---
license: mit
tags:
  - climate
  - downscaling
  - diffusion
  - pytorch
---

# ERA5→CONUS404 Diffusion Downscaling — Model + Full Dataset

41 yearly CONUS404 NetCDF files (WRF dynamical downscaling of ERA5 over
CONUS at 4km resolution, NCAR), 1980–2020. This is the model's raw training
target data, not preprocessed/cached tensors.

**For the current, verified-matching diffusion+VAE+DRN checkpoint set, use
[mudhil/era5-conus404-diffusion-downscaling](https://huggingface.co/mudhil/era5-conus404-diffusion-downscaling) instead.**


A three-stage "Latent CorrDiff" pipeline downscaling ERA5 (0.25°) to
CONUS404 (4km): a DRN predicts the conditional mean (49.6M params), a VAE
compresses the residual to a 64×64×8 latent (39.7M params), and an EDM
diffusion UNet (142.4M params) generates the latent residual. Six
variables: t2m/d2m/u10/v10/sp/tp → T2/TD2/U10/V10/PSFC/PREC_ACC_NC.