Show HN: SeedVR2 – One-Step High-Resolution Video Restoration (www.aiupscaler.net)

SeedVR2 is a new one-step, high-fidelity video restoration model that compresses the usual multi-step diffusion pipeline into a single forward pass using a diffusion-transformer. It promises stable motion, sharper edges, and reduced compression artifacts at 1080p+ resolutions, delivering near real-time inference suitable for 4K-ready outputs. The demo emphasizes temporal texture stability (no plastic over-sharpening) and artifact removal in a single pass, making it attractive for film restoration, creator content, sports, surveillance, and e-commerce assets. Technically, SeedVR2 combines an adaptive window attention mechanism to scale efficiently to large frames, an adversarial post-training stage to boost perceptual realism, and a feature-matching loss that replaces traditional perceptual losses to stabilize adversarial training and preserve structural consistency. By avoiding iterative diffusion steps it cuts computational cost and reduces accumulated refinement errors, while the adversarial + feature-matching combo aims to balance naturalness against hallucination risks. For practitioners, that means faster, higher-resolution restoration with improved spatiotemporal coherence and a workflow-friendly single-step API for upscaling, denoising, deblocking, and frame export.