[00:00]
You can now upscale your footage locally
[00:02]
to HD, 2K, or even 4K resolution. Our
[00:05]
workflow will break down your video into
[00:07]
smaller segments and upscale them tile
[00:09]
by tile, making it easier for your
[00:11]
computer to handle. It uses generative
[00:13]
upscaling, meaning it will actually
[00:15]
generate new information into your
[00:17]
video, making even pixelated footage
[00:19]
like this usable.
[00:22]
We built this workflow to switch over
[00:24]
from subscriptionbased paid tools, which
[00:26]
this workflow also outperforms in many
[00:28]
cases. and now we're giving it to you.
[00:30]
So, here's how you can upscale your
[00:31]
footage free and locally.
[00:37]
This workflow is based on the one 2.2
[00:39]
video model that lets you generate 720p
[00:41]
videos for a maximum of 81 frames, which
[00:44]
is by far not enough resolution or time.
[00:46]
So, to fix this, we built a custom
[00:48]
workflow for Comf UI, a free nodebased
[00:51]
interface for AI models. And to install
[00:53]
it, you can just follow the guide on our
[00:55]
website links below. Once you have it
[00:57]
installed, you just need to download the
[00:58]
JSON file and drag and drop it into the
[01:01]
Comfy UI interface. You'll need to
[01:03]
install these missing custom nodes. Do
[01:06]
that by going to the Confui manager,
[01:08]
install missing custom nodes, select all
[01:10]
of them, and click install. Once it's
[01:13]
done, restart Confui and the full
[01:15]
workflow is here. Now, we need to
[01:16]
download the actual models for this
[01:18]
workflow to work. And you can find all
[01:20]
of them to the left here in these yellow
[01:23]
notes. First, we need to decide which
[01:24]
kind of base model we want to use for
[01:26]
this workflow. If you have a good GPU
[01:28]
with like 24 GB of VRAM or more, you can
[01:31]
use this FP8 version of the model, which
[01:33]
will be just a tiny bit faster. To use
[01:35]
it, you need to download it from this
[01:37]
link and put it inside of CompuI models,
[01:40]
diffusion models. Then load it here.
[01:42]
Connect this model to this node right
[01:45]
here. But in most cases, I recommend
[01:47]
using the GGUF version of this model.
[01:49]
You can find all of the GGF models right
[01:52]
here. GGF is basically a way to compress
[01:54]
model size so that it can run on GPUs
[01:57]
with lower VRAM. This will cost you some
[01:59]
quality, but usually it's not that bad.
[02:01]
You can ignore all these high- noise
[02:03]
models here. We just need one low-noise
[02:06]
version. Check how much VRAM your GPU
[02:08]
has, and then you select one of these
[02:10]
versions that comfortably fits onto your
[02:12]
GPU. If you only have 12 GB of VRAM,
[02:14]
check out the Q5 versions. But I'm going
[02:17]
to use the Q8 version. Just download
[02:20]
that and put it in confusi models unit.
[02:23]
Make sure it's selected right here.
[02:25]
Next, we're going to use the light x2
[02:27]
vora which allows us to run this
[02:29]
workflow at a lower step count,
[02:31]
increasing performance by so much. You
[02:34]
just need to download the low-noise
[02:35]
Laura. Download that, put it inside of
[02:37]
Confui models and Loras, refresh, and
[02:40]
load it right here. The next two Las
[02:42]
here are completely optional, but if
[02:44]
you're upscaling a lot of realistic
[02:46]
footage, I recommend you get at least
[02:48]
this one right here, the stock
[02:50]
photography. What this one does is
[02:52]
basically it just pushes the style
[02:54]
towards more realistic video. Next, you
[02:57]
need the clip model. You can get that
[02:58]
right here, the VAE. Put it in VA and
[03:01]
load it here. And next, we need an
[03:02]
upscaling model. You can use any
[03:04]
upscaling model here that you can
[03:06]
install via the CompuI manager. I
[03:08]
recommend really going with this one. So
[03:10]
to install it, go to manager, model
[03:12]
manager, search for X2, then download it
[03:16]
right here. I already have it installed.
