AEmotionStudio/SeedVR2-models

SeedVR2 — Quantized Model Mirror

Quantized and converted checkpoints for SeedVR2 by ByteDance-Seed.

Original model: ByteDance-Seed/SeedVR

Available Checkpoints

File	Size	Description
seedvr2_ema_3b_fp8_e4m3fn.safetensors	3.39 GB	3B model, FP8 quantized
seedvr2_ema_3b-Q4_K_M.gguf	1.91 GB	3B model, GGUF Q4_K_M quantized
seedvr2_ema_7b_fp8_e4m3fn.safetensors	8.24 GB	7B model, FP8 quantized
seedvr2_ema_7b-Q4_K_M.gguf	4.64 GB	7B model, GGUF Q4_K_M quantized
ema_vae_fp16.safetensors	501 MB	VAE (FP16)

Model Details

SeedVR2 is a high-quality video super-resolution model that upscales low-resolution video to high resolution using diffusion-based super-resolution. Available in 3B and 7B parameter variants.

• Architecture: DiT-based diffusion model
• Task: Video super-resolution / upscaling
• Paper: SeedVR: Seed Video Restoration

Usage with ComfyUI-FFMPEGA

These models are automatically downloaded and used by the ComfyUI-FFMPEGA extension via the AI Upscale mode.

License

Apache 2.0 — see LICENSE file. Same license as the original model.