Skip to the content
Know your tech before you buy it. Makers: get listed from $49/mo

SeedVR2

github.com

  • Free plan
  • Windows
  • macOS
  • Linux
  • Self-hosted
7.9 Club score Visit github.com

What SeedVR2 is

SeedVR2 is an open-source model for restoring video at arbitrary resolutions. Its project describes a one-step diffusion-transformer method that restores video without an additional diffusion prior. The official ComfyUI release can upscale both video and images, with four nodes for controlling the process. A standalone command-line interface is also available for multi-GPU and batch workflows. Inputs can be RGB or 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 available to manage memory during inference. The research paper describes adaptive window attention and adversarial post-training for high-resolution restoration; the attention method adjusts its window size to fit output resolutions. Standalone installation instructions cover Windows, macOS, and Linux, with Apple Silicon MPS support also documented. Input height and width must be multiples of 32, and the demo crops resized inputs to meet that constraint. SeedVR2 is licensed under Apache 2.0, and the repository links to inference code and model weights.

Who it suits

It suits people restoring or upscaling video and images who can work with ComfyUI or a standalone command-line interface. The documented standalone platforms include Windows, macOS, and Linux.

Good to know, and what to watch

  • Upscales both videos and images
  • ComfyUI release includes four pipeline-control nodes
  • Standalone interface supports multi-GPU and batch workflows
  • Offers 3B and 7B model variants
  • Memory controls include offloading, VAE tiling, and BlockSwap
  • Input height and width must be multiples of 32

SeedVR2

Free

Open source

The Club's review

SeedVR2 is an open-source diffusion-transformer for restoring video and upscaling images or video locally. It is best for technically confident creators, developers, and operators who can use ComfyUI or a command line. The trade-off is control and no software fee in exchange for setup work and strict input dimensions.

Overview

SeedVR2 performs restoration in one inference step rather than adding a separate diffusion prior. Its design combines adaptive window attention, which changes window size to suit the requested resolution, with adversarial training on real video and a feature-matching loss. The project describes efficient 1080p restoration with faithful detail, making it a focused option for high-resolution enhancement rather than a general video editor.

Images and video must have height and width divisible by 32. The supplied demo crops resized inputs to satisfy that rule, so edge framing can change. Apache 2.0 licensing, published inference code, and released model weights make it practical to adapt or operate in private infrastructure.

Key features

  • Two operating routes: Four ComfyUI nodes expose controls for a node-based upscaling pipeline. The standalone command-line interface adds batch and multi-GPU workflows, which is useful for repeatable production jobs but assumes users can install and maintain a technical environment.
  • Image and video support: The official ComfyUI release upscales both media types and accepts RGB or RGBA inputs. That covers common still and video assets without requiring separate tools.
  • Model and precision choices: SeedVR2-3B and SeedVR2-7B variants are documented alongside FP16, FP8, and quantized GGUF options. More choices let operators balance model capacity and hardware constraints, while selecting an appropriate build becomes part of setup.
  • Memory controls: Model offloading, VAE tiling, and BlockSwap provide ways to manage inference memory. These controls can make local deployment more feasible on constrained systems, though they add configuration decisions rather than delivering a one-click experience.
  • Resolution-aware research: Adaptive attention is intended to fit different output resolutions, and the project reports single-step 1080p results. Those claims point to a restoration-first design for demanding resolutions, not a broad suite of editing functions.
  • Safer model loading: A release note restricts loading to tensor-only deserialization to block code execution from malicious .pth files. That is a meaningful safeguard when model files come from outside a tightly controlled environment.

Pricing

SeedVR2 costs nothing. It is open source, has no paid plans or stated usage limits, and has no free trial because the software itself is free rather than time-limited. The saving is substantial for teams processing locally; users instead supply compatible hardware, storage, and the expertise needed to configure models and workflows. No seat, renewal, or hosted quota terms are stated.

Platforms

Standalone installation instructions cover Windows, macOS, and Linux, with Apple Silicon MPS support documented. The project is self-hosted, so processing stays within the machine or infrastructure you configure rather than relying on a hosted web service. ComfyUI and command-line operation make it suitable for workstation and automated environments, but neither provides the polished desktop workflow of a conventional editor.

Who it's for

Choose SeedVR2 if you need repeatable, local restoration, want ComfyUI nodes or batch and multi-GPU automation, and can accommodate model and memory configuration. It is particularly relevant to developers, technical artists, and teams handling sensitive footage that should remain on their own systems.

Look elsewhere if you expect a guided consumer interface, cloud processing, or unrestricted image dimensions. The 32-pixel divisibility rule and possible cropping require attention before large jobs, and the project does not present itself as an end-to-end editing application.

