Welcome to the ComfyUI-SeedVR2 Video Upscaler repository! This project offers a non-official video upscaling tool designed specifically for ComfyUI. With this tool, you can enhance your video quality, making your visual content more engaging and clearer.

ComfyUI-SeedVR2_VideoUpscaler

Configure block swapping to reduce VRAM usage during video upscaling.

BlockSwap dynamically moves transformer blocks between GPU and CPU/RAM during inference, enabling large models to run on limited VRAM systems with minimal performance impact.

Configuration Guidelines:
    - blocks_to_swap=0: Disabled (fastest, highest VRAM usage)
    - blocks_to_swap=16: Balanced mode (moderate speed/VRAM trade-off)
    - blocks_to_swap=32-36: Maximum savings (slowest, lowest VRAM)

Advanced Options:
    - use_non_blocking: Enables asynchronous GPU transfers for better performance
    - offload_io_components: Moves embeddings and I/O layers to CPU for additional VRAM savings (slower)
    - cache_model: Keeps model in RAM between runs (avoids model loading time on subsequent generations)
    - enable_debug: Shows detailed memory usage and timing information

Performance Tips:
    - Start with blocks_to_swap=16 and increase until you no longer get OOM errors or decrease if you have spare VRAM
    - Enable offload_io_components if you still need additional VRAM savings
    - Note: Even if inference succeeds, you may still OOM during VAE decoding - combine BlockSwap with VAE tiling if needed (feature in development)
    - Enable cache_model for batch processing to skip model reloading between runs
    - Keep non_blocking=True for better performance (default)
    - Combine with smaller batch_size for maximum VRAM savings

The actual memory savings depend on your specific model architecture and will be shown in the debug output when enabled.
    

ComfyUI Node: SeedVR2 BlockSwap Config

Category

Inputs

Outputs

Extension: ComfyUI-SeedVR2_VideoUpscaler

Run ComfyUI workflows in the Cloud!