to convert it into this simplified form

14:44So, if I have a photo made from pixels, I cannot input it directly into the K sampler

14:51We need to use VAE encode first

14:54We'll explore this process in another episode to create an image-to-image workflow

14:59However, for text-to-image, we need an empty latent image, even when using text to generate an image, the system still requires a placeholder image to work with and add noise on top to perform its task effectively

15:11As you can see, we have width and height options here

15:16Since we are using an SDXL model, we can create a larger image, such as a 1024 pixel image

15:22So, basically, most of the workflow is complete

15:25We have the model, the prompts, and, um, everything entering the K sampler, uh, including the empty image that allows us to choose the size

15:37Um, theoretically, the image exists in the latent space

15:39Um, we just need to extract it to view it

15:43Remember I mentioned that everything going into and out from the latent space must be either latent or converted to get pixels? Since we want to visualize the pixels, we use the load VAE decode node to extract the image

15:57from the latent space

16:00A VAE (Variational Autoencoder) decode node is a component in machine learning models, especially in image processing, that transforms latent space representations back into their original visible form

16:16Think of it like compressing and decompressing a file

16:18The VAE encode node compresses the image into a compact form, the latent space, and the VAE decode node decompresses it back to the original image

16:32Now, I know from the model description that Juggernaut has VAE baked in, so I can simply connect the VAE from the decoder to the VAE from the load checkpoint node

16:44But if you're using another model that requires a separate VAE file, you'll need to search for a node called load VAE first, download the VAE and place it into your model folder, then into the VAE folder

17:00After refreshing, select the model from the list and connect it to the VAE decode

17:04In my case, I don't need this step, so I'll just connect it back to the model

17:10Now, the final pixel image is here

17:12We just need to save it

17:15We could use a preview node if we want to preview the image first, um, and then save it with a right-click

17:22However, what I prefer to do is use a save image node

17:24Now, that way, everything I generate is saved, and I can delete what I don't need later from the output folder

17:29And that's it

17:33Um, let's test it to see if it works by clicking on Q prompt

17:38Um, you'll notice each node gets a green outline when the flow is active in that area, which is a good sign

17:44Um, at the end, we'll have our Generated image all from a workflow created from scratch

17:50Now, you might ask, why does it need so many components to make a simple image? I might ask you back, why do you need so many Lego pieces to build a castle? Consider a, a real-life example, a desktop PC

18:04You have the unit itself, um, then you connect the power cable to the socket, connect the internet cable, um, to the router, which is connected to other cables, um, and you also connect the, the mouse, monitor, and maybe a printer

18:16All these components are, are necessary for everything to function properly

18:21It's the same concept here

18:22If you want control over the workflow, you need more components to build exactly what you want

18:28I was curious about what's inside this latent image node, so I tried Adding a preview image node

18:35However, as you can see, we cannot connect them directly because the latent image is not pixel-based

18:43Remember, we need to decode it first

18:46So, I will use the VAE decode node, and, uh, now I can connect the image to the preview image node

18:53As you can see, I intentionally left the VAE input disconnected so you could observe what happens when I run this workflow

19:03When I cue the prompt, you'll notice the other nodes turn green, but this one remains red with a little circle around the VAE input, indicating this branch, uh, was, uh, interrupted here

19:16And to fix it, we simply, uh, connect the VAE to VAE and try again

19:23This is how the latent image looks

19:29You know, if I right-click on it, I can, um, open the image in a new tab, and it appears as a gray with subtle purple hues, um, matching the size we defined in width and height

19:35You can do the same for the generated image

19:38With a right-click, you can open, save, or copy the image

19:40I'll, I'll delete this part since, you know, we've seen what's there and, um, it's not particularly exciting

19:49Let's change the width and height to generate a landscape ratio image

19:54I'll adjust the values accordingly and also create a portrait ratio image

20:00Remember to keep it around 1024 pixels

20:03Larger images can capture more details, but if they're too big, you might encounter issues like duplicate heads or slower processing speed

20:13When you're constructing a complex workflow, you might find it useful to visualize how all the nodes are connected

20:19In such cases, a node called reroot can be very handy

20:25Think of a reroot node like a cable organizer for your computer desk

20:29Instead of having all your cables tangled and messy, the cable organizer, reroot node, helps you manage and direct each cable neatly to its destination

20:38After completing the workflow, you can arrange it to fit better on your screen and make it more comfortable for you

20:42For example, if you're used to interfaces like Automatic 1111 or Forge, you can organize it similarly, um, if that's more convenient for you

20:53In Forge, um, we have the stable diffusion checkpoint, which is equivalent to the load checkpoint node in ComfyUI

21:01Similarly, we have the positive and negative prompt nodes, which we can also arrange under the checkpoint if we prefer and resize them accordingly

21:11Then we have, um, the sampling method, steps, CFG settings, as well as the width and height parameters, all of which are, uh, also present in ComfyUI within the case sampler node

21:23Uh, additionally, there's the empty latent image node for setting the width and height

21:28Let's remove these reroute nodes and concentrate on the VAE code

21:32We can collapse the nodes we don't adjust settings for

21:36I'll place it on top to mimic the layout in Forge and position the save image node on the right, similar to how the preview is arranged in Forge

21:48If you don't like how the cables look in the background, you can go to settings and choose a different link render mode, such as straight, uh, then close the settings

21:55Next, I'll go to the manager and click restart

21:59After the restart, the links are now straight

22:02I'll try moving a node to see if the link stays straight

22:04You can also group nodes together

22:06Right-click on the canvas and select "Add Group." I suggest resizing the group before moving it close to the nodes

22:13Any node that touches or is within the group area will stick to that group, like a super strong magnet that pulls in any node it gets close to, creating, uh, an organized bundle

22:28Then I can arrange the items within that group as I like

22:30Uh, I can also rename the group

22:33Now I can move all items at once just by selecting the group instead of individual nodes

22:38After playing around with it for a while, I managed to make it look like this

22:43Now it almost resembles A regular interface and doesn't look so intimidating anymore, right? To change the links back to the default style, go to settings, choose spline, then close from the manager

22:58Restart, start the interface, and voila, we have spaghetti links again

23:02Now, there are custom nodes that can combine multiple nodes into one to make the interface even clearer, but that's for another episode

23:11I hope you enjoyed this video

23:13Please leave a like and a comment to support my channel

23:17Thank you, and see you soon with a new episode

23:24[Music]