00:01One 2.2 Lightning
00:02Hell yeah, Hackerman is back
00:04I hear you were asking for me
00:06I'm back, baby
00:09Let's get started
00:12So, my friends, there's a little bit of Hackerman here because the Lightning Model is out
00:14They have a workflow here, the first file, the JSON file, and you don't really need that
00:20So, you don't need to download that, but I want to show you what it is
00:24When you open up the workflow, they have these two files here, and they are safe tensor diffusion files, and they are 25 gigabytes each
00:30Ain't nobody got time for that, and that's why we do none of that
00:37So, instead, what we do is download these two models here—high noise and low noise
00:43These are the LoRA models
00:46Here's the download button—one, two
00:50You want to rename them at
00:54The start is to write a 2.2 to image-to-video, that you know what it is, right? Just so to check it out
00:58Right, you go into the Laura folder, by the way, and then inside of ComfyUI, we're going to use the workflow that I showed you last time, and we're going to use it with the GGUF model that we have loaded last time
01:10If you do not have this GGUF model, it's going to go into the units folder
01:15It's pretty important to know that there is a high noise model and a low noise model, so you should already have that from last time
01:22I'm using the Q4 model here
01:24And then over here we have the Laura loader
01:27This is where we load these two Laura models that I just showed you, and then the rest is basically the same for that workflow
01:39One more thing you need is the UMiT5 XXL FP8, and so on
01:41Model
01:43You can see it here on the screen as big as possible
01:45Of course, also the VAE from version 2.1 for some reason
01:48It works
01:50Use it
01:53And then, of course, you need an image that you want to use here
01:56And then down here, you can set the size, which is going to become very important in a second
02:00So, where are we going from here? Up here, you have a prompt
02:06You can use a prompt, or you don't have to
02:09If the image suggests motion, you might be fine without a prompt
02:13If you want a specific motion, you might want to add a prompt
02:17Right? And then down here, we have some Chinese for the negative prompt
02:21Leave it as it is, don’t touch that
02:23And then here is very important
02:26Here we have the KSampler
02:30Now, the difference here compared to my..
02:34The last workflow is that this is four steps, and we are rendering zero to two steps in the first KSampler with the Uler sampler this time
02:39Then, in the second one, we go from the second step to the end, which is the fourth step, and this also uses the Uler sampler
02:46So, this is also different
02:49I'm going to save the workflow a second time, upload it, and it's linked below the video
02:52Check that out
02:55And then, of course, we have here our initial video
02:57If you want to, you can also use this interpolation to get some more frames and create a smoother video down here
03:02You don't have to, but you can do that
03:06It doesn't use too much render time
03:11Now, let's talk for a second about these render times
03:16You can see here I'm using 640 by 640
03:18For the resolution and 49 frames initially, and then, of course, this is going to be interpolated for more frames after that
03:25And this is what these render times refer to
03:28I have a 4080
03:30So, when we look at this, this is what we did last time in my previous video
03:36You can also watch that where we used eight steps
03:38All of that is rendering in 135 seconds for that, with this resolution
03:45Now, with this new Lightning Laura, what we have is 80 seconds
03:48So, that's a considerable time save for the same length, the same resolution
03:57However, I have a pretty cool hack for you if you want to save a ton of time on all of that
04:01What you want to do is go in here for the image-to-video one, this thing where you set the size, and you halve the size to 320x320
04:08This is I'm going to give you a low-resolution video, and you're not going to want to use that low resolution video, but it gives you a very quick preview, and that can save a lot of time for you because if you find something you like, you can still then double the resolution and render it again
04:26For that, you have to fix the seat here so you get the same result
04:32The render time for that on a 320x320 video with 49 frames is only 36.4 seconds on a 4080, which is super fast
04:37And that's already it
04:41Hell yeah, baby
04:46And don't forget to comment "Hacker Man" under the video
04:48Peace out.