[0:00] In today's video, I'm going to show you
[0:02] a completely free way to use AI for
[0:04] upscaling low-resolution videos, old
[0:07] clips, and movies, turning them into
[0:10] 1080p or even 4K videos with sharp
[0:13] details and high-quality results. Now,
[0:15] it's worth mentioning that there are
[0:17] already plenty of other video upscaling
[0:19] tools [music] out there. However, most
[0:21] of those tools come at a high price or
[0:23] require subscription. So, instead, we're
[0:26] going to use a completely free upscaling
[0:28] model called SeedVR 2.5. I'm going to
[0:31] show you exactly how to use it, test its
[0:33] output quality and upscaling speed, and
[0:35] see how it compares to other paid tools
[0:37] [music] in the market. Before getting
[0:39] started, it's worth mentioning that the
[0:41] hardware you use will directly impact
[0:44] upscaling speed and your ability to
[0:46] tweak certain [music] settings. I will
[0:48] put my own specs on screen for your
[0:50] reference. Keep in mind that you can
[0:52] still use SeedVR on GPUs with as little
[0:55] as 8 GB of VRAM, but results will vary,
[0:58] and I will explain this in more detail
[1:00] later in the process. Now, if you do not
[1:03] meet the minimum hardware requirements,
[1:05] or if you want to use SeedVR at its full
[1:08] potential, you can run it entirely
[1:10] online. There are dedicated websites
[1:12] that let you run ComfyUI in the cloud
[1:15] with access to high-end GPUs at
[1:17] different hourly rates, and I will leave
[1:20] direct links to those websites in the
[1:22] description box. With those options, you
[1:24] can skip all the installation steps, and
[1:26] you can still follow the exact same
[1:27] process as running ComfyUI locally. And
[1:30] if you want to run the workflow on your
[1:32] computer, you will first need to install
[1:34] ComfyUI. If you haven't done that yet,
[1:36] make sure you watch this video first. It
[1:38] will take about 5 minutes, then you can
[1:40] come back and continue this tutorial.
[1:42] Now that you have ComfyUI installed, go
[1:44] to the GitHub page that I linked in the
[1:46] description and download the SeedVR
[1:49] video upscaling workflow. Make sure you
[1:51] get the JSON file, then drag and drop it
[1:54] into ComfyUI to open it. As soon as you
[1:57] open the workflow, you will likely see a
[1:59] pop-up window telling you that you're
[2:01] missing some [music] required nodes. To
[2:02] fix this, open the manager, go to all
[2:05] extensions, and search [music] for this
[2:07] node, then click install. It only takes
[2:10] a couple of seconds, and once it's done,
[2:12] click apply changes. Wait until it says
[2:15] extensions successfully installed, then
[2:18] close the manager, and the workflow
[2:20] should now be ready to go. But before we
[2:22] go through the settings, I want to share
[2:24] something with you. If you are someone
[2:25] who creates AI content, [music]
[2:27] you probably already know how scattered
[2:29] everything can get. Different websites,
[2:31] new model drops, it's all over the
[2:33] internet, and you're just sitting there
[2:34] trying to figure out where to even
[2:36] [music] use it. And that's exactly why I
[2:38] want to tell you about today's sponsor,
[2:40] openart.ai.
[2:42] It brings everything into one place. You
[2:44] get access to powerful video models like
[2:46] One, Kling, LTX, and everything runs
[2:50] online, so your GPU can finally relax.
[2:52] I've been using it a lot over the past
[2:54] couple of years, especially for
[2:56] generating images and [music] thumbnail
[2:58] ideas using models like Flocks or Nano
[3:00] Banana Pro. And using OpenArt, you don't
[3:03] burn through a ton of credits, which is
[3:04] huge when you're just trying things out.
