[0:00] this is the ultimate AI image generator
[0:04] it's free and open source and the
[0:06] awesome thing about it is you don't even
[0:08] need an Nvidia GPU you can run it with
[0:10] just a CPU or an apple M1 or M2 Chip and
[0:15] it gives you total control not only can
[0:18] you do text to image but also image to
[0:21] image plus upscaling plus you can
[0:24] control the poses of your character or
[0:27] you can control the positions of objects
[0:30] in your image or you can control the
[0:32] depth of your composition and much more
[0:35] you can also do face swapping and create
[0:38] consistent characters you can also link
[0:41] to popular AI tools such as mimic motion
[0:44] which makes your character dance or live
[0:47] portrait which lets you animate any
[0:50] photo of a face or tun crafter which
[0:53] allows you to input a start frame and an
[0:56] end frame and it would generate an anime
[0:59] scene in between those two frames the
[1:02] possibilities are just endless now the
[1:05] tool that we're going to go over today
[1:07] is called comfyi and the interface looks
[1:10] like this so if you're seeing this for
[1:12] the first time this might look very
[1:14] complicated to you and that's exactly
[1:17] why I'm making this tutorial I'm going
[1:19] to show you step by step how to install
[1:21] it and then how to do text image image
[1:24] to image upscaling how to control the
[1:28] pose or the composition or for the depth
[1:30] of your images plus how to download and
[1:33] use external tools like face swapping or
[1:36] toun crafter I'm going to make it as
[1:38] easy as possible so that even if you
[1:40] don't have any technical background on
[1:43] AI or stable diffusion you can still
[1:46] follow along easily so let's get started
[1:48] all right first let's go over how to
[1:51] install comfy UI you simply have to go
[1:53] to their GitHub which I'll link to in
[1:55] the description below and installing it
[1:57] is really easy all you got to do is
[2:00] scroll down and then you'll see this
[2:02] installing comfy UI section and then
[2:04] note that you don't need an Nvidia GPU
[2:07] to run this you can also run it on CPU
[2:10] only and it also supports Apple M1 or M2
[2:15] although of course if you do it this way
[2:17] or if you run it using CPU it's going to
[2:19] be way slower compared to an Nvidia GPU
[2:23] so if you're serious about running image
[2:25] generation locally definitely do get an
[2:27] Nvidia GPU it's just going to make your
[2:29] life a lot more easier and it's going to
[2:31] be compatible with a lot more other
[2:33] open- source AI tools by the way I'm
[2:36] using a Dell Precision 5690 you can
[2:39] integrate a powerful RTX 5000 Ada into
[2:42] this huge thanks to Dell and Nvidia for
[2:45] sponsoring this anyways I'm using
[2:47] Windows so under Windows all you got to
[2:49] do is click on this direct link to
[2:51] download so this is going to install a
[2:54] 7z file which we can unzip now note that
[2:57] this is 1.4 GB in size so it's going to
[2:59] take a few minutes to download depending
[3:01] on the speed of your internet all right
[3:03] so once you've installed the 7zip file
[3:06] you can unzip it with 7zip or WinRAR and
[3:08] then you're going to see this folder now
[3:10] you can extract this folder anywhere so
[3:13] I'm going to extract this on my desktop
[3:16] and this is a very large folder so it's
[3:19] going to take a while for this to fully
[3:21] extract all right so once that is fully
[3:24] extracted we can now open up the folder
[3:27] and then depending on whether you have
[3:29] an Nvidia GPU or something else double
[3:32] click on the appropriate bat file so in
[3:35] my case I do have an RTX 5000 so I'm
[3:37] going to run this and then you might see
[3:39] this message so we got to click more
[3:41] info and then run anyway and so that's
[3:44] how you would install comfy UI very
[3:46] simple to use now if you click on Q
[3:49] prompt you're going to see this error
[3:51] message we don't actually have any check
[3:53] points or models yet so we got to
[3:55] download a model to use for our image
[3:58] generation now there are are thousands
[4:00] of models you could potentially use a
[4:03] model basically defines the style of
[4:05] your image so for example you can choose
[4:07] from realistic to Disney Pixar style to
[4:11] watercolor to anime there are like
[4:13] thousands of checkpoints or models that
[4:16] you can choose from and you can browse
[4:18] all of these models in a site called
[4:20] civit AI which I'll link to in the
[4:22] description below now be warned though
[4:25] there's a lot of NSFW content here so
[4:27] make sure you have your filters on or
[4:30] it's going to be very not safe for work
[4:33] now because there are like thousands of
[4:35] models how do you know which one is the
[4:37] best of course you want to filter out
[4:38] all the bad ones and just use the best
[4:40] ones well there's an awesome site called
[4:43] image cyst which I'll also link to in
[4:45] the description below this is basically
[4:47] a rankings list where users can kind of
[4:50] blind test different models and then all
[4:52] these results are accumulated into this
[4:54] ranking list so you can see for example
[4:57] that real viz has the most points so far
[5:00] followed by colorful XL followed by
[5:03] playground version 2.5 followed by
[5:06] Juggernaut which is also a very popular
[5:08] model and note that most of the ones at
[5:10] the top are using stable diffusion XL
[5:13] which is basically a higher quality
[5:16] version compared to stable diffusion 1.5
[5:19] so let's say we want to use this one
[5:21] real viz XL I'm going to copy this name
[5:23] and in civii I'm going to Simply search
[5:25] for this so here is real viz XL version
[5:28] 4 and here are the results you can see
[5:31] if you scroll down here are some sample
[5:33] images from other users so let's just
[5:36] download the most recent one version 4.0
[5:39] lightning so let's go ahead and download
[5:41] this it is 6.4 GB just a warning so it's
[5:45] going to take a while to download and
[5:46] then you would place your download which
[5:49] is a safe tens file in comfy UI and then
[5:52] models and then checkpoints so let's
[5:55] save it directly in this folder so once
[5:57] you've downloaded the checkpoint or the
[5:59] model file make sure that it is located
[6:02] in comfy UI then in models and then in
[6:05] checkpoints so notice that I have
[6:07] downloaded this real viz XL safe tensors
[6:10] file in this folder and so the next time
[6:12] you start up comfy UI which I'm going to
[6:15] do now you should be able to select the
[6:17] model so again I'm going to open run
[6:20] Nvidia GPU dobat and then wait for this
[6:22] to give me a link all right perfect so
[6:26] now if you go back here in the load
[6:29] checkpoint point if it doesn't
[6:30] automatically select one for you you can
[6:33] click on this and note that I can select
[6:35] this real viz safe tensor file in here
[6:38] now note that for other tutorials they
[6:40] suggest that you download the sdxl base
[6:43] file and also the sdxl refiner file but
[6:46] this is just the default stable
[6:49] diffusion XL model actually this is not
[6:52] necessary and right now sdxl has gotten
[6:54] so good that you don't really need to
[6:56] use the refiner file anymore in most
[6:58] cases and and this Bas sdxl model it's
[7:02] not great so you don't have to download
[7:04] this and you can save a few gigabyt of
[7:06] dis space again if you look at the
[7:08] rankings from image Cy the sdxl base
[7:11] files at least the lightning model seems
[7:13] to be in 12th Place so I wouldn't
[7:15] actually recommend you download this
[7:17] base model and the refiner model so you
[7:19] can actually save a few gigabytes of
[7:21] disk space so anyways back in here all
[7:23] we need to do is download whatever model
[7:25] you want and then here is the positive
[7:28] prompt and then here's the negative
[7:29] prompt you can use your Mouse's scroll
[7:32] wheel to zoom in and out and I'll
[7:34] explain what all these boxes mean in a
[7:36] second but first let's just generate an
[7:38] image to see if this all works so for
[7:41] the positive prompt I'm going to enter a
[7:43] castle in a forest and then for the
[7:46] negative prompt this is basically all
[7:48] the things you don't want to see in your
[7:50] image for example I don't want to see
[7:53] people cars KN and that's pretty much it
[7:56] I'll cover some more advanced prompting
[7:58] techniques later in this video but let's
[8:00] just click Q prompt for now and you can
[8:02] see right now it's running this box
[8:05] which is highlighted in green and then
[8:07] it's running the positive prompt the
[8:08] negative it's going through the K
[8:09] sampler and then now you can see the
[8:11] progress bar is up here so it's in the
[8:14] process of generating the image and then
[8:16] it goes through this VA decode and
[8:18] finally we get an image of a castle in
[8:20] the forest now this image is not great
[8:23] and we're going to go over how we can
[8:24] make this better now first I want to
[8:26] really give you a solid understanding of
[8:29] what all these boxes mean because I
[8:31] think it'll help you when you build more
[8:33] complex workflows so first of all let's
[8:35] just ignore all of this and start from
[8:37] scratch so if I move down here you can
[8:41] start a new box or node by right
[8:43] clicking and then you can see add node
[8:45] and you can select whatever you want now
[8:47] there are so many options here I
[8:49] wouldn't recommend this way so another
[8:51] way to do it is to doubleclick and then
[8:53] after you double click you can actually
[8:55] search for whatever node you want so if
[8:57] I type in checkpoint for example you you
[8:59] can see that we have a low checkpoint
[9:01] node here so I'm going to select this
[9:03] and because the real viz model is the
[9:05] only model I have in this models folder
[9:08] it selects this by default so again this
[9:11] checkpoint basically defines the style
[9:13] of your image for example there's going
[9:15] to be some that work particularly well
[9:17] for anime some that work well for
[9:19] realistic photos and then some that work
[9:21] well for Disney Pixar like characters
[9:24] all right so next we need to add in our
[9:27] positive and negative prompts now again
[9:29] and we can either rightclick and try to
[9:31] find the positive prompt node here I
[9:34] would not recommend that it's really
[9:35] hard to find a node in these options or
[9:38] you can doubleclick and then search for
[9:40] the node here in which case we see it
[9:43] here or another way is to drag from one
[9:46] of these connectors so the prompts
[9:47] should be linked to this clip connector
