SeedVR2 for ComfyUI — Easy Video Upscaling With GGUF & BlockSwap
Upscale videos locally with SOTA SeedVR2 upscaling models.
SeedVR2 is ByteDance’s next-gen AI video restoration model designed for high-resolution upscaling with cutting-edge temporal consistency and one-step inference. Unlike older frame-by-frame tools, SeedVR2’s diffusion-based approach was specifically trained on video sequences, so it preserves motion and details from frame to frame for professional-looking results. It’s fully compatible with ComfyUI and runs efficiently on consumer GPUs with both GGUF quantized models and traditional diffusion models. This post includes everything you need for a seamless install: preloaded models, custom nodes, and automated workflows ready to upscale your AI or standard footage in minutes.
Preloaded Models within the Installer (Low VRAM)
The portable installer comes with SeedVR2 GGUF models optimized for low VRAM use. Standard diffusion and SeedVR2 GGUF weights automatically download after your first run—no manual setup required. Just click to install and start upscaling.
The workflow currently supports the 3B and 7B fp16, fp8, and GGUF models.
Users with less than 12GB VRAM should opt for GGUF models, which are specifically optimized for memory efficiency while retaining quality.
Speed
Upscale clips to 480p videos in roughly 3-10 minutes on an RTX 4050 6GB VRAM GPU using GGUF models. The workflow dynamically adapts to your system, which also depends on the batch size and model you select so faster render times and higher resolutions (720p/24fps or even 4K) are achievable with higher-end GPUs (e.g., RTX 4090).
System Requirements
Nvidia RTX 30XX, 40XX, or 50XX series GPU (FP16 support required; GTX 10XX/20XX not tested)
CUDA-compatible GPU (minimum 6GB VRAM; more recommended)
Windows OS
At least 40GB free storage.
What’s Included
Portable, pre-configured ComfyUI and WAN2.2 install—no manual ComfyUI setup.
Automated download of all custom nodes and required models.
Complete SeedVR2 Upscaling workflows supporting both diffusion and GGUF models.
Usage Notes
Compatible with both GGUF quantized models and standard diffusion models.
If your workflow runs out of memory during generation, try reducing the batch size. Lower batch sizes use less VRAM but can introduce flickering in the final output. For best image quality, keep the batch size as high as your GPU can handle.
The workflow also comes equipped with blockswapping, a feature allowing memory savings and offloading of unused model blocks. The more VRAM you have available, the lower you can set the number of blocks for speed and quality gains. If you run out of memory, increase the block swap value (up to a max of 36) — this allows SeedVR2 to swap portions of the model in and out of memory, making big models feasible even on midrange VRAM.