SeedVR 2.5 ComfyUI Tutorial | FREE AI Video Upscaler to 4K

[00:00] In today's video, I'm going to show you

[00:02] a completely free way to use AI for

[00:04] upscaling low-resolution videos, old

[00:07] clips, and movies, turning them into

[00:10] 1080p or even 4K videos with sharp

[00:13] details and high-quality results. Now,

[00:15] it's worth mentioning that there are

[00:17] already plenty of other video upscaling

[00:19] tools [music] out there. However, most

[00:21] of those tools come at a high price or

[00:23] require subscription. So, instead, we're

[00:26] going to use a completely free upscaling

[00:28] model called SeedVR 2.5. I'm going to

[00:31] show you exactly how to use it, test its

[00:33] output quality and upscaling speed, and

[00:35] see how it compares to other paid tools

[00:37] [music] in the market. Before getting

[00:39] started, it's worth mentioning that the

[00:41] hardware you use will directly impact

[00:44] upscaling speed and your ability to

[00:46] tweak certain [music] settings. I will

[00:48] put my own specs on screen for your

[00:50] reference. Keep in mind that you can

[00:52] still use SeedVR on GPUs with as little

[00:55] as 8 GB of VRAM, but results will vary,

[00:58] and I will explain this in more detail

[01:00] later in the process. Now, if you do not

[01:03] meet the minimum hardware requirements,

[01:05] or if you want to use SeedVR at its full

[01:08] potential, you can run it entirely

[01:10] online. There are dedicated websites

[01:12] that let you run ComfyUI in the cloud

[01:15] with access to high-end GPUs at

[01:17] different hourly rates, and I will leave

[01:20] direct links to those websites in the

[01:22] description box. With those options, you

[01:24] can skip all the installation steps, and

[01:26] you can still follow the exact same

[01:27] process as running ComfyUI locally. And

[01:30] if you want to run the workflow on your

[01:32] computer, you will first need to install

[01:34] ComfyUI. If you haven't done that yet,

[01:36] make sure you watch this video first. It

[01:38] will take about 5 minutes, then you can

[01:40] come back and continue this tutorial.

[01:42] Now that you have ComfyUI installed, go

[01:44] to the GitHub page that I linked in the

[01:46] description and download the SeedVR

[01:49] video upscaling workflow. Make sure you

[01:51] get the JSON file, then drag and drop it

[01:54] into ComfyUI to open it. As soon as you

[01:57] open the workflow, you will likely see a

[01:59] pop-up window telling you that you're

[02:01] missing some [music] required nodes. To

[02:02] fix this, open the manager, go to all

[02:05] extensions, and search [music] for this

[02:07] node, then click install. It only takes

[02:10] a couple of seconds, and once it's done,

[02:12] click apply changes. Wait until it says

[02:15] extensions successfully installed, then

[02:18] close the manager, and the workflow

[02:20] should now be ready to go. But before we

[02:22] go through the settings, I want to share

[02:24] something with you. If you are someone

[02:25] who creates AI content, [music]

[02:27] you probably already know how scattered

[02:29] everything can get. Different websites,

[02:31] new model drops, it's all over the

[02:33] internet, and you're just sitting there

[02:34] trying to figure out where to even

[02:36] [music] use it. And that's exactly why I

[02:38] want to tell you about today's sponsor,

[02:40] openart.ai.

[02:42] It brings everything into one place. You

[02:44] get access to powerful video models like

[02:46] One, Kling, LTX, and everything runs

[02:50] online, so your GPU can finally relax.

[02:52] I've been using it a lot over the past

[02:54] couple of years, especially for

[02:56] generating images and [music] thumbnail

[02:58] ideas using models like Flocks or Nano

[03:00] Banana Pro. And using OpenArt, you don't

[03:03] burn through a ton of credits, which is

[03:04] huge when you're just trying things out.

