SeedVR

SeedVR is an AI video restoration project for developers and researchers processing real-world footage or AI-generated video on their own GPUs. Its SeedVR2 model restores footage in a single diffusion step, with substantial hardware demands in the supplied implementation.

• One-step restoration: SeedVR2 generates restored video without the repeated diffusion steps that make high-resolution processing costly.
• High-resolution footage: The model supports 1080p restoration and adjusts its processing to the output resolution, with multi-GPU inference extending to 2K video.

The project is open source under Apache 2.0. The supplied inference setup uses NVIDIA H100 GPUs with 80 GB of memory: one handles 100 frames at 720p, while four support 1080p and 2K processing. These requirements make it a closer fit for GPU servers and research workstations than ordinary desktop hardware. Input height and width must be multiples of 32.

Heavy degradation and large motion can leave damage unresolved or cause unwanted generated details. Lightly degraded footage can become oversharpened, particularly at smaller output resolutions such as 480p.