[3:07] You can also use OpenArt to replace
[3:09] characters in videos, generate
[3:11] voiceovers, and creating accurate lip
[3:13] sync. Wow. It's honestly become a solid
[3:16] part of my workflow, and [music] I
[3:17] definitely recommend trying it out. I
[3:20] will leave a link in the description,
[3:21] and make sure you use my code [music]
[3:23] MDMZ20 to get 20% off when you choose
[3:26] any monthly subscription. And now, let's
[3:28] get back to the tutorial. First, let's
[3:31] go to the load video node and click here
[3:34] to import the low-quality video you want
[3:36] to upscale. For this demo, I'm going to
[3:38] use [music] this clip of a parrot. I
[3:40] specifically chose this to see how well
[3:43] SeedVR recovers feather details and
[3:45] whether it keeps the blurred background
[3:47] [music] soft or tries to sharpen it. The
[3:49] resolution of this video is 360p, which
[3:52] is pretty low and pixelated, as you can
[3:55] see, but you can still use the same
[3:57] workflow to upscale even
[3:59] lower-resolution videos. Here, you can
[4:01] choose from a list of model variations.
[4:04] They are all SeedVR2 models, but they
[4:06] come in different sizes, they have
[4:08] different hardware requirements, and
[4:10] they are suited to slightly different
[4:11] use cases. So, yes, your choice here
[4:14] matters, [music] and it will affect
[4:15] generation speed and even quality. But
[4:18] to simplify things, the 7B models
[4:21] generally produce better quality, but
[4:23] they also require more VRAM. I recommend
[4:25] using them only if you have more than 24
[4:28] GB of VRAM. The 3B models, however,
[4:32] should work well on lower VRAM GPUs such
[4:35] as 8 GB or 12 GB. Quality might be
[4:38] affected. In some cases, you will notice
[4:40] a big difference. Other times, the
[4:42] results will be very close to what you
[4:44] get with 7B models. Still, use the 3B
[4:47] models if you have low VRAM or if you
[4:50] want to upscale long videos. You will
[4:52] also notice that some models are labeled
[4:54] sharp. I do not recommend using these,
[4:56] since they tend to produce less
[4:58] natural-looking results. That said, you
[5:01] can use them if you want extra crisp
[5:03] edges and more detail, and they can be a
[5:05] good fit for cartoons or animated
[5:07] videos. Choosing the right model depends
[5:09] on many factors, so I recommend
[5:11] experimenting based on your input and
[5:13] computer specs. If a model is too heavy
[5:15] for your machine, you will get an
[5:17] out-of-memory error when running the
[5:19] workflow. That said, if you want the
[5:21] best results, go with the 7B FP16
[5:24] version. It is the most demanding model
[5:26] of the bunch, but if your machine can
[5:28] handle it, it's the one I recommend.
[5:30] Personally, I found that it hits its
[5:32] limit on my machine when videos exceed
[5:35] 10 seconds. So, for longer videos, I
[5:37] switched to the FP8 version. Now, let's
[5:40] move over to this node. [music] Here,
[5:42] you can choose the target resolution you
[5:44] want to upscale your video to. By
[5:46] default, it is set to 1080p, and I
[5:49] recommend sticking with this resolution
[5:51] for a good balance of quality and
[5:53] upscaling speed. Personally, I did not
[5:55] have any issues upscaling to 1080p using
[5:58] the 7B models. I did try pushing it to
[6:01] 4K, and unfortunately, that did not work
[6:04] on my machine, but it worked just fine
[6:07] with the same video on RunComfy using a
[6:10] GPU with 48 GB of VRAM. There are still
[6:13] a bunch of other settings here that
[6:15] directly affect output quality and speed
[6:17] that you might want to adjust based on
[6:19] your input and needs. For example,
[6:21] [music] the batch size should ideally
[6:24] match the total number of frames in your
[6:26] video. Not only will you achieve better
[6:28] quality, but you will also get better
[6:31] temporal consistency. With a small batch
[6:33] size, you might notice some flickering
[6:36] in the upscaled video. However, a bigger
[6:38] batch size requires more VRAM, so there
[6:41] is a limit to how high you can push it.
