Seed VR 2 is an advanced model via Bybustence seed Video restoration and licensed under the Apache -2.0 license. This shows that if you want to be used in a lower capacity.
One thing to note is that this model is a GPU Waram Hungary model and you will struggle to run it on the user grade GPU. I have 16GB VRAM on my RTX4080 and has lost the results with the smallest model in the FP8 (Floating Point 8) version. So I recommend that you start at least 24 GB but the best success with 32GB and above. I rented RTX5090 with 32GB VRAM On Ranpid platform For less than $ 0.90 per hour, which is stock on your nearest store is incredibly cheaper than buying $ 5000 cards.
The quality of the upsclar is very high, and many sharpens and vice versa, and can be incredibly well -accelerated. Sometimes it can feel a lot.
They have a Quick Start Directions that you can use and install to run using the script on your computer.
Cofuyi in Seed VR 2
If you are more comfyui user you can install custom nodes from comfyui-seedvr2_videoScaler Using Comphyui Manager. If you are going to run it on Rinpad, check out our Custom notebook This helps you easily start with a comphi.
Customs nodes will automatically download models based on your choice. Two main types of model:
- 3B – 3 Billion version
- 7b – 7 billion versions
Obviously the high parameter version is the best quality producing the quality but it is most VRAM hunger. Fortunately, the maker of these nodes has provided a block sweep node, which means you can control how the VRAM and the blocks are filled.
A huge video of models to describe and their working method is provided by the AinVFX channel. If you want to learn the tech aspect of this model, check it out.
During my experience I made 3 workflows with Seed VR2. You can download my workflows that contain three. I am using some additional nodes in my workflose so if you get some red boxes and notices that the nodes are missing, use Comphyui Manager To install Customs nodes deprived.
Seed VR 2 Video High (2 Download)
The latest workflow is taking in the upcoming video and comparing it to see if the video resolution is larger than 0.5 megapixel. If in this case it means that the size of the Vidu is huge and will potentially cause OOM (out of memory), the workflow has turned it into a resolution based on an advanced factor.
After that the workflow will not promote the full advanced element, but will remain up to 80 % of the desired size. The image then uses a standard image upsclar to carry the image to 100 % desired size. In this way, once again you manage the use of VRAM and also ensure that the video is not monitored. Which is one of my major observations that this model oversees videos.
In order, make sure that there is a temporary consistency between the frames, the size of the batch is very important. The default is on its set 5, which will introduce Flickr in the output video, try to increase it when you do not get OOM errors. I have experienced 121 frames and RTX4090 has been able to handle it well.
I share my full process and explanation in my YT video that you can see below. I will share another sample video where you can watch before and after comparing the Seed VR 2 videos, which has increased by 2X.
Try my workflows but as previously stated that they cannot work well in the local GPU environment. Can have the best use Run pod Or like a cloud GPU environment.
If you want to support our site please consider buying us FiFor, for, for,. Catch a product Or Subscribe. Fast GPU is required, access the fastest GPUS for less than $ 1 per hour runpod.io