[9:49] so I'm going to drag a noodle out and
[9:51] then once I release it you can see that
[9:53] it gives me several options which it
[9:56] thinks is most relevant and indeed the
[9:59] prompt window is this one clip text and
[10:02] code so we actually need to drag two of
[10:06] these one for the positive prompt and
[10:08] one for the negative prompt now there
[10:09] are several ways to clone this node so
[10:13] one way is to right click on this and
[10:15] then click clone in which case we'll
[10:18] have another node here but then we'll
[10:19] need to connect this ourselves so I'm
[10:21] just going to delete that and then
[10:23] another way is you can just click on
[10:25] this and press crl C to copy it and if
[10:27] you press crl V to paste it this is what
[10:30] you get but if you want this to have the
[10:32] same connectors as this original node
[10:35] then instead of pressing contrl + V you
[10:37] can press contrl shift V and so now when
[10:40] I paste in this node it's automatically
[10:42] connected to this now just to avoid
[10:44] confusion you can also rename this so
[10:47] for example you can set the title to
[10:50] positive prompt and and then press okay
[10:53] and then here you can right click and
[10:55] then set the title to negative prompt
[10:58] and then press okay you can also set the
[10:59] color if you want so for example I can
[11:01] right click and then in colors I can set
[11:04] this to red and this is important
[11:06] because when you build more complicated
[11:08] workflows this graph is going to be
[11:10] really complicated so if you color code
[11:12] things and rename these nodes it just
[11:14] keeps things more organized and helps
[11:16] you avoid confusion all right next step
[11:19] is we need something called a case
[11:21] sampler and it's basically an algorithm
[11:24] that takes in your prompts and takes in
[11:26] a latent image which we'll go over in a
[11:28] second and create an image from that so
[11:31] if we click on this conditioning
[11:34] connector and drag it out we should see
[11:37] K sampler here so I'm going to select
[11:39] that this is for the positive and then
[11:40] for the negative prompt we will connect
[11:42] this to here and then for the model we
[11:45] just drag this one all the way to here
[11:48] where it says model and then for latent
[11:50] image let's drag a line out and then
[11:54] here it would give us the option to
[11:56] create an empty latent image so so let
[11:59] me just drag this over here to keep
[12:01] things cleaner all right so what on
[12:04] Earth does this mean what exactly is an
[12:07] empty latent image so how stable
[12:09] diffusion works is it doesn't create an
[12:12] image from a blank canvas actually what
[12:15] it does is it starts with an image of
[12:18] just random noise and then at every step
[12:20] it removes some of that noise and if you
[12:23] remove enough noise you get whatever
[12:25] image you prompted it with so this empty
[12:27] latent image is basic basically an image
[12:30] of random noise and you can select the
[12:33] width and the height so for example we
[12:36] can change this to 800 if you want and
[12:38] then same with the height 800 by the way
[12:40] for sdxl it's best to create an image of
[12:44] 1024x 1024 because that is what it's
[12:47] optimized for and then if you are using
[12:50] stable diffusion 1.5 or earlier it's
[12:53] best to use 512 x 512 and then the batch
[12:56] size is how many images you want so if
[12:59] you set this to two basically it will
[13:02] create two images at once but for now
[13:04] let's just set this to one all right so
[13:07] next let's go over all these settings
[13:10] the seed is basically the starting point
[13:13] of this random image this image is just
[13:16] random noise but there could be an
[13:18] infinite number of images of random
[13:19] noise each of them would be slightly
[13:21] different so the seed basically defines
[13:23] the starting point and usually we keep
[13:26] the seed at random so that's what this
[13:28] one does however if you keep it at fixed
[13:31] and you set the C to a certain value for
[13:34] example 69 then if all of the other
[13:37] settings are the same you're going to
[13:39] generate the same image every time
[13:41] because the seed or the starting number
[13:43] is the same but for now let's leave it
[13:45] at zero and then for here let's set it
[13:48] to randomize and then the number of
[13:50] steps is basically again if we go back
[13:53] to how stable diffusion works is how
[13:56] many steps of noise removal do we want
[13:59] so if it's just a few steps you're only
[14:01] going to get something like this you
[14:03] haven't removed enough noise yet and
[14:05] after enough steps you're going to get a
[14:08] very clean image of whatever you want to
[14:10] generate so generally few steps would
[14:13] give you a lot of noise and then after a
[14:15] point like if you exceed 50 to 100 steps
[14:18] then you're not going to get a better
[14:20] image in fact you might get some noise
[14:22] or artifacts because there's not much
[14:24] remaining noise to be removed so I'm
[14:26] just going to keep the number of steps
[14:27] at 20 for now and then CFG is basically
[14:31] how well do you want this algorithm to
[14:33] follow your prompt so if you set this to
[14:37] one for example it's not going to follow
[14:39] your prompt and it's going to try to
[14:41] generate whatever it wants basically it
[14:43] can be more creative however if you set
[14:45] it to 15 for example then it listens to
[14:48] your prompt very literally and it tries
[14:50] to generate whatever you specify in your
[14:52] prompt which is sometimes not what you
[14:54] want sometimes if it's too literal
[14:55] you're going to get some weird results
[14:57] so generally a value of like seven or
[15:00] eight would work best and then the
[15:02] sampler name this is basically the
[15:04] algorithm that's used to remove this
[15:07] noise and generate the image for you so
[15:09] ooler is a common one this is one of the
[15:12] fastest ones and then if you want to go
[15:13] for Quality I would suggest one of these
[15:16] so DPM PP 2m or 2m SD but for now let's
[15:20] just go with ooler and then for
[15:22] scheduler this basically gives even more
[15:25] quality so usually if you want a really
[15:27] good quality image would select Caris or
[15:30] exponential these settings are just very
[15:32] subtle differences you can add to the
[15:34] image so just play around with it so you
[15:36] can get a feel for what settings work
[15:38] best for your particular use case and
[15:41] then Den noise what this is is because
[15:43] again our latent image or starting image
[15:46] is basically an image of random noise
[15:49] the den noising strength is basically
[15:51] how much noise do we want to remove from
[15:53] this initial image so if we're starting
[15:56] from scratch then obviously we want to
[15:58] remove 100 % of the noise so that's why
[16:01] this is set as one all right so the next
[16:04] step is we next need to connect this
[16:07] latent connector so I'm going to drag
[16:09] this out and we need to connect it to
[16:11] what is called a VA decode a vae
[16:15] basically encodes an image into a latent
[16:18] space but right now we want to do the
[16:20] reverse of that so we need to decode
[16:23] this latent image into an image that we
[16:25] actually want to see so that's why we
[16:28] need this final step to decode this
[16:30] latent image and then we need to drag
[16:33] the vae from our checkpoint to this vae
[16:37] connector that we see here so in most
[16:39] cases your checkpoint should come built
[16:42] in with a vae so all we got to do is
[16:44] just connect this vae to this vae
[16:47] connector here and then we are almost
[16:50] done right now we just need to produce
[16:52] the image so once we drag this out we
[16:54] actually have two options we can either
[16:56] preview the image so if we preview the
[16:59] image it's not going to save the image
[17:00] in our computer or we can select save
[17:03] image which also shows us a preview of
[17:05] the image but it also automatically
[17:07] saves the image so I'm going to select
[17:09] preview first because I don't want it to
[17:12] save every image it generates I only
[17:14] want to select the good ones to save on
[17:16] my computer so I'm going to show you how
[17:18] to save a previewed image in a second
[17:21] there are very few software tools that I
[17:23] use every day but this is one of them
[17:26] thanks to our sponsor turbo type it's
[17:28] free forever and it saves me so much
[17:30] time basically you can create custom
[17:33] keyboard shortcuts so that you don't
[17:34] need to keep typing out repetitive
[17:36] things for example if there's a prompt
[17:38] that I use in chat GPT very often I can
[17:40] make a shortcut here and then when I go
[17:43] to chat GPT or anywhere else I just need
[17:45] to type in the shortcut and voila or
[17:49] let's say I have a very long email
[17:51] address I can also make a shortcut for
[17:53] that so that whenever I need to enter in
[17:55] my email I can simply type in the
[17:58] shortcut and it types out the email
[18:00] finally it also supports Rich Text for
[18:03] example you can add in bold and italics
[18:05] and add links to your text as well so
[18:07] let's say I need to send out a lot of
[18:09] cold emails with the following template
[18:12] well I can just create a shortcut for
[18:14] that and then whenever I start an email
[18:16] I just need to type in the shortcut and
[18:18] voila the text is already styled and
[18:21] linked for me there are hundreds of
[18:23] pre-existing templates that you can
[18:25] choose from including common prompts for
[18:27] chat GPT business finance medicine and
[18:31] more this tool saves me so much time
[18:34] every day there's absolutely no reason
[18:36] not to use this because they have a free
[18:38] forever plan so definitely check it out
[18:41] and download the free Chrome extension
[18:43] in the link below but that is pretty
[18:46] much it so for the positive prompt let's
[18:49] put a med evil Warrior realistic 8K
[18:55] Masterpiece these are just some keywords
[18:58] that I tend to use a lot to give it more
[19:00] detailed and then also Ultra detailed is
[19:03] another good one and then for negative
[19:06] prompts again these are all the things
[19:07] we don't want to see in our image so for
[19:09] example I don't want it to painting or a
[19:12] cartoon or anime drawing I don't want it
[19:15] to have any copyright or watermarks and
[19:19] I think we are good to go for now so
[19:22] before we click run note that I had this
[19:26] previously this is the default so I
[19:28] don't want to to run both of these at