[03:18]
That's all you need to do to set up this
[03:20]
workflow. If you want to skip all these
[03:21]
installation steps, or maybe you don't
[03:23]
have the most powerful computer, you can
[03:25]
also run this workflow on RunPod. Runpot
[03:28]
is a cloud GPU platform that we know a
[03:30]
lot of you guys use. So we built this
[03:32]
template for you that just starts up the
[03:34]
workflow with everything installed for
[03:36]
you. You just need to follow the link in
[03:38]
the description. Set up your account on
[03:39]
Runport, add some credits to rent a GPU.
[03:42]
I usually run this on a 5090, which
[03:44]
costs around 89 cents an hour. Now,
[03:47]
let's upscale a video. The cool thing
[03:48]
about this workflow is that we made it
[03:50]
as plugandplay as possible. There is
[03:52]
actually a lot of stuff happening in
[03:54]
these groups here in the background, in
[03:56]
these subgraphs that you never have to
[03:57]
check out or open. There's a bunch of
[03:59]
math happening here, but you don't need
[04:01]
to care about that at all. You can just
[04:03]
load the workflow and choose a video
[04:05]
right here. Let's take for example this
[04:07]
here. This should be a very good
[04:09]
challenge. You can see the quality is
[04:11]
just horrible and we're just trying to
[04:12]
be able to use it. Next, you can set the
[04:14]
frame load cap and this should
[04:16]
theoretically work. You can upscale
[04:18]
videos with any length, but this is
[04:19]
really intense. So, I'm just going to
[04:21]
cap that at 81 frames. Next, you can
[04:24]
come down here and this is where you set
[04:25]
up the whole workflow. Put your final
[04:27]
output resolution. And I'm just sticking
[04:29]
with 2K here. I feel like 2K is a very
[04:32]
good starting point where it adds a lot
[04:34]
of detail, but it also doesn't take too
[04:36]
long. And then you can choose if you
[04:38]
want to enable this here for higher
[04:40]
quality. I would say this adds about
[04:41]
like maybe 30% of quality, but the
[04:44]
problem is it takes like nearly twice as
[04:46]
long. So I usually do this for smaller
[04:48]
upscaling sizes like this one right
[04:50]
here, 2K or HD. But when I go to 4K, it
[04:53]
really is not worth the time. So I
[04:55]
usually switch that off for 4K video. So
[04:58]
here we have this creativity slider
[05:00]
which just tells the workflow how much
[05:01]
it can change in the image. And I
[05:03]
usually start pretty low, something like
[05:05]
this right here. Finally, these are the
[05:07]
iterations that the workflow will
[05:08]
automatically create. So if you have a
[05:10]
video that's longer than 81 frames, it
[05:13]
will create multiple 81 frame batches
[05:16]
and then blend them together so you
[05:18]
don't realize that there are any seams.
