SeedVR2 is an upscaler model by ByteDance-Seed for Video Restoration and is licensed under Apache-2.0 license. This implies that can be used in commerical capacity if you want.
The one thing to note is that this model is GPU VRAM hungry model and you will struggle to run it on a Consumer Grade GPU. I have 16GB VRAM on my RTX4080 and it has missed results with the smallest model in FP8 (floating point 8) version. So I recommend that you start with atleast 24gb but best success with 32GB and above. I rented the RTX5090 with 32GB VRAM on Runpod platform for less than $0.90 per hour which is incredibly cheap than buying a $5000 card if it is stock at your nearby store.
The quality of the upscaler is very high, and can accentuate the details of the orginal video incredibly well, with lots of sharpnes and contrast. Sometimes it may feel like its too much.
They have a quickstart guide that you can use and install on your computer to run it using scripts.
SeedVR2 in ComfyUI
If you are a ComfyUI user than you can installed the custom nodes by numz called ComfyUI-SeedVR2_VideoUpscaler using the ComfyUI Manager. If you are going to run this on RunPod then check out our Custom Notebook that helps you get started with ComfyUI easily.
The custom nodes will automatically download the models based on your selection. There two main types of models:
3B – 3 billion version
7B – 7 billion version
Obviously the higher parameter version is going to produce the best quality but it is also the highest VRAM hungry. Luckily numz the maker of these nodes has provided a Block swap node which means you can control how the VRAM and Blocks are loaded.
A great in depth video explaining the models and how they work is provided by AInVFX channel. Check it out if you want to learn the tech side of this model.
During my experimentation I created 3 workflows with SeedVR2. You can download my workflows below which contains all three. I am utilising some additional nodes in my workflows so if you get some red boxes and notification that nodes are missing, use ComfyUI Manager to install Missing Custom Nodes.
SeedVR2 Video Upscale (14 downloads )
The advanced workflow is taking in the incoming video and comparing it to see if the video is larger than 0.5 Megapixel in resolution. If thats the case it means the vidoe is too large in size and will likely cause OOM (out of memory) error, hence the workflow resized it down to the manageable resolution based on Upscale factor.
The workflow then will not upscale the full upscale factor but keep to 80% of desired size. Then uses standard image upscaler to take the image to the 100% desired size. This way again you manage the VRAM usage and also ensure the video does not appear to be oversharpened. Which is one of the big observations of mine that this model oversharpens the videos.
In order make sure there is temporal consistency between the frames, the Batch size is crucial. By default its set to 5 which will introduce flicker in the output video, try to increase this to a point when you don’t get OOM errors. I have tested up to 121 frames and the RTX4090 has been able to handle it well.
I share my complete process and explanation in my YT video which you can watch below. I will share another sample video where you can see the before and after comparison of SeedVR2 Videos upscaled by 2x.
Try out my workflows but as stated before they may not work well in local GPU environment. Best could be to use Runpod or like similar cloud GPU environment.
If you’d like to support our site please consider buying us a Ko-fi, grab a product or subscribe. Need a faster GPU, get access to fastest GPUs for less than $1 per hour with RunPod.io