00:01Hello ladies and gentlemen
00:04I'm back after a small hiatus, and we're going to talk about Wan 2.1, and we're going to actually build this workflow right here because, why not? That's what we do on this channel
00:15We show you how things work
00:17And the greatest thing about this is that it allows us to go back to our good old buddy, the K sampler
00:23How it all really started on this channel, or at least the tutorial started on this channel
00:27This version of the one I'm going to show you is the ComfyUI native version
00:31So, it allows us to use stuff like this because it works directly with it
00:36And that's what I like about it the most
00:38That's why I've chosen this one
00:40And it's very simple and straightforward to me, and I'll show you how
00:44It could be very simple and straightforward for you
00:46Plus, we're going to just talk about how everything works in there, too
00:49That's the goal of this, and we're going to focus on the image-to-video workflow because I feel like people want to see that more than they want to see the text-to-video one
01:00So, that's the one we're going to build on today
01:04We're going to start just like we used to, just like the old-fashioned model
01:06Just take it out here, and it's already loaded up
01:12Oh, by the way, to be able to utilize this, you need to upgrade ComfyUI to the latest version
01:20If you don't do that and you're not up to date, you won't be able to have this option for utilizing it, so no matter what workflow you're using
01:26Somebody out there, you gotta do that
01:28You have to update your ComfyUI to the latest version
01:29And we're going to use, I think, yeah, we're going to use that one
01:34This specific one right here, this model
01:36Also, when I load up all these things, I'll show you — I'll leave a link in the description to where you can find all the things you'll need, the models you'll need for this particular workflow
01:41So don't worry about that
01:46All right
01:51I can show you where to get what you need to make this thing work
01:53So the model comes out here
01:55The greatest thing I love about this, too, is that we get our positive and negative prompts back
02:02They come back, and I love this new updated AI because I can just change the color
02:05Red means negative
02:09And green means positive
02:12I love the new workflows on this thing
02:15Right, so we're just going to do this right here for now
02:17This is not exactly how the workflow is going to look, but for the sake of building it out really quickly, that's what we're going to do
02:23And because I love to do this on the other side too, VAE decode
02:28And then you get to here and you load your VAE, and it's going to be the 2.1 version
02:40Then come out here, regular CLIP loader, and this one's going to be this weird one called UMP
02:49I'm going to use the scale version, or the T5XXL text model, or T5XXL, blah blah blah
02:59Part of this you'll probably be familiar with
03:01Change this to T5XXL
03:05If you've seen any of the Flux tutorials, you might remember that we used a type of..
03:10The clip model that was called T5XXL
03:13This one's called UMP T5XL
03:16It basically works the same way
03:19The difference is that this one is like some universal multilingual model
03:22And I believe you can put Chinese characters in this thing, obviously, because, you know, that's where the model—like the whole thing—comes from
03:30But this thing, you know, maybe a bit bigger
03:33I don't know, because it can understand multiple languages
03:36I'm not quite sure about that; perhaps the other one could too
03:38I just never tried because I don't really speak Mandarin
03:41But the idea is that it works the same kind of way
03:45But here's the really cool part about that: since we're using something that is a large language model, just like we had in Flux, That means we can speak to it like in regular sentences, not a whole bunch of tags and stuff like that
03:58But now we get the added bonus of having the negative prompt back, which we didn't have in the Flux dev models
04:04So that's pretty freaking cool, right? All right
04:11So I know I modified this workflow from the original one they have available just a little bit
04:14Not too much
04:16One of the major things I changed was I wanted to take this and turn it into video output because we want to see a video
04:23Like, isn't that what everybody really wants to see, right? That's what I wanted to see
04:29And I think the frame rate this thing commonly works with is around 16 frames
04:36I'm going to change this to DDIM because I prefer it for..