[6:43] This is where understanding the settings
[6:45] and experimenting comes in. You can read
[6:47] more about what each of these inputs
[6:49] does by simply hovering over it. For
[6:52] simple upscaling tasks, you do not need
[6:54] to worry too much about the other
[6:56] settings. But let me know if you would
[6:58] like me to make another video where I
[6:59] dive deeper into the settings, and I
[7:01] will also link some useful resources in
[7:04] the description box. Okay, at this
[7:06] stage, we can go ahead and click run to
[7:08] start upscaling the video. Keep in mind
[7:10] that if you're running this for the
[7:12] first time, the workflow will use some
[7:14] time to download the models before it
[7:16] starts processing. [music] And real
[7:18] quick, if you're finding this tutorial
[7:19] useful so far, please support me by
[7:21] liking the video and subscribing to the
[7:23] channel. I make a lot of ComfyUI and AI
[7:26] tutorials that you don't want to miss.
[7:28] Once processing is done, you can preview
[7:30] the upscaled video in this node, and as
[7:32] you can see, SeedVR did an impressive
[7:35] job enhancing our low-quality video and
[7:38] restoring detail. I also like that it
[7:40] retained the softness of the blurred
[7:42] background. Some upscaling models overdo
[7:44] it and try to sharpen out-of-focus
[7:47] background, which would look terrible.
[7:49] You might argue that it looks a little
[7:51] over-sharpened and unnatural, especially
[7:53] around the parrot's eye, as you can see,
[7:56] but let's not forget that the input
[7:57] quality was very low, which makes the
[8:00] upscaling process more challenging than
[8:02] working with a medium-quality video.
[8:04] That being said, I'm confident this is
[8:07] one of the best, if not the best,
[8:09] open-source [music]
[8:10] upscaling model out there. During my
[8:12] tests, it performed really well with
[8:15] human faces. It has preserved the
[8:17] original colors and contrast. I also
[8:19] found it very useful for enhancing
[8:22] low-light footage like this, as well as
[8:24] black and white videos, which is truly
[8:26] impressive. [music]
[8:27] And by the way, if you want to learn how
[8:28] to colorize black and white videos using
[8:31] AI, check this video out. Another great
[8:33] use case of SeedVR is AI-generated
[8:36] videos. You can generate videos at a
[8:38] lower resolution to save time and money,
[8:41] then run them through SeedVR to upscale
[8:43] them for free. Now, I also tried to see
[8:46] how SeedVR performs against the video
[8:49] upscaling giant, Topaz [music] Video.
[8:52] This has always been my favorite
[8:53] upscaling tool, and I use it all the
[8:56] time. It's fast, works really well on
[8:58] all sorts of footage, and gives you
[9:00] access to a bunch of upscaling models,
[9:02] both local and on the cloud. It also has
[9:05] frame interpolation, advanced exporting
[9:07] settings, and much more. Personally, I
[9:10] find Topaz much more intuitive. The UI
[9:13] makes the process easy and simple. You
[9:15] can preview short segments before
[9:17] upscaling, and with the right
[9:18] combination of settings, I'm able to get
[9:20] more natural-looking results. However,
[9:23] when you put the price into perspective,
[9:25] things change. Topaz Video requires a
[9:28] subscription, and prices start at $25 a
[9:31] month. For a tool that's absolutely
[9:34] free, SeedVR stands really strong.
[9:36] Personally, Topaz Video has become a
[9:38] crucial part of my workflow, and I will
[9:40] keep using it, but I also like having
[9:42] the option of using SeedVR inside
[9:45] ComfyUI. And for a free tool, I find it
[9:47] very capable. [music] Let me know what
[9:48] you think in the comments below. Is
[9:50] SeedVR good enough for what you need? I
[9:53] would love to hear your thoughts. Stay
[9:55] creative, and I'll catch you in the next
[9:57] video. Peace.
[9:59] >> [music]