comfyui-seedvr2_videoupscaler

ComfyUI-SeedVR2_VideoUpscaler is the official ComfyUI packaging of ByteDance's SeedVR2 model for diffusion-based video and image upscaling. It aims to preserve temporal coherence across frames while enhancing detail, and supports RGB and RGBA content, including transparent images. In ComfyUI, the project uses a modular node setup with separate nodes for the DiT model, VAE model, torch.compile settings, and the main upscaler. It follows the ComfyUI V3 stateless node design and can share loaded models across multiple upscaler instances through a global cache. Models can be downloaded automatically from Hugging Face on first use. Supported model variants include 3B and 7B parameter models in FP16, FP8, and GGUF formats, with 4-bit and 8-bit GGUF quantization intended for lower-VRAM systems. Users can control batch processing, temporal overlap, prepended frames, maximum resolution, and deterministic seeding. The upscaler supports outputs at resolutions divisible by 2 and uses padding rather than restrictive cropping. Memory and execution features include BlockSwap transformer-block swapping, tiled VAE encoding and decoding, model and tensor offloading, model caching, streaming VAE decoding, and chunked processing for long videos. Attention backends include PyTorch SDPA, Flash Attention 2/3, SageAttention 2/3, and cuDNN attention, with fallback behavior when an optional backend is unavailable. torch.compile integration is optional. The repository can also be used independently of ComfyUI as a standalone CLI. The CLI supports single images, individual videos, and batch directory processing, with automatic MP4/PNG output detection, multi-GPU workload distribution, temporal-overlap blending, model caching, and memory-bounded streaming chunks. An FFmpeg video backend provides 10-bit x265 output as an alternative to 8-bit OpenCV output. Platform compatibility work covers NVIDIA CUDA, AMD ROCm, and Apple Silicon MPS, including older-GPU bfloat16 fallback handling. Color-processing features include LAB color transfer, HSV saturation matching, wavelet-based processing, and hybrid approaches. The project is released under the Apache 2.0 license.