SeedVR2

SeedVR2 is a one-step diffusion-based video and image restoration model developed by the ByteDance Seed team in collaboration with S-Lab at Nanyang Technological University. Released as open weights under the Apache 2.0 license, the model specializes in upscaling low-resolution, blurry, or highly degraded media into sharp, high-quality outputs. Unlike conventional diffusion-based upscalers that require multiple sampling steps (typically 15 to 50), SeedVR2 utilizes a one-step process to perform super-resolution tasks efficiently without relying on a pre-trained diffusion prior.

Model Architecture and Variants

Under the hood, SeedVR2 is built on a large Diffusion Transformer (DiT) architecture and is released in two primary configurations: a 3B parameter version optimized for faster inference and lower VRAM environments, and a 7B parameter version designed for maximum detail recovery. The model utilizes Diffusion Adversarial Post-Training (DAPT) to align its single-step outputs with real-world distributions. To handle arbitrary high-resolution inputs, it incorporates an adaptive window attention mechanism. This dynamically adjusts the window size based on the output resolution, eliminating spatial-temporal inconsistencies and patch-boundary artifacts often produced by static window-attention architectures.

Capabilities and Limitations

The model supports processing for both single images and video sequences. Beyond standard upscaling, it can preserve transparent alpha channels in image sequences and maintain temporal consistency across consecutive video frames. Different checkpoint variations are available, including "sharp" editions (such as seedvr2_ema_7b_sharp) designed to boost fine details, though these may increase ringing artifacts on highly degraded sources. While highly effective at recovering details from low-resolution inputs, SeedVR2 has some constraints: it can struggle with extreme degradations or rapid physical motions, and its strong generative capacity may occasionally oversharpen or introduce unwanted details on clean, high-resolution original footage.