04:40This type of thing
04:45Although you can use Uler a lot, actually try out all of these if you have the time, and you can see how much you know
04:49You have these different schedulers too
04:51There's a lot of combinations
04:53Let's go back to the KSampler, try out some stuff
04:55The Uler thing—the reason why I don't use it—is because a lot of times it seems a lot more stiff when it comes to video, but it does maintain a little bit of quality, I think
05:05Maybe a little bit better, not quite sure
05:08I have to really test that a little bit more
05:09It's, you know, still kind of new for me
05:10I haven't tested everything out, but fair enough
05:15We'll keep moving forward, right? All right
05:17So, it's a little bit different from what we're trying to do
05:20So here
05:22The thing that we're going to have to change is, you see, it's the latent image
05:25Now, we're not going to put just a latent image down here
05:29The main reason why is because we want it to be image-to-image, right? So, that won't work
05:32But they have created an image-to-video workflow
05:42Think of this as a pre-processor that takes all the information from the left side of our workflow, pre-processes it, and gets it ready for the actual processing in the KSampler
05:55So, we can just basically disconnect this part right here because this is going to run through
06:01The negative input goes through here, the positive through here, the VAE output comes out here, and the other goes to your decoder
06:09So, that doesn't change
06:10Then, take this, put it through
06:12Here
06:14Right? Negative through the negative
06:18Matching colors
06:21You've done this before, right? Okay
06:24Now, we take this clip Vision output
06:28We're going to need this clip Vision encode
06:31Clip Vision, which is going to be the clip Vision loader
06:34We're going to use this one called clip Vision H
06:38All right
06:41Then, image Load Image
06:44Now, this is basically the part that interprets how you're, how to like compute the objects and things inside of your image
06:53That's what this clip Vision thing is about
06:54We also have to take this image out and set it as a start image because we want the video to start with that, right? So, it goes two different ways
07:03This is basically the entire workflow
07:06This is all that you need right here to make
07:08This thing works
07:10You don't need anything else
07:12Well, you need an image in here; otherwise, it's going to give you an error
07:17In other things that we've created, we don't put a prompt in there, right? It's like, what happens if you don't prompt it? Will it do anything? Well, maybe, maybe not
07:25I actually haven't tried that on the video models yet
07:27Well, at least not this one
07:29I might want to try just for the sake of it
07:32I'm going to lower this to make it a little bit faster
07:34I think 65
07:37Yeah, 65 is around
07:39It has these weird estimates
07:41So, this is the number of frames, by the way, it's going to produce
07:43Batches— you can make more than one batch, right? This is the height and width
07:47You know, try to make it match up with your image if..
07:51You can
07:54If not, again, you can always use this crop tool if the image is too big
07:56But again, that's what it will do — it will crop off the edges
07:58I'm going to remove this crop now
08:01So, this is a working workflow; it will work
08:07Now, before we run it, I just want to show you some things again about where to find this stuff
08:11These locations are found within the directories on your computer
08:18Usually, they'll be under ComfyUI backslash or forward slash, depending on whether you're using Windows or Linux
08:22I'm using Linux, so it might be a little bit different in how everything's laid out or how the slashes work
08:30But these specific folders are written this way on purpose
08:34So, as you can see for me, I..
08:36I have ComfyUI models, and then you'll see I have the Fusion models section
08:43I have one, and where are the other ones? The other ones are named Text Encoders
08:47There's one Text Encoder, another named VAE, and up here is CLIP Vision
08:53That's where you're going to put all these different models that you see, like the CLIP Vision H and all that kind of stuff
09:00So it shouldn't be too hard to figure out where to find this stuff
09:03But I just want to take a quick look at this
09:08These are all the different types
09:09Now, the one we're using right now for this tutorial is this one, version 2.148, or whatever it is
09:15I usually don't talk about installation details, but I will mention it just for the sake of clarity
09:18I am going to cover this because you can get caught up in this
09:22Depending on how your system works, you may or may not be able to run all these processes
09:25I'm not really sure about everybody's specs and who can run what, but the heavier tasks, of course, are the 720 processes and the ones that are not FP8 scaled or FP8 E4 M3 F, blah blah blah
09:38All right, so if you see this big number 14B, the simpler it looks, like BF16, these are the ones that are going to be really intense on your systems
09:46Keep in mind, I2V means image to video, and T2V means text to video
09:56So everything we're utilizing in this tutorial will be from here and upward
10:00Again, specifically, I'm going to be using this one right here: the 1.2.1 Image to Video 480p 14B FP8E4
10:08Know, there you go
10:17Anyway, I will leave the link right here, in the description of this video, so you can find this stuff and get what you need for your system
10:25Everything else should be pretty straightforward
10:27I'm still using this as well, because, as you can see, this is only 6.74 GB, so you can see it's a lighter version of it
10:36And the VAE is VAE—you got to get that
10:37That's lightweight
10:39Anyway, text encoder, blah blah blah
10:40I think I just looked at that
10:42All right, I think I'm moving in circles now
10:43Anyway, back to the tutorial
10:48So, all those things are required
10:50The CLIP vision again is what interprets what's in the image, to let you know or let it know what it is
10:58It's looking at, so that way it can devise a way to animate it
11:00This is basically your text encoder, and it's a large language model, so you can use regular speech
11:07We also get back the ability to have negative prompts, so we can tell it what we don't want in the output
11:12This is a VAE, it's your variation encoder, blah, blah, blah
11:17I'm going to remember what that A means at one point
11:18One day, I'm going to remember what that means again
11:20And of course, you need your big model here that understands everything
11:24Your pre-processor, because that's basically what it is
11:29It takes into account this
11:30You're using text-to-video, you don't need this, of course
11:32It's going to be quite different again
11:34Like I said, this idea..
