SeedVR2 Cog (3B & 7B)

Overview

SeedVR2 Cog packages ByteDance-Seed's one-step diffusion transformer for restoring soft, noisy, or artifact-heavy footage. This repository packages the official release as a Cog project so you can run the model on Replicate or locally with the same configuration.

Status (Nov 9 2025): The predictor now hot-swaps between the SeedVR2 3B and 7B checkpoints on a single GPU. The build targets 80 GB accelerators (A100) and larger cards (H100/H200+) without relying on multi-GPU parallelism.

Highlights

• One-step restoration for both videos and single-frame images with optional wavelet color correction to match the official Gradio demo.
• Audio passthrough keeps the original soundtrack when producing MP4 outputs.
• Dual-model cache: 3B loads by default, and 7B can be selected at runtime.
• All large assets download from Replicate's CDN via pget for reproducible builds.
• CUDA 12.4 / PyTorch 2.4.0 environment mirrors Replicate's production base image.

What this repository includes

• predict.py wraps the original inference code for both video and single-frame inputs, returning WebP (by default) or PNG/JPG stills and MP4 video with the source audio stream passed through when available.
• cog.yaml defines the CUDA 12.4 build with the dependencies Replicate expects, including pre-installation of flash-attn and runtime checks for Apex.
• cache_manager.py helps bundle large weights into tarballs and publish them to the Replicate CDN, keeping cog predict downloads fast and deterministic.

Try the model

• In the browser: open replicate.com/zsxkib/seedvr2

• From the CLI: install Cog and run:

  cog predict \
    -i media="https://replicate.delivery/pbxt/O0Iv6e7e6eXVjeuEy6bai5SYhkU2EEBxAgYQHp6Lc9kWUHBf/q6e4gnsd5xrma0ct4qt9zt4vd8.mp4" \
    -i sample_steps=1 \
    -i cfg_scale=1 \
    -i sp_size=1 \
    -i fps=24 \
    -i output_format="png" \
    -i output_quality=90

  The call detects whether media is a video or image and selects the correct preprocessing and output branch automatically.

Inputs

Name	Type	Description	Default
media	file / URL	Video (.mp4, .mov) or image (.png, .jpg, .webp).	–
sample_steps	integer	Diffusion steps (1 matches the public demo).	1
cfg_scale	float	Guidance strength; raise for stronger sharpening.	1.0
sp_size	integer	Keep at 1 for single-GPU runs; higher values only adjust padding heuristics (no multi-GPU).	1
fps	integer	Frame rate for video outputs.	24
seed	integer?	Optional random seed (auto-randomized when omitted).	random
output_format	string	Image output format when media is an image.	webp
output_quality	integer	JPEG/WebP quality when using lossy formats.	90
model_variant	string	Choose between the 3B and 7B checkpoints ("3b" or "7b").	3b
apply_color_fix	boolean	Apply wavelet color correction (matches the official demo).	false

GPU expectations

• The predictor is optimized for a single GPU: 1× A100 80 GB, 1× H100 80/94 GB, 1× H200 141 GB, or future single-card equivalents.
• sp_size remains fixed at 1; sequence-parallel sharding is not used in this deployment. Longer clips are padded internally to satisfy the sampler layout.

Outputs

• Video inputs return an MP4 file whose audio track is copied from the source when ffmpeg is present (the Cog image installs it by default).
• Image inputs return a single restored frame in the requested format.

Run locally

pip install cog
git clone https://github.com/zsxkib/cog-ByteDance-Seed-SeedVR2.git
cd cog-ByteDance-Seed-SeedVR2
cog predict -i media=@test_videos/01.mp4

The first run downloads the cached weights, Apex wheel, and flash-attn build from the Replicate CDN into model_cache/. Subsequent runs reuse those files.

Weights and caching

• Weight archives live at https://weights.replicate.delivery/default/seedvr2/.
• cache_manager.py can regenerate the tarballs and upload replacements if the underlying checkpoints change.
• Environment variables (HF_HOME, TORCH_HOME, etc.) are set to model_cache/ before any downloads so cached files persist across runs.

Development notes

• The predictor removes the need for Hugging Face downloads by mirroring the release assets at https://weights.replicate.delivery/default/seedvr2/.
• mux_audio_stream keeps audio in sync by copying the source stream onto the generated video without re-encoding.
• CUDA 12.4 + PyTorch 2.4.0 matches Replicate's cuda12.4-python3.10-torch2.4.0 base image, so no extra drivers are needed for deployment.

Deploy to Replicate

Create or update a Replicate model pointing at r8.im/zsxkib/seedvr2, then:

cog push r8.im/zsxkib/seedvr2

The push builds the Docker image, validates the schema, and publishes the artifacts.

Citation

Please reference the original SeedVR/SeedVR2 work if you build on these models:

@article{wang2025seedvr2,
  title   = {SeedVR2: One-Step Video Restoration via Diffusion Adversarial Post-Training},
  author  = {Wang, Jianyi and Lin, Shanchuan and Lin, Zhijie and Ren, Yuxi and Wei, Meng and Yue, Zongsheng and Zhou, Shangchen and Chen, Hao and Zhao, Yang and Yang, Ceyuan and Xiao, Xuefeng and Loy, Chen Change and Jiang, Lu},
  journal = {arXiv preprint arXiv:2506.05301},
  year    = {2025}
}

@inproceedings{wang2025seedvr,
  title     = {SeedVR: Seeding Infinity in Diffusion Transformer Towards Generic Video Restoration},
  author    = {Wang, Jianyi and Lin, Zhijie and Wei, Meng and Zhao, Yang and Yang, Ceyuan and Loy, Chen Change and Jiang, Lu},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year      = {2025}
}

License

• All code in this repository is released under the MIT License (see LICENSE).
• Upstream SeedVR/SeedVR2 weights and research assets remain under Apache 2.0 per ByteDance-Seed’s terms. Bring your own weights or follow their license when you redistribute the checkpoints.