SeedVR2
What it is
SeedVR2 is a free, open-source model for restoring video at arbitrary resolutions through a single-step diffusion-transformer approach. The project describes restoration without an additional diffusion prior. Its official ComfyUI release can upscale images as well as videos and provides four nodes for controlling the process. A standalone command-line interface is also available, including multi-GPU and batch-processing workflows. Supported inputs are RGB and RGBA images and video. The release documents 3B and 7B model variants, with FP16, FP8, and quantized GGUF options. Model offloading, VAE tiling, and BlockSwap are documented as ways to manage memory during inference. The research paper describes adaptive window attention and adversarial post-training for high-resolution restoration; the attention method adjusts window size to the output resolution. Standalone installation instructions cover Windows, macOS, and Linux, and Apple Silicon MPS support is documented. Input height and width must be multiples of 32; the demo crops resized inputs to meet this constraint. The project is licensed under Apache 2.0 and links to inference code and model weights. Its paper identifies the work as ICLR 2026.
For and against
For
- Upscales both images and videos.
- Offers ComfyUI nodes and a standalone CLI.
- Includes multi-GPU and batch workflows.
- Documents FP16, FP8, and quantized GGUF options.
- Licensed under Apache 2.0.
Against
- Input dimensions must be multiples of 32.
- Standalone use requires installation and model weights.
The verdict
SeedVR2 provides image and video upscaling in ComfyUI and through a command-line interface, with several model and memory-management options. Users should account for its input-dimension constraint when preparing material.
Who it suits
SeedVR2 suits people seeking image or video upscaling through ComfyUI or a command-line workflow. Its documented model variants and memory controls may be relevant to users setting up inference on Windows, macOS, Linux, or Apple Silicon.
Plans and prices
SeedVR2
No paid plans or usage limits stated on the project pages
Read 10 Oct 2026 · github.com
On the record
- Adaptive attention
- Its adaptive window attention adjusts window size to fit output resolutions.github.com · read 10 Oct 2026
- Authors’ affiliations
- The listed affiliations are S-Lab at Nanyang Technological University and ByteDance Seed.iceclear.github.io · read 10 Oct 2026
- ComfyUI integration
- The release provides four ComfyUI nodes for controlling the upscaling pipeline.github.com · read 4 Oct 2026
- Downloads
- The repository links to released inference code and model weights.github.com · read 10 Oct 2026
- High-resolution results
- The project page describes efficient single-step restoration at 1080p with faithful details.iceclear.github.io · read 10 Oct 2026
- Image and video upscaling
- The official ComfyUI release supports upscaling both videos and images.github.com · read 4 Oct 2026
- Inference
- The model restores video in a single step without relying on an additional diffusion prior.github.com · read 10 Oct 2026
- Input constraint
- Input height and width must be multiples of 32; the demo crops resized inputs to meet that constraint.iceclear.github.io · read 10 Oct 2026
- Integration
- The maintainers state that a ComfyUI integration was released in November 2025.github.com · read 10 Oct 2026
- License
- The project is released under the Apache License 2.0.github.com · read 10 Oct 2026
- Model options
- The release documents 3B and 7B model variants, with FP16, FP8, and quantized GGUF options.github.com · read 4 Oct 2026
- Performance controls
- The release documents model offloading, VAE tiling, and BlockSwap to manage memory use during inference.github.com · read 4 Oct 2026
- Platform support
- Standalone installation instructions cover Windows, macOS, and Linux; Apple Silicon MPS support is also documented.github.com · read 4 Oct 2026
- Purpose
- SeedVR2 is a one-step diffusion-transformer model for video restoration at any resolution.github.com · read 10 Oct 2026
- Research
- The project page identifies the work as ICLR 2026.iceclear.github.io · read 10 Oct 2026
- Research affiliation
- The paper lists the authors’ affiliations as S-Lab at Nanyang Technological University and ByteDance Seed.iceclear.github.io · read 4 Oct 2026
- Research method
- The paper describes adaptive window attention and adversarial post-training for high-resolution video restoration.iceclear.github.io · read 4 Oct 2026
- Security
- A release note says model loading was restricted to tensor-only deserialization to prevent code execution from malicious .pth files.github.com · read 4 Oct 2026
