Self-Supervised Test-Time Adaptation for Spatio-Temporal Video Super-Resolution

Abstract

Spatio-Temporal Video Super-Resolution (STVSR) reconstructs high-resolution frames in space and time from low-resolution inputs. Existing models often degrade when motion dynamics at test time differ from training data. We propose a Test-Time Adaptation (TTA) framework that fine-tunes a pretrained STVSR decoder using only the low-resolution test clip. Our self-supervised method exploits recurring patches in adjacent input frames, reducing domain-shift artifacts and improving robustness. Applied to VideoINR and BF-STVSR, it consistently boosts test-time performance without external supervision.