SeedVR2 for ComfyUI: Video Upscaling With a Side of Memory Engineering

What it does This project wraps ByteDance’s SeedVR2 diffusion-based upscaler into ComfyUI nodes for high-quality video and image restoration. It also exposes the same engine as a standalone multi-GPU CLI that can stream long videos in memory-bounded chunks instead of loading entire sequences at once.

The interesting bit The real work here is platform gymnastics. The changelog reads like a driver team’s backlog: Apple Silicon unified-memory optimizations, Flash Attention 3 support for Blackwell/RTX 50xx, automatic fallback chains when CUDA extensions are corrupted, and even a security patch to stop malicious .pth files from executing code during model load.

Key highlights

• Runs as both ComfyUI custom nodes and a standalone multi-GPU CLI with frame-level parallelism and temporal overlap blending
• Streams arbitrarily long videos via chunking with independent model caching between segments
• Supports multiple attention backends—Flash Attention 2/3, SageAttention 2/3, CUDNN—with automatic fallback to PyTorch SDPA
• Ships cross-platform fixes for Apple Silicon MPS, AMD ROCm, older NVIDIA cards, and PyTorch 2.9+ memory regressions
• Optional 10-bit FFmpeg output backend to reduce gradient banding compared with the default 8-bit OpenCV writer

Caveats

• The release history shows recurring regressions in color accuracy and memory handling—reverted bfloat16 detection, MPS cache leaks, and LAB color artifacts—so mileage varies by version and hardware
• BlockSwap auto-disables on Apple Silicon with a warning; the README notes unified memory “makes it meaningless” there
• VRAM hard-limit enforcement was rolled back; the tool now warns on overflow rather than preventing shared-memory paging slowdowns

Verdict Grab it if you are already in the ComfyUI ecosystem and need serious video upscaling without writing your own chunking and multi-GPU orchestration. If you want a simple one-click GUI or are tight on VRAM, this wrapper won’t shrink the model’s appetite for GPU memory.