Snowpack in the western United States serves as a critical seasonal freshwater reservoir, yet Earth System models (ESMs) lack the spatial resolution to resolve snow water equivalent (SWE) at scales relevant to water resource management. We present SWEDE (Snow Water Equivalent Downscaling Emulator), a dual-encoder Swin-Transformer U-Net that downscales coarse atmospheric output to 9 km SWE across the western United States without receiving coarse SWE as input. We assess how SWEDE’s performance varies across a collection of diverse training setups. When trained on the Western U.S. Dynamically Downscaled Dataset (15 CMIP6 simulations from 14 ESMs from 1980–2100), SWEDE performs best (post-2050 MAE of 0.93 mm). In the WRF-ERA5 experiment, SWEDE achieves interannual mean SWE correlations of 0.89 to 0.99 and annual maximum SWE biases below 10 % across all six subregions during 2014–2024. When trained on a single ESM, performance degrades, but we find that the choice of ESM has an impact here, with high-warming ESMs outperforming low-warming ones (post-2050 MAE of 1.50 versus 1.91 mm). Lastly, both the multi- and single-ESM training setups substantially outperform a historical reanalysis setup, which struggles to emulate conditions under future warming, illustrating the importance of exposing the emulator to future climate states during training. These findings establish that climate-aware training data composition, particularly the inclusion of warm-climate sources, is essential for deep learning downscaling systems intended for use beyond the historical record.