[19:29] once so I'm going to select all of these
[19:32] nodes and delete them so to select
[19:34] multiple nodes at once what you can do
[19:36] is hold down control and then drag to
[19:39] Encompass all the nodes that you want to
[19:41] select and then if you want to move this
[19:43] group of nodes around you need to hold
[19:45] down the shift key and then you can drag
[19:47] this wherever you want and then to
[19:49] delete everything at once all you got to
[19:51] do is press delete all right so moving
[19:54] back down here everything else looks
[19:57] good so I'm going to press Q prompt so
[20:01] note that it's starting here it's
[20:02] loading the checkpoint and then it's
[20:05] moving over to K sampler now it's
[20:06] generating the
[20:09] image and then it's decoding the image
[20:12] and then voila we have a medieval
[20:15] warrior so basically this is like the
[20:18] standard workflow for a simple text to
[20:22] image generation and then let's do
[20:24] another one so let's say I want to
[20:26] generate two images at once all we got
[20:28] to do is increase the batch size to two
[20:31] and then press Q prompt again and note
[20:33] that it starts off here it starts in the
[20:36] case sampler it doesn't start here or
[20:38] here and that's because we haven't
[20:40] changed any of these other nodes in a
[20:42] previous step so all of this is already
[20:45] saved in memory it only loads from here
[20:48] and so this makes comy UI very efficient
[20:52] and then so now we have two images this
[20:54] is the first one here's the second one
[20:57] and remember this is not saving your
[20:58] image this is just the preview image
[21:01] node so to save it all you got to do is
[21:03] right click and then press save image so
[21:06] this is the most basic text to image
[21:09] workflow for stable diffusion hopefully
[21:12] this gives you a better understanding of
[21:14] what a k sampler is what an empty latent
[21:17] image is and what a VA decode is because
[21:20] once you set up more complex workflows
[21:23] you're going to need to understand what
[21:25] these nodes actually do so I hope this
[21:27] gives you a good understanding all right
[21:29] before we move on to the next section
[21:31] let's go over some tips and tricks for
[21:34] navigation and organization and
[21:36] productivity so first thing is I'm not
[21:39] sure if you can see it on my screen
[21:42] share right now but there is a very
[21:44] faint dark blue frame on your comfy UI
[21:48] canvas so I'm hovering my cursor over
[21:51] that frame right now I wish they could
[21:53] make the contrast higher so you can
[21:54] actually see the blue line but anyways
[21:56] there's this very faint Blue Line and so
[21:59] whenever you start an interface the
[22:02] default location would be within this
[22:05] blue frame so it's always best to have
[22:07] your workflow within this blue frame so
[22:10] that whenever you load up comfy UI your
[22:12] workflow will show up right away and you
[22:14] don't need to like try to find it within
[22:16] this huge canvas all right next thing
[22:18] you can do is let's say this is a
[22:21] workflow that you want to use again in
[22:23] the future you can click this to save it
[22:25] or press crl s so let's name this this
[22:29] as temp. Json and then you can save this
[22:32] wherever you want I'm just going to save
[22:34] this in the compi folder another
[22:36] keyboard shortcut to be aware of ISR a
[22:38] which is select all and then delete or
[22:41] backspace which will delete everything
[22:43] you've selected all right so right now
[22:45] this whole workflow is gone now I can
[22:47] undo it by pressing contrl Z which would
[22:50] undo my delete of the workflow or I can
[22:52] also redo the action by pressing control
[22:54] Y which deletes the whole workflow again
[22:57] now to load up a previously saved
[22:59] workflow you can press load here or
[23:02] press crl o so if we go into our compy
[23:05] UI folder and then select this temp.
[23:08] Json which we just loaded you can see it
[23:11] has loaded our workflow back up now a
[23:13] few shortcuts for navigation you can
[23:15] either press on your mouse and then move
[23:18] the canvas around or you don't actually
[23:20] need to click on your mouse you can also
[23:21] hold down the space bar and then move
[23:23] your mouse around and it would still
[23:25] move the canvas now you can zoom in and
[23:27] out by using the scroll wheel and then
[23:30] to select multiple nodes you just simply
[23:32] hold down control and then select this
[23:35] one and then let's say I want to select
[23:36] this one and select this one so I'm
[23:38] holding down control for each of these
[23:40] and then let's say I want to delete
[23:41] these I can just press delete and then
[23:43] let me undo that by pressing contrl Z
[23:46] you can also select multiple nodes by
[23:48] holding down control and then dragging a
[23:51] frame around all the nodes you want to
[23:53] select and then if you want to move
[23:55] these nodes around simply hold shift and
[23:58] then you can drag this group of nodes
[24:00] that you've selected wherever you want
[24:03] another thing you can do is let me
[24:05] delete this first now every time you
[24:08] drag a connector out there's also an
[24:10] option called reroute so if I click this
[24:13] it's basically just an extra blank node
[24:16] which extends your connection further so
[24:17] it's basically the same thing as just
[24:20] connecting this vae to this vae but the
[24:22] nice thing about this is let's say you
[24:24] don't want this line to be hidden behind
[24:26] this node well you can d it out like
[24:29] this so you can clearly see that this
[24:31] line is being connected here and then
[24:34] one more thing is right now you see that
[24:37] for example in K sampler these values
[24:40] are set in this node but what if you
[24:43] want to set this value somewhere else
[24:46] and then link it to here so for example
[24:48] for CFG if you want to set this value
[24:51] somewhere else and then link it to the
[24:52] cas sampler you can right click on this
[24:54] and then under convert widget to input
[24:58] you can set any of these options to an
[25:00] input so let's say we want to set CFG to
[25:04] an input now you can see that CFG has
[25:06] disappeared from here and it is now an
[25:08] input connector so let's drag this out
[25:11] and then we need to actually use the
[25:13] node called primitive so right now this
[25:16] value is set to seven it's connected to
[25:18] this CFG input so the CFG of this case
[25:21] sampler is 7 so that's how you would use
[25:25] it all right one final thing I want to
[25:27] share with you and this is really cool
[25:29] let me press crl a and delete all of
[25:32] these let's say I made an image
[25:34] previously using com UI well if I drag
[25:38] that image onto the canvas what happens
[25:41] is it actually gives me the entire
[25:42] workflow that I used to create the image
[25:45] how cool is that so like if you go
[25:47] online and other users have shared their
[25:51] comy UI generations and assuming they
[25:53] haven't deleted that metadata you can
[25:55] actually download their image and then
[25:57] drag and drop that image onto comy UI to
[26:00] look at the entire workflow that they
[26:02] used to generate that image this allows
[26:05] you to learn really quickly so those are
[26:08] like the basic things you need to know
[26:10] for organization and productivity using
[26:12] comfy UI so let's move on to the next
[26:14] section first of all we need to install
[26:17] this plug-in called comfy UI manager it
[26:20] will make your life a lot easier for
[26:22] installing extensions and plugins and
[26:25] missing models so I'm going to link to
[26:27] this GitHub repo in the description as
[26:29] well and if you scroll down a bit they
[26:31] will give you some installation
[26:33] instructions so in our comfy UI folder
[26:37] and then in our custom nodes folder we
[26:39] just need to open command prompt here so
[26:42] in this bar up at the top type in CMD
[26:45] and this will open up our Command Prompt
[26:46] and you can see that we are now inside
[26:48] our custom nodes folder next all we got
[26:50] to do is get clone this repo so we will
[26:55] paste it in here now you do need to have
[26:58] get installed first if you don't here's
[27:00] how to install git if you already have
[27:02] git installed feel free to skip to the
[27:04] next section so all we got to do is
[27:06] download the latest release for whatever
[27:08] operating system you're using so I'm
[27:10] using Windows so I'm just going to click
[27:11] on download for Windows I'm running 64
[27:14] bit so I'm going to click on this to
[27:15] download and it's now downloading this
[27:17] exe file so once that's completed all we
[27:19] got to do is open that exe file and then
[27:22] follow the steps so I'm going to click
[27:24] on next I'm just going to go with the
[27:25] default install location which is
[27:27] program file SLG so I'll click next for
[27:30] that and then I'm just going to leave
[27:31] this at the default and then I'm going
[27:33] to click next again and click next here
[27:36] we're just going to use the default
[27:37] settings for all of these there's a lot
[27:39] of settings that you need to go through
[27:41] so I'm just going to click next for all
[27:42] of these all right and then it should go
[27:44] ahead and install all the files so this
[27:47] might take a few
[27:48] minutes perfect so now we have git
[27:50] installed all right so assuming you have
[27:53] get installed already you simply copy
[27:55] this line and then paste it in here and
[27:57] then press it enter and you'll see that
[27:59] now we are cloning this comfy UI manager
[28:02] into the custom nodes folder and then if
[28:04] you actually open up your custom nodes
[28:06] folder you can see this new folder
[28:08] called comfy UI manager all right so
[28:10] next we need to restart comfy UI so
[28:13] going back in our Windows portable
[28:15] folder we are going to run comfy UI
[28:18] again and you can see that it has
[28:20] detected that we have comfy UI manager
[28:22] installed so it's now installing
[28:24] dependencies all right so after you open
[28:27] your comfu up I now you should see this
[28:29] manager button in the right menu so next
[28:32] I'm going to show you how to upscale an
[28:34] image and then we're also going to move
[28:35] on to some more advanced workflows like
[28:37] image to image and control net and
[28:40] installing other plugins so first let's
[28:42] click on this manager button and then
[28:45] the main buttons that you will use is
[28:48] this one custom nodes manager there's
[28:50] also a button where you can like update