11:35That takes all these different things into account
11:37You're what you're telling it to do
11:38What are the starting images? What it interprets through the CLIP vision model of what it is looking at, then taking that through the K sampler, sampling it up, processing the thing, and then decoding it down here, which will give us a combined image
11:58This is all you need to make it work, and let's see what happens with just the image zero prompt
12:07I want to see this happen just as much as you do, but I'm going to have to pause this video to do this because this thing is going to take up a massive amount of processing power to do
12:16So, that being said, let's get to it
12:20I forgot one thing
12:24I stopped it because I just remembered I wanted to output
12:27Not as an image GIF
12:30I wanted to output as a video, so just changing that to MP4
12:33Everything else I'm going to keep exactly the same
12:35That's the only thing I want to change
12:37But, let's rerun it
12:43[Music] All right
12:46So, we have an output, and well, it looks like something
12:51It does look like I didn't really describe what it is they're looking at, but it does look like a successful animation despite the fact that I wasn't really descriptive about what it was
13:01And it looks like it might be underwater
13:03I don't know
13:05But that's where the descriptions come into play
13:08So, I'm just going to do a little bit of prompting, just to show you a little bit of how you could talk to a video versus talking to an image
13:16Uh, a static camera shot of a Pulsating, how do you say pul-? Oh my god, I'm so terrible at spelling
13:31I'm not going to ask Google this time
13:33I swear I'm not going to ask Google
13:34A static camera shot of a pulsating..
13:36Oh, no
13:38Okay, let's do this
13:39A red pulsating aquatic creature being held by an underwater diver
13:46Underwater
13:53I'm just going to say underwater again
13:59I probably don't need to do that, but anyway, let's just run with that
14:04So, the idea is that we're going to get it to not move around so much because you can see this one
14:11Just make this a little bit larger
14:14You can see the camera just moves around a little bit more, and we're not really talking about anything else
14:18And I'm not being too descriptive
14:22I could be more descriptive about the diver
14:28Themselves
14:34And you know, a static camera shot of two viruses underwater, blah blah blah blah blah, but I'm going to show that that's not necessarily something you need to talk about so much to get what we wanted to do, because it'll still recognize them
14:40As you can see, it still seems to recognize at least that that's an arm, and maybe it's understanding these as bubbles
14:47It seems like it might be understanding them as bubbles, but who knows
14:50I don't know if it understands this aquatic creature, but we're going to talk about it that way because we made a reference to it by it being red, and it's the only thing in the actual image that is red
15:00So, let's go with that
15:04Right back in the flash
15:07Okay
15:09And there we have it
15:12Now, it's still a..
15:14A little bit of movement, but it is following the subject we talked about
15:18We discussed the red pulsating aquatic creature as the most important element, and we did talk about it underwater
15:24Now you see the bubbles start to flow more actively
15:26You see they begin to pay attention to this creature, and you notice how it now has these little, you know, fins or whatever that are wiggling around
15:35 Another way we could approach this is by changing the focal point
15:44So, the focal point for this scene was the red pulsating aquatic creature
15:49Let's do this: let's first focus on the divers
15:51That will give us a better perspective
15:55So, let's..
15:58Say two underwater divers
16:02One diver
16:05Oh, let's make this a period
16:07End of a statement
16:10The diver is a red pulsating aquatic creature being held in a static camera shot of two underwater divers
16:18One diver is holding a red pulsating aquatic creature underwater
16:26Let's go with that
16:34So, the idea is still to have a static camera shot where there won't be a lot of movement, and the focal point will be on the divers because we talked about the divers first
16:42We did mention a bit more description because we said pulsating, so maybe we should describe the divers more
16:48A static camera shot of two underwater divers breathing, with air bubbles released from their masks
17:02All right
17:05How about that? Yeah
17:09A static camera shot of two underwater divers with air bubbles being released from their breathing masks
17:12One diver is holding a red pulsating aquatic creature underwater
17:14Let's go with that
17:16Let's just go with that
17:19I want to refine a little bit more, but let's just see how much more it will focus on the other thing
17:22Let's go
17:25Okay, so we have our result
17:29Now, you can see that, as much as we specified static, right? It almost seems more like a tracking shot
17:37And I think the real reason, especially in this situation, you'll notice that unlike the last video, you see more of the divers, even though it didn't really do much with the bubbles
17:48You'll see that it seems to want to keep them in the frame
17:51Frame as well
17:53But we did talk about the thing that they have in their hand
17:55So, it shows bubbles, but it doesn't show them coming out of the breathing mask or the mask
18:01Did I say breathing mask? I didn't say breathing mask, right? Oh, I didn't spell it right
18:03Holy crap
18:05God, me and spelling, you know, I would..
