RiskFuse: Suppressing Unsupported Detail in Generative Super-Resolution

Generative super-resolution (SR) produces sharp images with rich high-frequency detail, but it can also introduce textures and edges unsupported by the low-resolution (LR) observation. These errors impair visual fidelity and can corrupt the evidence seen by downstream models when SR is used as preprocessing. We introduce RiskFuse, a post-processing method that suppresses unsupported detail while preserving much of the generative output's high-frequency texture. RiskFuse pairs the generative output with a frozen distortion-oriented output. A dense predictor locates regions where the generative result is less faithful; calibration on a disjoint validation split converts its predictions to a stable intervention scale; and LR consistency protects sharp detail that remains compatible with the input. The resulting fusion-weight map drives frequency-aware Laplacian-pyramid fusion without retraining either SR model. Across nine representative SR models spanning generative adversarial network and diffusion formulations, RiskFuse improves average PSNR-Y by 0.96 dB and LR-consistency PSNR by 1.03 dB, while reducing false-edge density by 56.3%. LPIPS increases modestly from 0.1134 to 0.1171. Across downstream case studies in detection and recognition, RiskFuse also improves representative task metrics over both direct LR input and the unmodified generative output. These results show that existing generative SR models can be made markedly more faithful without surrendering the visual detail that motivates their use.