[28:53] comfy UI or update all or install
[28:55] missing custom nodes for example if you
[28:57] import ort a workflow that was made from
[28:59] another user you might have some missing
[29:02] models or nodes or dependencies so
[29:04] clicking this button will just
[29:05] automatically install all of them so
[29:07] that this other user's workflow will
[29:09] work on your computer and then here this
[29:11] is model manager so instead of going to
[29:14] civit Ai and browsing through all the
[29:15] models you can just easily download the
[29:18] model here through this interface so
[29:19] let's click on this and then let's
[29:22] search upscale and you should see a lot
[29:26] of different upscaler algorithms so
[29:29] usually the ones that I find work best
[29:32] are real sran X4 this is for realistic
[29:36] photos as the name implies and then
[29:38] there's also 4X Ultra sharp which works
[29:40] pretty good as well just to keep it
[29:42] simple for this tutorial I'm just going
[29:44] to download these two but definitely you
[29:47] can install all these and play around
[29:48] with it and see which algorithm works
[29:50] best for you so I'm going to select
[29:52] these two and then click install all
[29:55] right so after it is finished installing
[29:58] note that you need to click the refresh
[30:00] button on the main menu to apply these
[30:03] installations so we're going to click
[30:05] close And then close again and then
[30:07] click refresh all right let's dive in to
[30:09] see how we can upscale images so we're
[30:12] going to use the same prompts a medieval
[30:14] warrior realistic 8K Masterpiece Ultra
[30:17] detailed and then for the initial image
[30:22] width and height let's set it to 512 now
[30:25] I understand that for sdxl the optimum
[30:28] dimensions are 1024x 1024 but in this
[30:31] example I'm just going to show you a
[30:33] really blurry and low resolution image
[30:36] and then we're going to upscale it by
[30:37] four times so you can clearly see the
[30:39] before and after and then for the batch
[30:42] size let's leave it at 1 for now number
[30:44] of steps we can also leave it at 20 and
[30:47] then here instead of preview image let's
[30:49] delete that and then we can just drag
[30:52] this image out and then we don't see any
[30:54] upscale here so we need to click search
[30:57] and then we can search upscale now there
[31:00] are a few options one is upscale image
[31:03] which is kind of the same as upscale
[31:04] image bu and then we have upscale image
[31:07] using model now I'll show you what this
[31:09] one does first upscale image bu I would
[31:12] not recommend this because this is
[31:14] basically just upscaling your initial
[31:17] image but it's not adding any details
[31:19] it's basically just increasing the size
[31:22] so let's say we want to upscale this by
[31:24] two times so instead of 52 12 x 512 it's
[31:28] going to be 1024x 1024 and then for the
[31:32] image let's drag it out and we will have
[31:35] a preview image node and also I want to
[31:38] preview the image before we upscale it
[31:41] so you can compare the before and after
[31:43] so over here I'm also going to drag a
[31:45] preview image node all right now let's
[31:48] run this and see what we get so I'm
[31:51] going to press Q
[31:52] prompt all right so let's look at the
[31:54] initial image this is only 512 x 512 so
[31:59] I'm zooming in quite a bit and you can
[32:01] see the details of his face are very
[32:03] blurry and then let's look at the
[32:05] upscaled image yes this is 1024x 1024
[32:09] but you can see the details are the same
[32:12] this is pretty much the same image the
[32:14] same blurriness we're not adding any
[32:16] details here so again it's not
[32:19] recommended to use this upscale image by
[32:21] Method because you're simply just
[32:23] resizing the image but you're not adding
[32:25] more details to the image so I'm going
[32:27] to delete this one and also this one so
[32:30] next let's drag out another node and
[32:33] this time I'm going to search upscale
[32:35] again and then we are going to use
[32:36] upscale image using model and then after
[32:39] that we need to actually input an
[32:41] upscale model so I'm going to drag this
[32:43] node out and then we should get the one
[32:46] and only option which is upscale model
[32:48] loader so because I've downloaded for X
[32:51] Ultra sharp and real s again we should
[32:54] automatically see this over here now
[32:56] both of these are for X so your output
[32:59] image will be four times the resolution
[33:02] if you want 2x for example well you can
[33:04] go into manager again and then click on
[33:05] model manager and then find an upscale
[33:08] model that is only 2x like this real
[33:10] sran X2 all right and then just one last
[33:13] step is we need to drag out another node
[33:16] to either save the image or preview the
[33:19] image so if I run this
[33:25] again you can see that
[33:28] if I zoom in on both of these this is
[33:31] 512 x 512 so the details are very blurry
[33:34] but if you expand this image which is
[33:37] like over 2,000 * 2,000 you can see that
[33:40] the details are a lot sharper especially
[33:43] the patterns on his helmet and on his
[33:46] armor now this isn't the best way to
[33:48] upscale it's actually best to do one
[33:51] round of image to image first before we
[33:53] upscale so I'll show you that in a
[33:55] second so one more thing I want to
[33:57] mention is that sometimes you don't want
[34:00] to upscale all the images that you
[34:02] generate you want to decide which image
[34:04] you want to upscale because let's say
[34:06] for this initial image you don't like
[34:08] the design you don't like the
[34:09] composition you don't want to proceed
[34:11] further with upscaling and waste
[34:13] Computing resources so how do we decide
[34:16] if we want to upscale or not well first
[34:18] of all you can break off a workflow by
[34:21] selecting the node where you want to
[34:23] break it off in this case I want it to
[34:25] pause here so it only generates the
[34:27] initial image but it doesn't proceed
[34:29] further to the upscale unless I want it
[34:31] to do that and then I'm going to press
[34:33] crl M and this will mute the note and
[34:36] basically everything that goes after
[34:38] this point is going to be paused until I
[34:41] unmute this node now when I press Q
[34:43] prompt and it generates an image I can
[34:46] decide whether I want to proceed further
[34:48] with this upscale method and if I do
[34:50] then I would unmute this node and press
[34:52] Q prompt again however one thing to note
[34:54] is in this case you do need to set the
[34:57] same seed otherwise if you don't set the
[34:59] same seed when you press Q prompt again
[35:02] it's going to generate a completely
[35:03] different image and then it's going to
[35:05] upscale that image so going back to here
[35:08] we need to actually set this to fixed so
[35:11] it's going to be a fixed seed and then
[35:13] so if we run Q prompt again you can see
[35:15] that it's generating a new image and
[35:18] that image is being fed into here and
[35:20] then let's say I like this image I want
[35:22] to upscale it then I would click on this
[35:25] node which is now muted I'm going to
[35:27] unmute this by pressing crl M so now
[35:29] when I press Q prompt again you can see
[35:31] it's actually proceeding from here and
[35:33] then now it's upscaling this image and
[35:35] we can now see our upscaled image so
[35:39] that's one way to do it another even
[35:41] more efficient option is if you go back
[35:44] to manager and then click on custom
[35:46] nodes manager and then let's search for
[35:49] image Chooser and this is a node that's
[35:52] created by Chris Orange let's go ahead
[35:54] and install
[35:56] this all right all right so it says
[35:58] restart required so let's restart this
[36:02] all right so we are back here now let's
[36:05] say we want to generate four images and
[36:08] we would choose one of them to go
[36:10] through the upscale so let's increase
[36:12] the batch size here to four and then we
[36:16] can keep everything as is for the seed
[36:19] actually we can set this back to
[36:20] randomize and then yes this has to go
[36:22] through this vae decoder to decode the
[36:25] latent image and then instead of preview
[36:27] image let me just delete that let me
[36:29] also delete this node for now and then
[36:31] let's drag this image out and we will
[36:34] search for preview Chooser which is down
[36:37] here and then for any images that we
[36:40] select here we can then proceed to
[36:42] upscale it so let me just run this for
[36:44] you first so you can see what this does
[36:47] so right now it's loading the checkpoint
[36:50] again because we've restored the
[36:51] interface now it's inputting the
[36:53] positive negative prompts now it's
[36:54] generating four images through this case
[36:57] sampler then it's going to decode these
[36:59] four images and then so now we have four
[37:02] Images all right so let's say we want to
[37:06] select this one to upscale so we would
[37:09] click on this or if you want to select
[37:11] more you can always increase the count
[37:12] to two and then select this one for
[37:14] example but I'm going to decrease this
[37:16] to one and unselect this so we are only
[37:19] going to proceed with this image through
[37:21] the upscaler and then we simply click
[37:24] progress selected image and then this
[37:26] would go through the upscaler and it's
[37:28] using our forx Ultra sharp model and
[37:30] then voila here is our upscaled image so
[37:33] in a nutshell that's how you do
[37:35] upscaling now there is an even better
[37:38] method to upscale images which kind of
[37:41] uses image to image so next we're going
[37:43] to go over how to do image to image and
[37:45] then we're going to go back to this
[37:47] better upscaling method all right next
[37:49] I'm going to show you how to do image to
[37:52] image so what I'm going to do is first
[37:54] of all hold control and then Dr drag
[37:58] these nodes to select all these nodes
[38:01] I'm going to copy it and then paste it
[38:05] somewhere here and then I'm going to
[38:07] hold down shift and then move it to
[38:09] where I would like all right now going
[38:11] back to this another way instead of just
[38:14] deleting this workflow is to hold
[38:17] control and select all of this and then
[38:19] press contrl B to bypass so if you see
[38:22] the nodes highlighted in purple that
[38:24] means it will be bypassed this will not
[38:27] r run so only this will run all right so
[38:31] this is our standard text to image
[38:34] workflow right we have our checkpoint
[38:36] and then we have our positive prompt our
