SeedVR2: One-Step Video Super Resolution Model by ByteDance

SeedVR2 is ByteDance's one-step diffusion-based video super resolution model available in 3B and 7B variants with multiple quantization options.

S

SeedVR2

VideoVideo Super ResolutionVideo RestorationOne-StepByteDance

One-step video super resolution model by ByteDance Seed. Uses diffusion adversarial post-training for high-quality video restoration in a single inference step. Available in 3B and 7B parameter variants with FP16, FP8, INT8, MXFP8, and NVFP4 quantizations, plus Sharp variants for enhanced detail.

Hugging Face (Comfy-Org)Upstream 3BUpstream 7BPaperGitHubProject Page
DeveloperByteDance Seed
Release Date2025-06
ArchitectureDiffusion Adversarial Post-Training (One-Step)
Model Sizes3B, 7B (+ Sharp variants)
LicenseApache-2.0
CapabilitiesVideo super resolution, video restoration, one-step inference

Guides and workflows related to this model series.

No articles found.