The study asks whether the way a diffusion model receives conditioning information changes probabilistic daily precipitation downscaling in the Colorado River Basin. It compares conditioning mechanisms with an unconditioned baseline and assesses the generated precipitation in several ways: probabilistic scores, the shape of the precipitation distribution, spatial spectral structure, and the representation of extreme events. That gives the project a broader target than any one measure of agreement.
The task: a finer precipitation grid
The downscaling task starts with two linked data products. ERA5-Land supplies the target daily total precipitation at a 0.1-degree resolution. Dynamic atmospheric conditioning variables come from ERA5 at 0.25 degrees. In practical terms, the model is being asked to use atmospheric information on one grid to produce a precipitation target on another, finer grid.
That distinction matters because the study frames the comparison across several kinds of model behavior. It examines not only probabilistic output, but also the overall precipitation distribution, spatial spectral structure and behavior at the high end of rainfall. These are the four evaluation lenses named in the research question: probabilistic, distributional, spectral and extreme-event performance.
Three routes into the model
Three conditioned versions form the main comparison. The first uses channel concatenation, abbreviated CC. The second uses cross-attention with a learned convolutional encoder, or CA-CE. The third uses cross-attention with a frozen Prithvi WxC encoder, or CA-PWC. In this setting, conditioning means the information supplied to guide the precipitation output.
The two cross-attention versions differ in how their conditioning information is encoded: CA-CE uses an encoder learned for the task, while CA-PWC uses an encoder that remains frozen. An unconditioned diffusion model provides the baseline for asking whether adding conditioning improves the downscaling task at all. The labels are technical, but the central editorial question is straightforward: does the route used to feed atmospheric information change the kind of precipitation detail the model can represent?
Keeping the periods apart
The time split is explicit. Training uses 1985-2009, validation uses 2010-2012, and testing uses 2013-2015; the periods do not overlap. Inputs are normalized with scale factors from the training set. That arrangement keeps the evaluation period separate from the years used to fit and tune the models.
Training runs for 100 epochs. Fixed random seeds are used for both training and inference, and inference uses 20 sampling steps. The setup ends the sampling process at a noise level of zero and specifies a noise-spacing parameter of 7.
Instead of producing just one output for each test day, the inference procedure produces five stochastic realizations per day. Those repeated outputs allow the evaluation to examine probabilistic behavior rather than relying on a single generated field. The five-member output is part of the stated test design.
What the design is built to show
The structure makes the paper a comparison of model design choices within a defined precipitation-downscaling task. Its question is bounded by the Colorado River Basin, the two named reanalysis products, the daily total precipitation target, and the specified resolutions. Within that frame, it asks whether the conditioning choice changes probabilistic, distributional, spectral and extreme-event performance.
For readers trying to interpret the work, the key distinction is between the ingredients and the test. The ingredients are the three conditioning routes and the unconditioned baseline. The test keeps the calendar periods, normalization approach and sampling recipe in view while the models are assessed across multiple kinds of output behavior. The resulting setup is a focused benchmark of how diffusion models can be conditioned for this particular daily precipitation downscaling task.
Paper data and sources
Original title: Precipitation Downscaling Using Foundation Model-Conditioned Diffusion
Authors: Victor Nascimento Ribeiro, Jorge Guevara, Jorge Sebastian Moraga et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text