numz / SeedVR2_comfyUI

ComfyUI-SeedVR2_VideoUpscaler

Official release of SeedVR2 for ComfyUI that enables Upscale Video/Images generation.

🆙 Todo

• Fixed unloading the 3B model when the process is finished (sorry about that, I'm trying to find out what's going on)

🚀 Updates

2025.06.24

• 🚀 Speed up the process until x4 (see new benchmark)

2025.06.22

• 💪 FP8 compatibility !
• 🚀 Speed Up all Process
• 🚀 less VRAM consumption (Stay high, batch_size=1 for RTX4090 max, I'm trying to fix that)
• 🛠️ Better benchmark coming soon

2025.06.20

• 🛠️ Initial push

Features

• High-quality Upscaling
• Suitable for any video length once the right settings are found
• Model Will Be Download Automatically from Models

Requirements

• A Huge VRAM capabilities is better, from my test, even the 3B version need a lot of VRAM at least 18GB.
• Last ComfyUI version with python 3.12.9 (may be works with older versions but I haven't test it)

Installation

1. Clone this repository into your ComfyUI custom nodes directory:

cd ComfyUI/custom_nodes
git clone https://github.com/numz/ComfyUI-SeedVR2_VideoUpscaler.git

1. Install the required dependencies:

load venv and :

pip install -r ComfyUI-SeedVR2_VideoUpscaler/requirements.txt

install flash_attn or triton if it ask for it

pip install flash_attn
pip install triton

or from https://github.com/loscrossos/lib_flashattention/releases and https://github.com/woct0rdho/triton-windows

Or use python_embeded :

python_embeded\python.exe -m pip install -r ComfyUI-SeedVR2_VideoUpscaler/requirements.txt

python_embeded\python.exe -m pip install -r flash_attn

1. Models

   Will be automtically download into : models/SEEDVR2

   or can be found here ( MODELS )

Usage

1. In ComfyUI, locate the SeedVR2 Video Upscaler node in the node menu.

1. things to know

temporal consistency : at least a batch_size of 5 is required to activate temporal consistency

1. Configure the node parameters:

   • model : Select your 3B or 7B model
   • seed : a seed but it generate another seed from this one
   • new_width : New desired Width, will keep ration on height
   • cfg_scale :
   • batch_size : VERY IMPORTANT!, this model consume a lot of VRAM, All your VRAM, even for the 3B model, so for GPU under 24GB VRAM keep this value Low, good value is "1" without temporal consistency
   • preserve_vram : for VRAM < 24GB, If true, It will unload unused models during process, longer but works, otherwise probably OOM with

Performance

NVIDIA H100 93GB VRAM (values in parentheses are from the previous benchmark):

NVIDIA RTX4090 24GB VRAM (preserved_vram=off)

Limitations

• Use a lot of VRAM, it will take all!!
• Processing speed depends on GPU capabilities

Credits

• Original SeedVR2 implementation

📜 License

• The code in this repository is released under the MIT license as found in the LICENSE file .