SeedVR2

SeedVR2 is a one-step diffusion-transformer model for generic video restoration from ByteDance Seed, published at ICLR 2026. It is trained with diffusion adversarial post-training against real data, so restoration takes a single forward pass rather than many sampling steps. It follows the earlier SeedVR model (CVPR 2025 Highlight).

The model uses an adaptive window attention mechanism, where the window size adjusts to the output resolution. This is meant to avoid window inconsistencies in high-resolution restoration, and the authors report restoration at up to 1080p in a single step. The paper also introduces a feature matching loss to stabilise adversarial training. The authors report results comparable to or better than existing video restoration methods.

Code and 3B weights are released under Apache-2.0 on GitHub and Hugging Face, with a Hugging Face Space for trying it. The authors describe the checkpoints as prototypes. They can fail on heavy degradations and very large motions, and they may over-generate detail and oversharpen lightly degraded inputs such as 720p AI-generated video. One YouTube reviewer comparing it with a newer image upscaler found SeedVR2 slow and prone to blurry edges and artifacts.