[38:38] negative prompt this goes into the K
[38:40] sampler the K sampler takes an El latent
[38:43] image which is just an image of random
[38:45] noise and then after going through this
[38:47] algorithm and going through this amount
[38:49] of steps then the final step is we need
[38:52] to use our vae to decode this latent
[38:56] space back into an image that we want to
[38:58] see now for image to image we don't want
[39:01] to start with an image of random noise
[39:03] right we want our input to be an image
[39:06] so let me select this node and delete
[39:09] that and then I'm going to double click
[39:11] anywhere and then search for load image
[39:14] so let me click on this so this is the
[39:17] default but you can click this button to
[39:19] upload an image I'm going to upload this
[39:22] image now we can't just directly drag
[39:25] this image to this Laton image connector
[39:28] that's because we need to First convert
[39:31] this image to a latent image and then
[39:34] connect the latent image into here so
[39:36] let me drag this connector out and you
[39:39] should see the first option here would
[39:41] be vae and code and that's exactly what
[39:44] we want to do we want to use a vae to
[39:48] encode this into latent space and then
[39:51] drag the latent image onto this
[39:54] connector all right and then where do we
[39:56] get our VA
[39:58] well we can just get the vae from the
[40:00] checkpoint that we loaded so let's drag
[40:03] this to here and then you can see that
[40:05] everything is connected now we do need
[40:07] to set up some additional settings for
[40:09] the K Samplers so one thing I forgot to
[40:12] mention is that if your checkpoint is
[40:14] lightning it only takes around like 5 to
[40:17] 8 steps you actually don't need 20 steps
[40:20] that's the awesome thing about lightning
[40:22] models and some models can even generate
[40:25] a decent image in as few as two two
[40:27] steps for us let's set the steps to
[40:31] seven sampler schedule we can leave it
[40:33] as is Den noise this is what we want to
[40:36] change if you remember for our text to
[40:39] image denoise this is saying to take our
[40:43] lat an image of random noise and replace
[40:45] 100% of that noise so that we get the
[40:48] image that we specified in our prompt
[40:51] but in this case because we're not
[40:53] starting with random noise we're
[40:55] starting with an image we don't want to
[40:57] remove everything we want to retain some
[41:00] of this image so in this case if you're
[41:03] doing image to image the noising
[41:04] strength means how much of this original
[41:07] image do you want to remove so if you
[41:10] set this to 100% or 1.0 in this case
[41:14] it's going to remove everything from
[41:16] this image you're going to get a
[41:17] completely different image conversely if
[41:20] you set this to zero then it's just
[41:22] going to produce this exact image
[41:24] nothing would change so depending on how
[41:26] similar you want the image to be it's
[41:28] better to have something in between so
[41:30] let's do 3 first so it's more similar to
[41:34] this image and then I'll show you an
[41:35] example of8 for comparison all right so
[41:39] now that we have everything in place
[41:41] there's just one final step which is to
[41:43] drag this image connector out and then
[41:46] we are going to preview image let me
[41:49] just reposition this here so you can see
[41:51] the entire workflow and we are good to
[41:53] go let's click Q prompt so it's our
[41:56] first time starting this workflow so it
[41:59] takes some time to load in the
[42:00] checkpoints and then it's going through
[42:02] the prompts it's taking this image and
[42:05] encoding it it's now going through this
[42:06] K sampler and then it's going to decode
[42:09] the image and then preview the image for
[42:11] us all right perfect so you can see if I
[42:14] drag this over here just temporarily for
[42:17] comparison this is image to image so
[42:20] here's our original image here's our new
[42:22] image and the denoising strength was set
[42:25] to. 3 all right let me now try one with
[42:30] 8 for example and you should see that it
[42:32] would be less similar to the original
[42:35] image so I'm going to set the denoising
[42:37] strength to8 press okay and then run
[42:40] this again and notice it's not starting
[42:42] from the beginning it's starting from
[42:44] here because this is the last note where
[42:46] we changed the settings so again this
[42:48] makes comfy UI very efficient and now
[42:51] it's decoding it and you can see for
[42:53] this one it's more different compared to
[42:55] the original image
[42:57] so basically in a nutshell this denoise
[43:00] value determines how similar do you want
[43:02] your output image to be compared with
[43:05] your uploaded image all right now
[43:07] remember how I said there's a much
[43:10] better way to upscale images using image
[43:13] to image well now that we've gone over
[43:15] image to image and specifically the den
[43:19] noising strength setting I can now show
[43:22] you a much better way to upscale images
[43:25] and this is one of the best ways to
[43:27] actually upscale images in order for
[43:30] this to work we need to download another
[43:32] node so let's go into manager and then
[43:35] click on custom nodes manager and this
[43:38] time we are going to search for ultimate
[43:40] SD upscale so let's click install here
[43:44] and after it has installed it says we
[43:46] need to restart com UI so let's click
[43:49] that all right so let's start with the
[43:51] default text to image workflow and I'm
[43:54] going to show you how to set this up so
[43:56] instead of the case sampler this new
[43:57] node is basically going to replace the K
[44:00] sampler so let me hold control and then
[44:03] select all three of these nodes and then
[44:05] click delete and then I'm going to drag
[44:07] a connector out here and then search for
[44:10] ultimate SD upscale and then we are also
[44:13] going to connect the negative prompt
[44:16] here we're going to connect the vae here
[44:18] and also connect the model here this
[44:21] time we are not going to use a latent
[44:24] image so I'm going to delete that and
[44:26] then instead we need to upload an image
[44:28] so I'm going to drag a connector out
[44:31] from image and then select load image
[44:34] and then I'm going to choose this image
[44:38] now this is a 512x 512 image that we
[44:41] generated previously you can see it's
[44:43] very blurry and then for the prompt we
[44:46] are going to set this to the same prompt
[44:48] that we used before which is medieval
[44:51] warrior 8K Masterpiece Ultra detailed
[44:55] realistic and then for the next negative
[44:57] prompt again same as before we are going
[44:59] to type in anime 2D cartoon painting
[45:03] Watermark all right so we are almost
[45:06] good to go one last thing is we need to
[45:08] select an upscale model so let's drag
[45:10] this connector out and then select the
[45:12] one and only upscale model loader and
[45:15] this will pull up the upscaler models
[45:17] that we've downloaded previously in this
[45:19] tutorial so let's go with 4X ultr sharp
[45:22] and now what this ultimate upscale does
[45:26] is called tiled upscaling so let's say
[45:29] we upscale this by two what this is
[45:32] actually going to do is break this into
[45:35] four sections so it's 2x two and then
[45:38] it's going to apply image to image for
[45:41] each quadrant so it's going to generate
[45:43] image to image for this one and then
[45:45] image to image for this one and then
[45:47] this one and then this one and then it's
[45:49] going to stitch all four quadrants
[45:52] together to give you your upscaled image
[45:54] and now the trick is if you use image to
[45:56] image image with this with a low Den
[45:59] noising strength again this is how much
[46:01] of your original image do you want to
[46:04] retain then this method is actually a
[46:06] lot better than than just upscaling with
[46:09] 4X Ultra sharp so again the key here is
[46:12] to set a denois strength to a relatively
[46:15] low value I think 0 2 is a good start or
[46:18] you could even go with 0.15 all right so
[46:21] all of these settings we've gone over
[46:23] before tile width and tile height just
[46:26] refer to the dimensions of each one of
[46:29] these quadrants so this one would be 512
[46:32] x 512 this one would also be 512 x 512
[46:36] or whatever you set here usually you
[46:38] would just set this to the width and
[46:40] height of your original image and then
[46:43] mask blur and tile padding this just
[46:45] refers to how well these tiles blend
[46:48] together after they are glued back
[46:50] together to form your final image so I
[46:53] just tend to leave it at the default but
[46:54] feel free to play around with these
[46:56] settings if for some reason you get a
[46:58] very obvious line in between tiles and
[47:02] then that's pretty much it the final
[47:04] step is to drag this connector out and
[47:06] then I'm going to select preview image
[47:09] all right so let's click Q prompt and
[47:12] I'll show you what that gives us all
[47:15] right so if you now compare these two
[47:18] images you can see that this is a lot
[47:21] more detailed and the details are a lot
[47:24] finer now previously we did a a clean 4X
[47:28] Ultra sharp upscale which looks like
[47:30] this right the details are not great you
[47:33] can see the hair doesn't really look
[47:35] like hair same with his facial hair same
[47:37] with his face it Still Remains blurry
[47:39] now this is 4 * 512 so this is 2048 *
[47:45] 2048 right now this is only 1024 x 1024
[47:48] since we upscaled it two times so to
[47:51] give you an Apples to Apples comparison
[47:54] we can either set this value to four
[47:57] which would upscale it four times or let
[47:59] me show you another trick you can do to
[48:02] upscale this further and this really
[48:04] shows the versatility of comy UI you can
[48:07] literally just customize the workflow to
[48:10] whatever you want so let's leave this to
[48:13] two and then get rid of this and then we
[48:15] can actually plug in another ultimate
[48:18] upscale here so I'm going to click
[48:20] search and then search ultimate again
[48:23] and then for model we can drag model
[48:26] here positive prompt we can use the same
[48:29] one negative prompt we can also use the
[48:31] same one vae we can drag that from the
[48:34] model and then upscale model we can also
[48:36] use the same one or a better way to do
[48:39] this is let me just select this node and
[48:41] delete it again is to click this press
[48:44] crl V and then over somewhere here press
[48:47] contrl shift V and it would
[48:49] automatically link everything that is
