SeedVR2
  • Pricing
  • Compare
SeedVR2

Source-verified model guide

SeedVR2 Hugging Face: Official 3B/7B & ComfyUI Files

Use the right publisher, file, and folder before downloading. ByteDance publishes the official 3B and 7B checkpoints; Comfy-Org provides a ComfyUI repack; NumZ maintains a separate community node.

Who publishes what

These projects are related, but they are not interchangeable. The labels below prevent a community integration from being presented as an official ByteDance release.

Official ByteDance Seed

The upstream repository and the SeedVR2-3B / SeedVR2-7B Hugging Face repositories are the official model sources.

Hugging Face / GitHub

Comfy-Org repack

Comfy-Org republishes model files in the layout used by native ComfyUI integration. It does not replace ByteDance as the upstream model publisher.

Hugging Face / GitHub

NumZ community node

NumZ maintains a community custom node with additional workflow and memory options. Treat its claims and files as community-maintained.

Hugging Face / GitHub

Which SeedVR2 file should you download?

Download order matters more than names. The native ComfyUI template expects one 3B INT8 diffusion file plus the shared VAE; the official 7B checkpoint, FP8, and GGUF are heavier or community-only options. ByteDance documents H100 inference and the ComfyUI integration maintainer reports testing from roughly 18GB VRAM minimum, so no row below is an 8GB guarantee.

FilePublished byComfyUI folderWhere it fitsDownload it when
seedvr2_3b_int8_convrot.safetensorsComfy-Org repackComfyUI/models/diffusion_models/The 8GB and 12GB tiers start from 3BAlways first: this is the diffusion model the ComfyUI 0.28.0 template selects.
seedvr2_ema_vae_fp16.safetensorsComfy-Org repackComfyUI/models/vae/Shared by the image and video templatesTogether with the diffusion model, before the first sample run.
ByteDance-Seed/SeedVR2-3BByteDance Seed, Apache-2.0Upstream repository, not a ComfyUI folderThe official 3B sourceWhen you want the publisher original, or you are not running ComfyUI.
ByteDance-Seed/SeedVR2-7BByteDance Seed, Apache-2.0Upstream repository, not a ComfyUI folder12GB conservative test; 16GB or more with FP8/FP16Only after a 3B baseline passes and the source still needs more detail.
FP8 repackComfy-Org and integration repackComfyUI/models/diffusion_models/12GB and 16GBWhen 3B INT8 fits but you want a precision step before FP16.
GGUF Q4_K_M / Q8_0NumZ community nodeModels folder named by the community node8GB experimentsOnly after the native template fails: quantization is community work with no 8GB guarantee.

FP8 and GGUF are not ByteDance releases; they exist because Comfy-Org and the NumZ community node repack the model. A community file can load and still look worse than the official 3B INT8 baseline, so compare one 100 percent crop before committing a batch.

8GB card: read the low-memory order of operations first

Official snapshot_download method

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="ByteDance-Seed/SeedVR2-3B",
    local_dir="./SeedVR2-3B",
)

3B: ByteDance-Seed/SeedVR2-3B
7B: ByteDance-Seed/SeedVR2-7B

ComfyUI file directories

For the Comfy-Org package, put diffusion model files under ComfyUI/models/diffusion_models/ and the VAE under ComfyUI/models/vae/. Keep filenames unchanged, restart ComfyUI, then select the exact files in the workflow.

ComfyUI/models/diffusion_models/
ComfyUI/models/vae/

Hardware evidence—not an 8GB promise

ByteDance shows an H100 inference setup. The ComfyUI integration maintainer reports testing at roughly 18GB VRAM minimum. An 8GB run may be possible only through community quantization, tiling, offload, or BlockSwap experiments; neither source makes it an official guarantee.

Official limitations to check before a large run

  • Heavy degradation can exceed what the model can restore reliably.
  • Large motion can cause restoration failure or temporal inconsistency.
  • Lightly degraded sources can be over-sharpened.

License

The ByteDance SeedVR2 repository and official checkpoints are published under Apache-2.0. Review the repository license and any dependency or input-media rights before commercial use.

Primary sources

  • github.com/ByteDance-Seed/SeedVR
  • huggingface.co/ByteDance-Seed/SeedVR2-3B
  • huggingface.co/ByteDance-Seed/SeedVR2-7B
  • huggingface.co/Comfy-Org/SeedVR2
  • github.com/comfyorg/comfyui_seedvr2/blob/main/README.md
  • github.com/numz/ComfyUI-SeedVR2_VideoUpscaler
Open the complete ComfyUI installation guideUse the hosted upscaler
SeedVR2

SeedVR2 is an independent AI upscaling workspace for image and video upscaling, clear job tracking, and delivery workflows.

Launched on FazierLaunched on FazierFeatured on ToolPilot.ai
Company
  • About
  • Blog
  • Tutorial
  • Support
Legal
  • Terms of Service
  • Privacy Policy
  • Refund Policy
  • Acceptable Use Policy
SeedVR2 resources
  • SeedVR2 model download (3B/7B)
  • SeedVR2 ComfyUI workflow
  • SeedVR2 Hugging Face models (3B/7B)
  • SeedVR2 GitHub repositories
  • SeedVR2 upscaler download (no app)

Get product updates

We'll email you once when a new feature or model launches. No spam.

SeedVR2 is an independent service that provides access to third-party AI models and workflows. We are not affiliated with, endorsed by, or sponsored by the model providers.

© 2026 SeedVR2 All Rights Reserved.