mirror of
https://github.com/hanzoui/workflow-templates-broken-zap-2026-06-08.git
synced 2026-10-30 20:44:35 +00:00
1.5 KiB
1.5 KiB
SeedVR2
SeedVR2 is a one-step diffusion-based video restoration model developed by ByteDance Seed and NTU S-Lab, published at ICLR 2026.
Model Variants
SeedVR2-3B
- 3B parameter DiT with one-step inference for video and image upscaling
- Available in FP16, FP8, and GGUF quantized formats
SeedVR2-7B
- 7B parameter model with Sharp variant for maximum detail
- Multi-GPU inference; supports 1080p and 2K on 4x H100-80GB
SeedVR (Original)
- Multi-step diffusion model (CVPR 2025 Highlight)
- Arbitrary-resolution restoration without pretrained diffusion prior
Key Features
- One-step inference achieving 10x speedup over multi-step methods
- Adaptive window attention with dynamic sizing for high-resolution inputs
- Adversarial post-training against real data for faithful detail recovery
- Hanzo Studio integration via official SeedVR2 Video Upscaler nodes
- Apache 2.0 open-source license
Hardware Requirements
- Minimum: 8-12GB VRAM with GGUF quantization and tiled VAE
- Recommended: 24GB+ VRAM (RTX 4090) for 3B model at 1080p
- High-end: 4x H100-80GB for 7B model at 2K resolution
Common Use Cases
- Upscaling AI-generated video to 1080p or 4K
- Restoring degraded or compressed video footage
- Image super-resolution and detail recovery
Key Parameters
- resolution: target shortest-edge resolution (720, 1080, 2160)
- batch_size: frames per batch, must follow 4n+1 formula (5, 9, 13, 17, 21)
- seed: random seed for reproducible generation
- color_fix_type: wavelet, adain, hsv, or none