[48:51] linked from the node that you are
[48:53] copying now we don't want this original
[48:56] image to be linked here so let me get
[48:58] rid of this linkage and instead we just
[49:01] want to pass our 2x upscaled image to
[49:04] here to upscale by 2x again all right
[49:08] now the only thing that we need to
[49:09] change is because this image that is
[49:12] being passed here is now 1024 x 1024 we
[49:16] should set the tile width here to 1024
[49:20] by 1024 all right so now if we drag this
[49:23] out and select preview image let me run
[49:27] this and I'll show you the insane
[49:30] quality that this can generate compared
[49:32] to if we just did a normal upscale
[49:35] method like this all right so here is
[49:38] our preview image let me just save this
[49:40] first and then I'm going to pull both of
[49:42] these side by side on the left this is
[49:45] only using the 4X Ultra sharp to upscale
[49:49] my image of 512x 512 to four times and
[49:52] then this one is using ultimate upscale
[49:55] to upscale my image four times now
[49:58] notice the insane difference if I zoom
[50:01] in on this guy's face and zoom in on
[50:03] this guy's face notice how much more
[50:06] details this ultimate upscaler is able
[50:09] to generate his facial hair his eyebrows
[50:12] his eyes are super detailed the lighting
[50:15] on his nose also super detailed whereas
[50:18] for this one on the left even though
[50:20] it's the same resolution his face is
[50:23] just blurry and his facial hair does not
[50:25] look realistic his eyebrows his nose are
[50:28] super blurry and then same with the
[50:30] crown here you can see the details are
[50:32] really lacking in this left photo
[50:36] whereas for this one with the ultimate
[50:38] upscale everything just looks super
[50:40] sharp and crisp now notice that because
[50:43] we are using image to image with a den
[50:45] noising strength of 2 there are going to
[50:48] be subtle differences from the original
[50:51] image so for example you can see this
[50:53] dude is looking straight at the camera
[50:54] now whereas this guy is looking slightly
[50:57] to the right and that's because it
[50:59] doesn't just take the original image and
[51:01] upscale it so if you really want to
[51:03] retain a 100% of your original image
[51:06] then I think this method is better
[51:08] however if you're okay with changing
[51:11] some of the details to get a much
[51:13] sharper and more detailed image than
[51:16] ultimate upscaler is one of the best
[51:18] options out there so yeah at least for
[51:20] now ultimate upscale or basically this
[51:22] is tiled upscaling this is one of the
[51:25] best methods to to upscale images and it
[51:28] basically takes your image and breaks it
[51:31] down into tiles and then for each tile
[51:34] it does image to image but with a very
[51:37] low denoising strength so that it
[51:39] retains most of the original image but
[51:41] it just adds more detail to that image
[51:44] and then it glues all these tiles back
[51:46] together to give you your upscaled image
[51:49] this is one of the best upscaling
[51:50] methods out there right now so that
[51:52] covers upscaling next let's move on to
[51:55] some more complex stuff
[51:57] all right next I'm going to show you how
[51:58] to use control net in comy UI now what
[52:01] is control net and why do we need to use
[52:04] it basically it's a tool to really help
[52:06] you customize your image if you're
[52:08] serious about image generation you got
[52:11] to learn control net so for example you
[52:14] can really control the pose of your
[52:17] Generation by using an open pose
[52:19] pre-processor so you would upload a pose
[52:21] like this and I'll show you how to do
[52:23] that in a second and all your
[52:25] Generations would all align with this
[52:27] pose so here's another example and you
[52:30] can see all these Generations follow
[52:32] this pose to some extent and you can
[52:34] also adjust well how much do you want
[52:36] your generation to follow your pose
[52:38] here's another example it's so very
[52:40] powerful tool instead of pose you can
[52:43] also create a depth map and use that as
[52:46] a reference so for example if you upload
[52:49] this depth map all your Generations
[52:51] would have the same depth map to some
[52:53] extent including the lights including
[52:55] the laptop top this is a really powerful
[52:58] way for you to control what objects show
[53:01] up in what areas in your image and then
[53:03] here's yet another example this is
[53:06] especially useful if you have multiple
[53:09] characters and the scene is very complex
[53:11] but you really want to control where
[53:14] those characters are in the scene then
[53:16] again this is a great tool to give you
[53:17] more control over those settings instead
[53:20] of a depth map you can also upload
[53:23] something called a cany pre-processor
[53:25] which is basically just lines and you
[53:27] can see your Generations would follow
[53:29] this canny image here is another example
[53:32] so control net is very versatile there
[53:35] are so many things you can do with this
[53:37] something that's very similar to canny
[53:39] is line art so it's essentially the same
[53:42] thing you take an image and you break it
[53:44] down into line art and then use that as
[53:47] a reference image for your future
[53:49] Generations so you can see all of these
[53:51] Generations follow this line art to some
[53:54] extent here are some additional examples
[53:56] and if you want to generate anime
[53:58] there's an even better line art
[54:00] pre-processor called anime line art and
[54:03] this is more optimized for anime so as
[54:05] you can see here if you want the same
[54:07] pose the same outfit for your character
[54:10] but maybe you want different colors
[54:11] different backgrounds well you can use
[54:13] this option to generate those images
[54:16] here's another example and then similar
[54:18] to line art there's also scribble where
[54:21] you can draw in some lines and then that
[54:23] would also influence your generations to
[54:26] some degree as you can see in these
[54:28] examples this is also a good one so you
[54:31] can take any image and break down that
[54:33] image into different segments and use
[54:35] that as a reference and you can see all
[54:37] your future Generations would also
[54:39] follow the guidance of this image here's
[54:41] another complicated example and you can
[54:44] see it's able to control for all these
[54:47] objects very nicely you can see with
[54:49] this segment pre-processor you're able
[54:51] to control the location of all these
[54:54] people all these objects very precisely
[54:56] l in your image all right so let's Jump
[54:58] Right In how do we use it so first of
[55:00] all let's start with a very simple text
[55:03] to image workflow this is just your
[55:05] positive prompt negative prompt you're
[55:06] taking in an image of random noise
[55:08] you're plugging it through this case
[55:09] sampler it's going to decode it and give
[55:12] you your final image now let's start
[55:14] with adding control net first of all in
[55:16] this manager section let's click on that
[55:19] and then we'll click on model manager
[55:21] and then we'll search for control net
[55:23] Union this is the newest contr control
[55:26] net model and it basically includes all
[55:29] of these options that we just talked
[55:31] about so you don't need to go in and
[55:33] install all of those pre-processors
[55:36] separately so let's go ahead and install
[55:39] this note that it is 2.5 GB so depending
[55:42] on the speed of your internet this might
[55:44] take a few minutes to download all right
[55:47] after we've installed it note that we
[55:49] need to click the refresh button so
[55:51] let's click close here and then close
[55:54] and then click refresh all right so how
[55:57] do we use control net it might be not
[55:59] intuitive but we don't actually link
[56:02] control net to the latent image we
[56:04] actually link it to the positive prompt
[56:07] so let me select this node and delete it
[56:09] first and then I'm going to move up here
[56:11] for a bit so let's drag this connector
[56:14] out and then search for apply control
[56:17] nit so again it's not intuitive but
[56:20] control net is actually applied to the
[56:22] positive prompt before going into the
[56:25] case sampler so we are going to drag
[56:27] this conditioning connector back into
[56:29] the positive connector of the K sampler
[56:32] all right and then the next step is we
[56:34] need to select a control net model so
[56:37] let's drag this out and then it's just
[56:40] the first option here which is control
[56:42] net loader so if that control net Union
[56:45] was the only thing you've installed this
[56:47] is the only model you should see if not
[56:49] you can click on this and it would have
[56:51] a drop down of all the compatible
[56:53] control net models that you've
[56:54] downloaded but again this is the only
[56:56] only one you need this is the newest one
[56:58] and it contains all of these options for
[57:00] you so you don't have to go ahead and
[57:02] download each of them separately all
[57:04] right so the next step is the image
[57:07] first of all I'm going to double click
[57:09] here and then type in load image and
[57:13] then select this node and then let's say
[57:16] I want to specify a certain pose for my
[57:19] generation so I'm going to upload this
[57:22] image and I want whatever I generate
[57:25] down here to follow this pose so what we
[57:28] need to do is actually load this image
[57:31] to a preprocessor now in order to do
[57:34] that we need to download another node
[57:36] and you know this is kind of messy I
[57:39] wish we can just merge all of these
[57:41] nodes together into one node just to
[57:43] keep things cleaner but anyways it is
[57:45] what it is let's click on manager and
[57:47] then click on custom nodes manager and
[57:49] then we'll search for art Venture and
[57:52] then you should see this one comfy UI
[57:54] art Venture let's click on this to
[57:56] install it and I'll show you what this
[57:58] node does in a second all right so after
[58:00] we've installed this node it says we
[58:03] need to restart comy UI so let's click
[58:05] on restart and then click okay all right
[58:08] so we are back after the restart
[58:10] everything is still here so what we need
[58:13] to do is link our uploaded image to a
[58:16] preprocessor which is the node we just
[58:19] installed so let me drag this out and
[58:22] then I'm going to search for control net
[58:25] preprocessor and we should see this AV
[58:27] control nit preprocessor so let's click
[58:30] on this and then why did we install this
[58:32] instead of all the other options we