[03:07] You can also use OpenArt to replace

[03:09] characters in videos, generate

[03:11] voiceovers, and creating accurate lip

[03:13] sync. Wow. It's honestly become a solid

[03:16] part of my workflow, and [music] I

[03:17] definitely recommend trying it out. I

[03:20] will leave a link in the description,

[03:21] and make sure you use my code [music]

[03:23] MDMZ20 to get 20% off when you choose

[03:26] any monthly subscription. And now, let's

[03:28] get back to the tutorial. First, let's

[03:31] go to the load video node and click here

[03:34] to import the low-quality video you want

[03:36] to upscale. For this demo, I'm going to

[03:38] use [music] this clip of a parrot. I

[03:40] specifically chose this to see how well

[03:43] SeedVR recovers feather details and

[03:45] whether it keeps the blurred background

[03:47] [music] soft or tries to sharpen it. The

[03:49] resolution of this video is 360p, which

[03:52] is pretty low and pixelated, as you can

[03:55] see, but you can still use the same

[03:57] workflow to upscale even

[03:59] lower-resolution videos. Here, you can

[04:01] choose from a list of model variations.

[04:04] They are all SeedVR2 models, but they

[04:06] come in different sizes, they have

[04:08] different hardware requirements, and

[04:10] they are suited to slightly different

[04:11] use cases. So, yes, your choice here

[04:14] matters, [music] and it will affect

[04:15] generation speed and even quality. But

[04:18] to simplify things, the 7B models

[04:21] generally produce better quality, but

[04:23] they also require more VRAM. I recommend

[04:25] using them only if you have more than 24

[04:28] GB of VRAM. The 3B models, however,

[04:32] should work well on lower VRAM GPUs such

[04:35] as 8 GB or 12 GB. Quality might be

[04:38] affected. In some cases, you will notice

[04:40] a big difference. Other times, the

[04:42] results will be very close to what you

[04:44] get with 7B models. Still, use the 3B

[04:47] models if you have low VRAM or if you

[04:50] want to upscale long videos. You will

[04:52] also notice that some models are labeled

[04:54] sharp. I do not recommend using these,

[04:56] since they tend to produce less

[04:58] natural-looking results. That said, you

[05:01] can use them if you want extra crisp

[05:03] edges and more detail, and they can be a

[05:05] good fit for cartoons or animated

[05:07] videos. Choosing the right model depends

[05:09] on many factors, so I recommend

[05:11] experimenting based on your input and

[05:13] computer specs. If a model is too heavy

[05:15] for your machine, you will get an

[05:17] out-of-memory error when running the

[05:19] workflow. That said, if you want the

[05:21] best results, go with the 7B FP16

[05:24] version. It is the most demanding model

[05:26] of the bunch, but if your machine can

[05:28] handle it, it's the one I recommend.

[05:30] Personally, I found that it hits its

[05:32] limit on my machine when videos exceed

[05:35] 10 seconds. So, for longer videos, I

[05:37] switched to the FP8 version. Now, let's

[05:40] move over to this node. [music] Here,

[05:42] you can choose the target resolution you

[05:44] want to upscale your video to. By

[05:46] default, it is set to 1080p, and I

[05:49] recommend sticking with this resolution

[05:51] for a good balance of quality and

[05:53] upscaling speed. Personally, I did not

[05:55] have any issues upscaling to 1080p using

[05:58] the 7B models. I did try pushing it to

[06:01] 4K, and unfortunately, that did not work

[06:04] on my machine, but it worked just fine

[06:07] with the same video on RunComfy using a

[06:10] GPU with 48 GB of VRAM. There are still

[06:13] a bunch of other settings here that

[06:15] directly affect output quality and speed

[06:17] that you might want to adjust based on

[06:19] your input and needs. For example,

[06:21] [music] the batch size should ideally

[06:24] match the total number of frames in your

[06:26] video. Not only will you achieve better

[06:28] quality, but you will also get better

[06:31] temporal consistency. With a small batch

[06:33] size, you might notice some flickering

[06:36] in the upscaled video. However, a bigger

[06:38] batch size requires more VRAM, so there

[06:41] is a limit to how high you can push it.

