In this video, I will show you everything you need
to know about upscaling in ComfyUI. I'll guide you through the various methods to enhance your
images, introduce you to my favorite upscaler, and finally share my personal workflow with you.
So, let's get started without any further delay. If you've seen my last video, you're already
familiar with how ComfyUI works and how we've set up this workflow with nodes. If you
haven't, I highly recommend watching that video first to better follow along. To
kick things off, I'll introduce you to the standard upscaler of ComfyUI. To do this, I'll
create an "Upscale Image By" node and connect it to a "Preview Image" node so we can compare
the original image with the upscaled version. I'll then adjust the settings to
upscale the image by a factor of 2, which determines the new size of the image.
As for the method, I'll keep it as is. It's important to note that the method affects how
the image is upscaled; usually, it's best to leave it unchanged. Let's start upscaling.
Oh, and one more thing – it's a good idea to keep the seed fixed to ensure we don't generate a
new image, but rather keep everything consistent. As you can see, the upscaling process
is complete. When I open both images in the browser, the upscaled one is now larger
than the original. However, this is just the basics. We're now moving on to the upscalers
that truly enhance the quality of the image. Next, we're taking our image upscaling
to the next level using models. For this, I'll set up an "Upscale Image (By Model)" node
and link it with the "Load Upscale Model" node to load our models. We'll then connect
this setup to the "Upscale Image By" node, where we can adjust the scaling factor as
needed. Important to note, these models require separate downloading. I'll guide you
through three ways to find and download them, highlighting the models I frequently
use and consider the best for starters. Firstly, you can download models directly from
the Manager. If you don't have the Manager yet, you can find a link in the description or refer
back to my last video where I downloaded it automatically with Pinokio. In the Manager, you'll
click on "Install Models" and find them under the "Upscale" category. You can download them directly
and they'll be available after restarting ComfyUI. Secondly, CivitAI offers a selection of
upscalers. Just filter for "Upscaler" under "Models" to find them. You can then compare
them here and choose the best ones for you, as they are visualized for you here,
as opposed to the manager. I'll show you where to save them in a moment, but let's
talk about the third way to find upscalers. Finally, there's OpenModelDB, primarily for
upscaler models. As with CivitAI, you can review and compare to find the best fit for your needs.
Now I'll show you the folder where these should be downloaded. Here's the path; mine might look
a bit different if ComfyUI was installed without Pinokio, but it's almost the same. Just navigate
to your ComfyUI folder and follow this path. Once you have your upscaler ready, you can load
it and upscale your image with it. Upon reviewing the results, you'll notice an improvement in
resolution. Later, I'll showcase all images with their respective upscalers for a clearer
comparison and to highlight the differences. Next we have the LDSR upscaler, a
relatively new node for ComfyUI. You can find and download it from
the Manager, or alternatively, download my workflow at the end and add it
using the "Missing Nodes" function through the Manager – the choice is yours. I'm
leaving the settings as they are; you're free to experiment and see what works best, but
for me, the default settings are perfectly fine. However, one thing you might want to try
is increasing the steps to make the image more detailed and clearer. Keep in mind,
though, more steps don't automatically mean better results for every image; it always
depends on the image you're upscaling. After upscaling, you can immediately see the
original image has become clearer. As mentioned, I'll provide a more detailed comparison towards
the end, showing other upscaled images so you can better appreciate the differences. One
more thing about the upscaler, it does not work well with all images and it requires a lot of
resources, which is why upscaling can take longer. Now, let's move on to the next upscaler,
"Iterative Upscale," which also requires a separate download. Personally, I'm not a huge
fan of this upscaler, but I've noticed better results on some images with this one. It's
best to test it on your images and form your own opinion. To use this upscaler, we need the
"Iterative Upscale Image" node. There's another way to use this upscaler, which I'll cover
in another video, but to keep things simple, I'll just continue with this method. This node
is then connected to "Pixel KSampler Upscale." Remember to link the "Pixel KSampler Upscale"
with the other necessary nodes as well. It is important that we also connect the Iterative
Upscale Image node to the VAE of our model. Next, we'll transfer the image settings to
"Pixel KSampler Upscale" to minimize changes to the original image. If you desire alterations,
feel free to adjust the settings. We set the steps to 50, higher than the original, to enhance
detail and clarity, and reduce noise to 0.1, aiming to keep changes to the original as minimal
as possible. Increasing this value will cause the upscaler to alter your original image more
significantly. For the iterative upscaler, I recommend setting the steps to 2 or 3. I keep
mine at 3, as it seems to work best for me. After upscaling, the image looks like
this – not perfect, but I'll show you later that this upscaler works better
on some images and worse on others. Now, let's move on to the best upscaler, in my opinion. I exclusively use this one for
my images because it achieves the best results. To do this, we'll create the Ultimate SD
Upscale node and link it to the necessary nodes. As I mentioned at the beginning of
this video, I'll provide my workflow for download at the end, so you can check out all
the settings and connections at your leisure. Then I'll take the settings from my
KSampler, set the Denoise to 0.1 and the Steps to 50 and we're done. We set up
this upscaler similar as the previous one. Because it works in a similar white way.
I leave the other settings as they are. I will now start the upscale process
only after I have stopped the recording, because this upscaler needs a lot of power and
I simply cannot record while it is running. As you can see, I got a
really good result with it. Now, I'm going to display the results side by side
with the different upscalers for comparison. Feel free to pause the video and take a closer look if
you'd like. As mentioned earlier, some upscalers perform better on certain images than others;
it really depends on the image. For example, I find that the Ultimate SD Upscaler works best
on people, while the Iterative Upscale and LDSR Upscaler excel with landscapes. But, of course,
I encourage you to test them out for yourself. Now let's take a closer look at my
upscaling workflow. On the left, you can see that I've separated the
image generation from the rest of the process. This is because upscaling consumes
a lot of resources. To optimize your time, I suggest generating 10 to 20 images first and
then selecting the best ones for upscaling. You can simply paste these images at the top
and start the upscaling process. If there are any nodes you prefer not to use, simply
bypass them to streamline your workflow. For those interested in learning more and using
my method of image generation and upscaling, I've shared both my workflow and the
ComfyUI document on Discord and Patreon. I hope you enjoyed the video. If you
did, please leave a like and don't forget to check out my latest ComfyUI
tutorial. See you in the next video!