[58:34] could choose from because this node
[58:36] allows you to select from a lot of
[58:38] different options so for example we can
[58:40] choose sdxl which is what we are using
[58:43] right our checkpoint is sdxl and so the
[58:46] reason why we downloaded this node in
[58:48] particular is because it contains all of
[58:51] the pre-processors you need all in one
[58:53] node so basically you can select things
[58:56] like open pose or depth or canny or line
[58:59] art or all of these other examples that
[59:01] I showed you previously so let's start
[59:04] with the simple one let's start with
[59:06] open pose so I need to convert this
[59:08] image into an open pose image and then
[59:12] for the SD version we are going with
[59:15] sdxl and then resolution let's set this
[59:18] to 768 and then let's set the width and
[59:22] height of our final image to 768 as well
[59:26] note that for sdxl it's actually best to
[59:28] use 1024x 1024 and it doesn't have to be
[59:31] square but just to make our generation
[59:34] faster let's go with 768 and then
[59:36] finally we just need to drag this
[59:39] pre-processed image into the image node
[59:42] of control net and then here the
[59:44] strength is well how strong of an
[59:46] influence do you want this control net
[59:49] or basically this pose to influence your
[59:52] final image so 1 is 100% 0 is 0% % if
[59:56] you set this to zero then you're
[59:58] basically not using control net at all
[60:00] and so let's set this to something like
[60:04] 0.8 for example and see what that gives
[60:06] us all right so just a quick summary how
[60:09] you would use control net is it actually
[60:11] goes in between the positive prompt and
[60:14] the case sampler and then for control
[60:16] net what you need to do is upload an
[60:18] initial image and you need to process
[60:21] that initial image into whatever
[60:23] pre-processor you select and then it
[60:25] would turn that processed image and add
[60:28] it to the control net so actually what
[60:30] I'm going to do just to show you what
[60:33] this actually looks like is I'm going to
[60:36] drag a connector out here and then click
[60:38] on preview image so you'll see what this
[60:41] open pose image looks like and then what
[60:44] I'm going to do actually is hold control
[60:46] and select all these nodes and also
[60:49] select this one and then press contrl B
[60:52] to bypass them so that we are only
[60:54] running this I want to show you what
[60:57] these steps actually do so let me click
[60:59] Q prompt and note that the first time
[61:01] this loads it might take a while because
[61:05] you can see that it's downloading this
[61:07] open pose model from hugging face all
[61:10] right so after everything is finished
[61:12] downloading you can see the preview
[61:14] image here so basically this
[61:16] pre-processor is converting our uploaded
[61:18] image into this pose image and then
[61:21] feeding this into control net and this
[61:24] would influence the pose of our our new
[61:26] image so I'm going to hold down control
[61:28] again and select everything Press contrl
[61:31] B to unbiassed and then this time
[61:34] instead of a medieval warrior let's try
[61:38] a princess arctic tundra snowing all
[61:43] right and then let's click Q prompt
[61:46] perfect so if I drag my load image next
[61:50] to this final image you can see that
[61:54] this princess is following the pose of
[61:57] my uploaded image to some extent and if
[62:00] we want to follow her pose completely
[62:03] then we would set this value or this
[62:06] control net strength to one so that's
[62:08] one example of how to use control net to
[62:11] control the composition of your image
[62:13] all right here's another example so
[62:16] let's say I want an image similar to
[62:19] this composition of mountains but I want
[62:22] this to be sunset instead of this
[62:24] lighting so I can use a control net I'll
[62:27] upload this image and then instead of
[62:30] open pose I would select something like
[62:32] canny or you can also select depth if
[62:35] you want you could also select line art
[62:38] it really doesn't matter it really
[62:39] depends on your use case and then we
[62:42] leave everything else the same and
[62:44] here's just a preview image so after the
[62:46] pre-processor it looks like this this is
[62:48] what the canny pre-processor does all
[62:51] right so we plug this into control net
[62:54] and then this time for the positive
[62:56] prompt I just put in mountains and
[62:58] Sunset and then all these other keywords
[63:00] and then our final image looks like this
[63:03] so again if I drag my original uploaded
[63:07] image onto here just for a comparison
[63:10] you can see that our final image matches
[63:13] the shape of these mountains to some
[63:15] degree but now it's Sunset instead of
[63:18] this lighting how cool is that basically
[63:20] there are so many different options you
[63:23] can choose from for control net feel
[63:25] free to just play around with all of
[63:27] these pre-processors there are just so
[63:29] many different pre-processor options you
[63:31] can choose in control net to really give
[63:34] you maximum control over the composition
[63:37] of your image so that sums it up for
[63:40] control net if you run into any errors
[63:42] or issues just let me know in the
[63:44] comments below and I'll try to help you
[63:46] troubleshoot as much as possible but it
[63:48] should be fairly easy to install
[63:51] everything in just one click using this
[63:53] manager button all right next I'm going
[63:55] to show you how to install and use
[63:58] external AI tools on comfy UI the
[64:01] awesome thing about comfy UI is that it
[64:03] supports a wide range of other
[64:05] open-source AI tools for example there's
[64:08] a comfy UI note for mimic motion which
[64:11] allows you to create dancing videos from
[64:13] a single photo or there's another comfy
[64:15] UI node for tun crafter which is a
[64:18] powerful tool for generating anime
[64:20] scenes if you're not familiar with Toon
[64:22] crafter check out this video where I did
[64:25] a deep Di dive on how to install and use
[64:27] it but basically you just need to enter
[64:30] in a start frame and an end frame and
[64:32] this tool will fill in an AI animation
[64:34] in between those two frames plus there's
[64:37] another comy UI node for another tool
[64:40] called live portrait in this tool you
[64:42] basically input an image of a face and
[64:45] then you input a video of another face
[64:48] talking or doing some expressions and it
[64:51] can map those expressions onto your
[64:53] input image if you want to learn more
[64:55] about live portrait check out this video
[64:58] anyways today I'm going to show you the
[65:00] process of installing and using one of
[65:02] these AI Tools in comi for us we're
[65:05] going to use this tool called instant ID
[65:08] at its core it's basically a face swap
[65:11] that takes in a reference image of a
[65:13] person's face and then Maps it onto your
[65:16] generation so there are a handful of
[65:19] face swap tools you can use for stable
[65:21] diffusion such as RP or reactor or this
[65:24] one instant ID which I find to be the
[65:27] most realistic and best Fidelity so
[65:31] here's the original instant ID page you
[65:34] can see that it is really good for face
[65:37] swapping as you can see here's Taylor
[65:39] Swift here's some Chinese actress this
[65:42] looks really good it really does
[65:44] preserve the details of that person's
[65:47] face even across all these different
[65:50] styles of generations like it doesn't
[65:52] have to be realistic you can also do
[65:54] face swap for painting or drawings and
[65:57] it even works with different angles so
[65:59] even though you just have one image of
[66:02] Taylor Swift's face this AI is able to
[66:05] miraculously kind of guess what that
[66:08] face would look like at these different
[66:10] angles so anyways let's jump right into
[66:12] how we would set this up on the GitHub
[66:14] page which is called comfy UI instant ID
[66:17] I'll link to this in the description
[66:19] below if you scroll down a bit here are
[66:22] the installation instructions so you can
[66:25] either download or get clone this repo
[66:28] into the custom nodes directory or use
[66:30] the manager of course since we have
[66:31] manager installed we're just going to
[66:33] use the manager so going back to our
[66:35] comfy UI instance I'm going to click on
[66:37] manager and then click on custom nodes
[66:40] manager and then search for instant ID
[66:43] and then we're going to go with this one
[66:45] the comy UI instant ID native support
[66:48] and the nice thing about this one as it
[66:50] says in the description is that it
[66:53] implements instant ID natively and fully
[66:56] integrates with comfy UI so let's click
[66:59] on install and then if you open up the
[67:01] command prompt while this is installing
[67:03] you can see that right now it's cloning
[67:06] the repo it's downloading all the files
[67:08] all right so after that it says we need
[67:11] to restart comy UI so let's do that I'm
[67:13] going to click on restart and then click
[67:15] okay and then after clicking on restart
[67:18] you can see in the CMD window that it's
[67:20] actually installing some additional
[67:22] dependencies such as inside face so
[67:25] depend ending on the speed of your
[67:26] internet connection this might take a
[67:28] while to download all right you can see
[67:30] now it's downloading Onyx runtime GPU
[67:34] all right so we've installed the nodes
[67:36] we've installed Insight face and Onyx
[67:38] runtime but we are not done yet so next
[67:41] we need to download these Insight face
[67:43] models so one of them is called Antelope
[67:47] V2 which we can download here so it
[67:50] seems to be a zip file on Google Drive
[67:53] I'm going to click download and then you
[67:55] get download that anywhere once it's
[67:57] finished downloading open up the zip
[67:59] file and then it says we need to move
[68:01] this into comy ui/ models SL Insight
[68:04] face/ models so let's go into our comfy
[68:08] UI folder and then in models we need to
[68:11] create a new folder and call that
[68:14] Insight face and then within Insight
[68:17] face we need to create another folder
[68:19] called models so I'm going to create new
[68:21] folder models and then within models we
[68:24] should have this entel V2 folder so I'm
[68:27] going to just drag and drop this into
[68:30] here all right so let me exit out of
[68:32] this and you can delete the zip file
[68:34] afterwards and then we also need another
[68:36] one so we need this main model which can
[68:39] be downloaded from hugging face and then
[68:42] placed into this directory and it says
