Animation Super-Resolution via Explicit Line Art Structure Loss

This paper proposes an optimization framework for animation super-resolution (ASR) to address line art structural collapse. Current state-of-the-art (SOTA) methods rely heavily on implicit supervision and often fail to preserve line integrity. To solve this, we introduce an explicit line art structure loss. Unlike prior approaches working solely in pixel or semantic domains, our loss directly constrains the geometric attributes of lines which explicitly ensures their sharpness and topological continuity. Specifically, our method integrates a differentiable, adaptive-threshold Soft XDoG operator and combine this with gradient profile priors (GPP) and gradient orientation consistency to guide the reconstruction. Experiments on the API dataset demonstrate the effectiveness of our approach. It significantly outperforms existing methods in maintaining line integrity and structural preservation.