From Simulation to Survey: Benchmarking Super-Resolution for LSST-like Lensing Data

High-resolution imaging of strong gravitational lenses have the potential to proved unique insights into dark matter substructure and galaxy evolution, but these data are currently limited to a handful of examples from expensive space-based observations. Ground-based surveys like LSST will deliver orders of magnitude more lens systems, but at substantially lower resolution and higher noise levels. To bridge this gap, we present a synthetic dataset of paired high- and low-resolution gravitational lens images, designed to mimic the differences between space-based and ground-based instruments. Our dataset, generated using the Mejiro simulation framework, integrates physically motivated lens and source models with realistic instrumental effects, yielding 20,000 paired samples suitable for training and benchmarking super-resolution methods. We evaluate four representative approaches - RCAN, SwinIR, a conditional diffusion model, and SatGAN - spanning CNN, Transformer, diffusion, and adversarial paradigms. Our benchmarks on the Mejiro data set establishs a controlled testbed for conditional super-resolution in astronomy, with potential to enhance the scientific return of upcoming wide-field surveys.