[06:43] This is where understanding the settings

[06:45] and experimenting comes in. You can read

[06:47] more about what each of these inputs

[06:49] does by simply hovering over it. For

[06:52] simple upscaling tasks, you do not need

[06:54] to worry too much about the other

[06:56] settings. But let me know if you would

[06:58] like me to make another video where I

[06:59] dive deeper into the settings, and I

[07:01] will also link some useful resources in

[07:04] the description box. Okay, at this

[07:06] stage, we can go ahead and click run to

[07:08] start upscaling the video. Keep in mind

[07:10] that if you're running this for the

[07:12] first time, the workflow will use some

[07:14] time to download the models before it

[07:16] starts processing. [music] And real

[07:18] quick, if you're finding this tutorial

[07:19] useful so far, please support me by

[07:21] liking the video and subscribing to the

[07:23] channel. I make a lot of ComfyUI and AI

[07:26] tutorials that you don't want to miss.

[07:28] Once processing is done, you can preview

[07:30] the upscaled video in this node, and as

[07:32] you can see, SeedVR did an impressive

[07:35] job enhancing our low-quality video and

[07:38] restoring detail. I also like that it

[07:40] retained the softness of the blurred

[07:42] background. Some upscaling models overdo

[07:44] it and try to sharpen out-of-focus

[07:47] background, which would look terrible.

[07:49] You might argue that it looks a little

[07:51] over-sharpened and unnatural, especially

[07:53] around the parrot's eye, as you can see,

[07:56] but let's not forget that the input

[07:57] quality was very low, which makes the

[08:00] upscaling process more challenging than

[08:02] working with a medium-quality video.

[08:04] That being said, I'm confident this is

[08:07] one of the best, if not the best,

[08:09] open-source [music]

[08:10] upscaling model out there. During my

[08:12] tests, it performed really well with

[08:15] human faces. It has preserved the

[08:17] original colors and contrast. I also

[08:19] found it very useful for enhancing

[08:22] low-light footage like this, as well as

[08:24] black and white videos, which is truly

[08:26] impressive. [music]

[08:27] And by the way, if you want to learn how

[08:28] to colorize black and white videos using

[08:31] AI, check this video out. Another great

[08:33] use case of SeedVR is AI-generated

[08:36] videos. You can generate videos at a

[08:38] lower resolution to save time and money,

[08:41] then run them through SeedVR to upscale

[08:43] them for free. Now, I also tried to see

[08:46] how SeedVR performs against the video

[08:49] upscaling giant, Topaz [music] Video.

[08:52] This has always been my favorite

[08:53] upscaling tool, and I use it all the

[08:56] time. It's fast, works really well on

[08:58] all sorts of footage, and gives you

[09:00] access to a bunch of upscaling models,

[09:02] both local and on the cloud. It also has

[09:05] frame interpolation, advanced exporting

[09:07] settings, and much more. Personally, I

[09:10] find Topaz much more intuitive. The UI

[09:13] makes the process easy and simple. You

[09:15] can preview short segments before

[09:17] upscaling, and with the right

[09:18] combination of settings, I'm able to get

[09:20] more natural-looking results. However,

[09:23] when you put the price into perspective,

[09:25] things change. Topaz Video requires a

[09:28] subscription, and prices start at $25 a

[09:31] month. For a tool that's absolutely

[09:34] free, SeedVR stands really strong.

[09:36] Personally, Topaz Video has become a

[09:38] crucial part of my workflow, and I will

[09:40] keep using it, but I also like having

[09:42] the option of using SeedVR inside

[09:45] ComfyUI. And for a free tool, I find it

[09:47] very capable. [music] Let me know what

[09:48] you think in the comments below. Is

[09:50] SeedVR good enough for what you need? I

[09:53] would love to hear your thoughts. Stay

[09:55] creative, and I'll catch you in the next

[09:57] video. Peace.

[09:59] >> [music]