Pros and cons

Pros

  • Free Apache 2.0 release: There is no software charge or stated processing quota, and the license supports private deployment and adaptation.
  • Flexible automation: ComfyUI nodes, batch execution, and multi-GPU support cover both interactive graphs and production pipelines.
  • Hardware-aware options: 3B/7B models, several precisions, offloading, tiling, and BlockSwap help match workloads to available memory.
  • Broad local compatibility: Windows, macOS, Linux, and Apple Silicon MPS documentation widen the range of machines that can host it.

Cons

  • Technical setup is unavoidable: ComfyUI or CLI operation and model management will frustrate users seeking a simple editor.
  • Dimension restriction: Inputs must be multiples of 32, and demo cropping can alter composition.
  • No hosted convenience: Local operation means the user provides compute and handles performance tuning.

Alternatives

AI Video Restoration Software is a broader directory starting point when you want to compare restoration tools rather than adopt a model toolkit.

AI Video Upscalers is the better comparison route when upscaling products and services are your priority.

TensorPix suits users who prefer a web, API, or self-hosted freemium service with a free-forever tier, weekly login credits, 720p input, and 1 GB storage.

Aiarty Video Enhancer is a fit for Windows or macOS users wanting a desktop product with a free version limited to videos under 120 seconds and watermarked exports.

HitPaw VikPea is an alternative for users seeking a freemium product across desktop, mobile, web, and API platforms.

Refocused is suited to local Windows, macOS, or Linux processing with a full-featured watermarked trial before its paid license.

KwaFlux Video File Repair is an alternative Windows desktop choice for video-file repair workflows.

Surveillant Forensic Video Enhancement targets forensic enhancement through hosted, API, mobile, and self-hosted deployments.

Vapourkit is a free, open-source Linux and Windows application for users who want a different local toolkit.

Klarity is another free, local, offline restoration tool for Linux and Windows.

Verdict

SeedVR2 is a strong choice for practitioners who want no-cost, self-hosted restoration with ComfyUI control or scalable command-line automation. Its main reason to choose it is the combination of open licensing, model variants, and memory-aware workflows. Choose another product if you need a guided desktop or cloud experience, or if enforcing 32-pixel dimensions and configuring local inference is more work than your workflow allows.

The verdict

SeedVR2 offers a single-step restoration workflow through ComfyUI or its command-line interface, with multiple model and precision options. The input dimension constraint and documented memory controls are relevant when planning a workflow.

What the maker publishes

Adaptive attention
Its adaptive window attention adjusts window size to fit output resolutions. source
Authors’ affiliations
The listed affiliations are S-Lab at Nanyang Technological University and ByteDance Seed. source
ComfyUI integration
The release provides four ComfyUI nodes for controlling the upscaling pipeline. source
Downloads
The repository links to released inference code and model weights. source
High-resolution results
The project page describes efficient single-step restoration at 1080p with faithful details. source
Image and video upscaling
The official ComfyUI release supports upscaling both videos and images. source
Inference
The project describes restoring video in a single step without an additional diffusion prior. source
Inference
The model restores video in a single step without relying on an additional diffusion prior. source
Input constraint
Input height and width must be multiples of 32; the demo crops resized inputs to meet that constraint. source
Integration
The maintainers state that a ComfyUI integration was released in November 2025. source
License
The SeedVR2 project states that it is licensed under Apache 2.0. source
License
The project is released under the Apache License 2.0. source
Model options
The release documents 3B and 7B model variants, with FP16, FP8, and quantized GGUF options. source
Performance controls
The release documents model offloading, VAE tiling, and BlockSwap to manage memory use during inference. source
Platform support
Standalone installation instructions cover Windows, macOS, and Linux; Apple Silicon MPS support is also documented. source
Purpose
SeedVR2 is a one-step diffusion-transformer model for video restoration at arbitrary resolutions. source
Purpose
SeedVR2 is a one-step diffusion-transformer model for video restoration at any resolution. source
Research
The project page identifies the work as ICLR 2026. source
Research affiliation
The paper lists the authors’ affiliations as S-Lab at Nanyang Technological University and ByteDance Seed. source
Research method
The paper describes adaptive window attention and adversarial post-training for high-resolution video restoration. source
Security
A release note says model loading was restricted to tensor-only deserialization to prevent code execution from malicious .pth files. source
Standalone use
The release also provides a standalone command-line interface, including multi-GPU and batch processing workflows. source
Support
The project README invites questions by email at [email protected]. source
Support
The repository invites questions by email at [email protected]. source

Checked against the maker's pages on 10 Oct 2026.