Video catalogue · Verified local path

SeedVR2 local guide

Apache-licensed one-step diffusion model for local video restoration, super-resolution and upscaling.

Choose an app

Start with the recommended app, or choose another supported option.

Compare video models
Open ComfyUI setupRecommended verified starting pointOpen in UnslothLaunches Unsloth Desktop on this modelOpen on Hugging FaceModel files, licence and model cardOpen advanced setupPython runtime for experienced usersCheck my machineSee whether this model fits your saved hardware

Desktop app links require the app to be installed. If nothing opens, LocalClaw will show app-download and model-file fallbacks.

What it does

Video To VideoVideo RestorationVideo Super ResolutionVideo UpscalingImage Upscaling

ByteDance Seed publishes Apache-2.0 SeedVR2 code and official 3B/7B Hugging Face checkpoints, including seedvr2_ema_3b.pth, seedvr2_ema_7b.pth and the VAE assets used by the reference PyTorch inference scripts. The research README documents Python 3.10, CUDA, FlashAttention, Apex, checkpoint downloads and torchrun inference, with 1 H100 80GB handling 100-frame 720p and 4 H100 80GB supporting 1080p/2K. For practical local use, Comfy Org republishes converted 3B and 7B FP16/FP8/INT8/NVFP4 safetensors plus VAE safetensors, workflow templates and a ComfyUI/standalone CLI path; the maintained ComfyUI integration documents Windows, Linux, macOS Apple Silicon, AMD ROCm and low-VRAM 8-12 GB workflows using GGUF, BlockSwap, CPU offload and VAE tiling. LocalClaw records 32 GB RAM and 8 GB accelerator memory as the conservative low-VRAM entry floor, with higher VRAM recommended for FP16 and longer clips.

Hardware figures are practical entry floors, not performance guarantees. Resolution, duration, precision, offloading and runtime versions can materially change memory use.

Strengths

  • Official ByteDance weights plus ComfyUI-packaged safetensors
  • Video and image restoration with ComfyUI workflows and standalone CLI
  • Low-VRAM path documents 8-12 GB optimization with offload and tiling

Limits to know

  • Original research inference targets H100-class hardware for high-resolution video
  • Low-VRAM workflows trade speed and resolution for fit
  • Restoration/upscaling model, not prompt-only scene generation