18:07it would suck if that's actually the reason why I didn't do it
18:11But in any case, you can see, like, a static shot, I believe, mostly just to focus on the characters or a static camera shot
18:24But if it's not a situation where there would be much movement, I think it would really stay static, like, really, really, really well, because I've done it in other situations
18:35But for the most part, usually static, maybe
18:37You can say "still shot," too
18:40There are other things you can say as well, but for the most part, you can see it does try to listen
18:46I'm going to run it again, too, because I want to see if it actually does that breathing thing I mentioned before
18:49Let's say this
18:52Let's do this
18:55Remaining still
18:57And we already have the breathing mask thing
19:01All right, so, let's run it again
19:03Let's see what we get
19:06Okay, so another iteration
19:09We got more bubbles out of them
19:12We didn't necessarily get it to come out of the mask, but it did recognize the need for bubbles
19:15I guess my description isn't the best
19:17They probably look more like snorkeling masks than breathing masks
19:22But as you can see, it remains still a little bit more because I did say
19:27The creature remains still, so that might have helped it like freeze frame a little bit
19:34But using keywords like that, I think, will help maintain the image in a certain way
19:39I have used the word "tracking" before, so I know that word works, especially if you want to follow something through actual motion as well
19:46And also like "crash zoom." Crash zoom actually works, so it'll just zoom in really quickly on a face or something like that
19:56Yeah
20:00Okay
20:05The camera pans to a close-up of two underwater divers' faces with air bubbles being released from the transparent masks they wear on their faces underwater
20:09I said the word "faces" again because I just want to make sure that they know that the transparent masks are on the
20:12Faces
20:15I don't, I can't think of another place
20:17Well, how did this happen? I think it's just random press
20:19I think that sometimes, like, you know, sometimes this thing can interpret things in ways I wouldn't necessarily imagine
20:25So, try to double down on certain things to make sure
20:28So, we'll go with that and see what happens
20:32Yep
20:35All right, we're back
20:39And it's pretty interesting
20:43It's a whole lot of bubbles, more than I was expecting, but it does pan
20:47It does show their faces
20:51Adds a new character in there, but eh, pretty good
20:55So, there's a lot of ways you can really think about controlling the camera
20:59I probably should have set a direction specifically
21:01I didn't specify a direction of panning, but whatever
21:05There's a lot
21:07Of things you can do with this
21:09You should definitely just test it out yourself
21:10Have fun
21:12See what you can come up with
21:13I'm just trying to show a few rudimentary things
21:15This thing right here
21:18Remember, when you're changing the frame rate, you're also changing how fast it goes through this length of frames
21:27So, keep that in mind, too
21:29If you were to change this to like 24 frames per second, then just realize that you're going to get a lot fewer seconds of footage, if that makes any sense
21:40Because, you know, you're using up 24 frames pretty quickly
21:42Probably, you're going to get less than three seconds; I'm pretty sure
21:48So, make sure if you increase the frame rate, you can increase the length of the video accordingly
21:54This too
21:59But also realize that just like with all of these video generators, at this point in time, there's a limit to how long you can really make them without the morphing issue becoming a problem
22:04And it's usually around 6 to 8 seconds before it really starts to happen
22:10Sometimes it happens sooner than that
22:11Also, when stuff moves out of frame, for the most part, I feel like Juan tends to remember, but of course, when you're adding more and more elements, angles, and so on, you want to make sure you create your source images outside of this process so they can match up
22:29That's what the Loras are really for
22:30I like using Loras on the images as the key frames more than using Loras within the video
22:35But since this setup is the way it is, it’s possible to put layers in this as well
22:42But that’s a different topic
22:44I just wanted to really highlight how to utilize the image-to-video workflow
22:48I will release this, just like every video I post on YouTube
22:56If I create a tutorial on it, I will also have a workflow for this on Patreon for free
23:02So, there’s also more in-depth content that I plan to include, especially about video generation and more creative applications
23:12I want to explore more topics like 3D animation and potentially game building
23:16There are some really cool things you can do with these AI generators that could help you create something amazing
23:25A lot larger and a lot more ambitious than just stuff like this
23:29But a lot of that, again, is a little bit more advanced topics
23:35I try to keep everything on YouTube with the basic stuff; more in-depth topics I'll put on Patreon
23:40But again, if you want this workflow, it will be on Patreon for free
23:44The link will be in the description of this video
23:48Thank you for watching
23:51I hope this was informative, and I hope you guys have fun with this, because I'm having a whole lot of fun with it
23:55I'm so glad we have a really, really good open-source video generator, and to me, that really stands up against all the other closed-source ones right now
24:04I can't even imagine where this is going to be in the next few months, let alone by the end of this year
24:11[Music] [Music]