[05:20]
For me, it's 81 frames. Don't go above
[05:23]
that, but you can go below that if you
[05:25]
don't have the best GPU. 41 frames is
[05:27]
also pretty good. And next are the
[05:29]
overlap frames. So, how many frames does
[05:31]
this workflow have to blend these
[05:33]
iterations together so you don't notice
[05:35]
that there are any seams? We realized if
[05:37]
you want to upscale really long videos
[05:39]
to like 4K resolution, in the end, it
[05:42]
can actually crash compi when it's
[05:44]
stitching everything together. So, if
[05:46]
you activate this setting right here,
[05:47]
you always have all the images that it
[05:49]
already upscaled as a fallback. So you
[05:51]
can import them into After Effects or
[05:53]
your video editing software of choice
[05:55]
and then just stitch the video together
[05:56]
there. So now we can just click run and
[05:58]
as you can see it will automatically
[06:00]
create a very detailed prompt for this
[06:02]
video. If it's not perfect or you want
[06:04]
to add more detail, you can come up here
[06:06]
and manually input something here. You
[06:09]
can also input a negative prompt, but
[06:11]
this is usually enough. So all I need to
[06:13]
do now is wait for this workflow to
[06:15]
finish. Now, while this is running, this
[06:16]
is a good time to mention that this
[06:18]
workflow and video are sponsored by our
[06:20]
amazing Patreon community. If you want
[06:22]
access to exclusive example files,
[06:24]
advanced versions, and our amazing
[06:25]
Discord community, consider supporting
[06:27]
us on Patreon. Your support makes
[06:30]
creating these workflows and sharing
[06:31]
them with you all for free possible. And
[06:33]
it also makes it possible that we don't
[06:35]
have to rely on sponsors in every single
[06:37]
video, which is just amazing. Thank you
[06:38]
so much for that. 10 minutes later and
[06:41]
the video is done. So, let's look at the
[06:43]
result. And you can see this was the
[06:45]
original video and this is after. Before
[06:49]
after. You can see how much detail it
[06:52]
was able to generate, especially in the
[06:54]
hair here. But it's still not perfect.
[06:56]
We could improve this result even
[06:58]
further if we allowed a bit more
[07:00]
creativity. You can see the hair is now
[07:03]
more detailed and the overall quality is
[07:06]
higher. If you use a high creativity
[07:08]
value, it can be more likely that you
[07:09]
see these tiles here. And sometimes you
[07:11]
just have to experiment what works best.
[07:13]
But here, for example, I tried the
[07:14]
highest possible value and I did not
[07:16]
have the problem with the tiles at all.
[07:18]
We also upscaled the same video using
[07:20]
different upscaling models by Topaz.
[07:21]
Even the 4K output from Topaz doesn't
[07:23]
compare to the better 2K output from our
[07:26]
workflow. Now, let's talk about the
[07:28]
creativity value. The higher you set the
[07:30]
value, the more it will modify your
[07:32]
video content, and this might change
[07:33]
your subject a bit too much. So, if you
[07:35]
need a super precise upscale that stays
[07:37]
faithful to every detail in your
[07:39]
footage, keep that creativity low. The
[07:41]
trade-off is that you'll get less
[07:43]
dramatic quality improvements. Overall,
[07:45]
we also tested this against Flash VSSR,
[07:47]
another local upscaler. There is a
[07:49]
version for Chromei, but it does not
[07:51]
produce the best results yet. We
[07:53]
recommend installing it via Pinocchio.
[07:55]
Pinocchio is this one-click installer
[07:57]
for AI tools, and it makes the
[07:58]
installation super easy. Flashvsr is
[08:01]
great at precision. It sticks very
[08:03]
closely to your source material.
[08:05]
However, it really struggles with
[08:07]
extremely lowquality videos. But when we
[08:09]
used 720p footage as input, it actually
[08:11]
produced sharper results than our
[08:13]
workflow could produce yet. But as you
[08:15]
can see in this shot, for example,
[08:16]
especially with the grass, it's also not
[08:18]
perfect. So when should you use our
[08:20]
workflow? It shines in situations where
[08:22]
perfect accuracy isn't your top
[08:24]
priority. For example, when you're
[08:26]
upscaling AI generated videos or videos
[08:28]
with very poor quality. In these
[08:30]
scenarios, our workflow outperforms
[08:32]
FlashVSR because it can actually
[08:34]
generate new plausible detail instead of
[08:36]
just trying to preserve what's barely
[08:38]
there. I hope you enjoyed this video and
[08:40]
found these workflows useful. If you
[08:42]
create something with it, feel free to
[08:43]
send it to me or tag me in your work. I
[08:45]
always love to see what you come up
[08:47]
with. Thanks for watching and thank you
[08:48]
to our lovely Patreon supporters for
[08:50]
making these videos possible. See you
[08:52]
next time.
[09:07]
Hey.