[68:45] you also need a control net and you need
[68:48] to place it in the comy UI control net
[68:50] directory instead of downloading it from
[68:52] hugging face there's a much easier way
[68:54] to download these which is through the
[68:56] manager so let me open manager again and
[68:59] then this time we are going to click on
[69:01] model manager and then we are going to
[69:03] search for instant ID and you should see
[69:07] that down here we have these two options
[69:10] the IP adapter and control net and this
[69:13] is for cubic instant ID so it's for this
[69:16] repo it's basically downloading this
[69:18] model which is based on IP adapter and
[69:21] this control net so we are going to
[69:23] select both of these and then click on
[69:25] install now note that this one is like
[69:28] 1.7 GB this is 2.5 GB so it's going to
[69:31] take a while to download all right so
[69:35] once we have these two installed let's
[69:37] click close And then close again and
[69:40] then let's click on refresh all right so
[69:43] let's start again with our very basic
[69:45] text to image workflow so we have a
[69:47] checkpoint here we plug in these prompts
[69:49] and then the these prompts go through
[69:51] this case sampler which takes in a
[69:53] latent image and then it decodes that
[69:56] image and it gives us our final image
[69:58] let me just adjust the placements of
[70:00] these to keep things more organized as
[70:02] we add the additional instant ID nodes
[70:05] all right so where the instant ID node
[70:08] goes is actually in between the prompts
[70:10] and the K sampler so if I drag this out
[70:13] here and then click search I will search
[70:15] for instant ID and I should see this
[70:19] apply instant ID all right so let's drag
[70:23] this over here I'm going to hold control
[70:26] and then hold shift and drag these nodes
[70:28] over here to keep things more organized
[70:30] all right so let's remove this and let's
[70:34] remove this so the negative prompt goes
[70:37] to here and then our model goes here and
[70:41] then the model is then connected back to
[70:44] this K sampler the positive is connected
[70:46] to the K sampler and then the negative
[70:48] is connected to the K sampler yes
[70:51] there's a lot of connections that needs
[70:52] to be made and I wish this process was
[70:55] simpler but it is what it is all right
[70:57] and then we need to also input an image
[71:01] so I'm going to drag this node out and
[71:04] then click on load image and then I will
[71:07] click upload and let me upload this
[71:09] image of Will Smith and then as the
[71:12] GitHub specified this also needs an
[71:14] instant ID control net so let me drag
[71:18] this out and then let's select control
[71:21] net loader and in your options if you
[71:24] search for instant ID you should only
[71:27] see this one instant id/ diffusion
[71:30] pytorch model so let's select this for
[71:32] the control net model and then for
[71:34] insight face we also need to drag this
[71:37] out and select instant ID face analysis
[71:40] and then for the provider since I have a
[71:42] Cuda GPU I will select Cuda and then
[71:45] finally for instant ID we also need to
[71:48] drag a connector out and the only option
[71:51] we see here is instant ID model loader
[71:53] so let's select this and by default it
[71:56] should just be this one IP adapter. bin
[71:58] which is the file that we installed all
[72:01] right so after all these things are
[72:03] connected let's also look at these
[72:05] values so the weight is basically how
[72:09] important do you want this face to be or
[72:11] how much influence do you want this pH
[72:13] to be in your final image and then for
[72:15] start and end again for stable diffusion
[72:18] it basically takes a latent image of
[72:20] random noise and through each step like
[72:23] right here we've set it to 20 steps and
[72:25] so for each step it removes a bit of
[72:27] noise until it gets to step 20 so for
[72:30] these start and end values it's
[72:32] basically saying well at what step do
[72:34] you want to start applying this face
[72:36] swap and at what step do you want to
[72:38] stop applying this face swap so let's
[72:40] say if we set this to like 0.5 or
[72:44] halfway basically this would stop
[72:47] applying the face swap at step 10 since
[72:49] we set the number of steps to 20 if we
[72:51] set this to 1 then it's going to apply
[72:55] the face swap at all steps right from
[72:57] Step Zero to the last step all right and
[73:00] then just a few more things we need to
[73:02] tweak for the width let's set this to
[73:05] 768 for the height let's set this to
[73:09] 1024 and then for the positive prompt
[73:12] let's say policeman realistic 8K
[73:15] Masterpiece Ultra detailed and then for
[73:18] the negative prompt we can say cartoon
[73:21] painting anime blurry Watermark all all
[73:25] right and if all is good we can click on
[73:27] Q prompt and see what that gives us so
[73:31] right now it's loading the checkpoint
[73:33] and then it's going through these
[73:34] prompts it's going through this apply
[73:36] instant ID which will take in our loaded
[73:39] image of Will Smith and all these other
[73:41] variables and then finally it's going to
[73:44] Output this image of Will Smith as a
[73:47] policeman so here we go now face
[73:49] swapping isn't just for generating deep
[73:51] fakes right generating fake images of
[73:53] real people you can also use face swaps
[73:56] to create consistent characters right if
[73:58] you want a certain character to have the
[74:00] same face throughout your video or
[74:02] throughout your animation or comic book
[74:04] or whatever face swap is a really good
[74:07] way to apply the same face to all your
[74:09] Generations let's try something else so
[74:13] the awesome thing about instant ID is it
[74:16] doesn't just work for realistic photos
[74:18] so you could also set this to let's get
[74:21] rid of painting and then instead of
[74:23] realistic let's set this to watercolor
[74:26] painting and then let's click Q prompt
[74:29] and see what that gives us all right
[74:31] perfect so now we have a watercolor
[74:34] painting kind of of Will Smith as a
[74:37] policeman so I hope this gives you a
[74:39] good understanding of how to install
[74:41] these different custom nodes of external
[74:43] tools and the awesome thing about comfy
[74:46] UI is there are a lot of other tools
[74:48] that it can support so for example
[74:50] there's also an animate diff node for
[74:52] comfy UI which helps you generate
[74:55] animations from images and inside the
[74:57] GitHub it should give you a
[74:59] demonstration of what a workflow should
[75:01] look like another person has created a
[75:03] comfy UI node for tun crafter and this
[75:06] allows you to take in one image as the
[75:09] input frame and then another image as
[75:11] the final frame and it would interpolate
[75:13] an animation in between these two frames
[75:16] so again here's the workflow for the tun
[75:18] crafter node and yet another user has
[75:21] created a comfy UI node for live
[75:24] portrait and this this basically allows
[75:26] you to take one input image and then one
[75:29] video of a person moving their heads and
[75:32] doing some strange expressions or
[75:33] talking and it would animate that input
[75:36] photo with this person's movements and
[75:39] here they shared with you how the
[75:41] workflow would look like in comfy UI so
[75:44] again there's just so many custom nodes
[75:46] that other users have created based on a
[75:49] lot of different external AI tools and
[75:51] that's what makes comfy UI awesome and
[75:54] with this manager you can easily search
[75:57] for all the nodes out there there are
[75:59] like tens of thousands of different
[76:00] custom nodes depending on what you would
[76:02] like to do and that's what makes comfy
[76:04] UI the most powerful free and open
[76:07] source image generator out there all
[76:09] right finally some of you might be
[76:11] wondering well how do I use comfy UI
[76:13] with the newest models such as oraflow
[76:16] or flux well for oraflow actually
[76:19] everything is the same this whole text
[76:21] to image workflow is the same and the
[76:23] only thing you need to change is the
[76:25] checkpoint you just need to change the
[76:27] checkpoint to this Aura flow saf enters
[76:30] file see this video on how to install
[76:33] and run oraflow with kyui and then for
[76:36] flux the workflow is quite similar you
[76:38] just need to tweak a few things like the
[76:41] model and a clip loader plus the K
[76:43] sampler see this video where I go in
[76:46] depth how to install and run flux on
[76:49] comfy UI however if you've watched this
[76:52] tutorial you should understand the basic
[76:55] of how to use comfy and how all these
[76:58] nodes work so changing between stable
[77:00] diffusion and flux and AA flow is
[77:03] actually very easy and for this tutorial
[77:06] I mostly used sdxl because it's still
[77:09] the most mature platform out there there
[77:12] are hundreds of models and luras you can
[77:14] choose from plus hundreds of plugins and
[77:17] tools such as control net and all of
[77:20] these only work with stable diffusion
[77:22] whereas for flux it's still quite new so
[77:26] there aren't a lot of tools and
[77:27] workflows built from the open-source
[77:30] Community yet but once we do have more
[77:32] of these tools let me know in the
[77:33] comments and if you want me to do an
[77:35] updated tutorial just for flux I'd be
[77:38] happy to make one as well so that covers
[77:41] my tutorial for comfy UI like I said in
[77:44] the beginning this is free and open
[77:46] source plus you don't even need an
[77:47] Nvidia GPU to use this and the
[77:50] installation is super easy if you
[77:53] followed all the steps as outlined in
[77:55] this video you should be able to get
[77:56] comfy UI up and running on your computer
[77:59] now in this tutorial we covered a lot of
[78:02] different topics from text to image to
[78:04] image to image to face swapping to a lot
[78:06] of different workflows so if you get
[78:09] stuck or you hit any errors along the
[78:11] way let me know in the comments below
[78:13] and I'll try to help you troubleshoot as
[78:15] much as possible as always I will
[78:18] continue to look out for the newest and
[78:20] coolest AI tools to share with you if
[78:23] you enjoyed this video remember to like
[78:24] share subscribe and stay tuned for more
[78:27] content also we built a site where you
[78:29] can find all the AI tools out there as
[78:32] well as look for jobs in AI machine
[78:34] learning data science and more so check
[78:37] that out at ai-
[78:38] search. thanks for watching and I'll see
[78:41] you in the next one