GenFocal — official pretrained climate debiasing and downscaling
Paper · Author code · Published checkpoints
GenFocal uses the author's unchanged JAX/Flax implementation and published
parameters. GenFocal_lite is a separate, untrained PyTorch approximation.
The main entry never falls back to lite or initializes random replacement weights.
from weatherai import GenFocal, GenFocal_lite
sr = GenFocal.from_pretrained("super_resolution")
flow = GenFocal.from_pretrained("debiasing")
# Author-normalized arrays and conditioning; this is a network evaluation:
y = sr(x, sigma, cond)
# Optional explicitly untrained experimentation:
lite = GenFocal_lite(channels=4)
Use Linux/Python 3.12 and the pinned environment in
hf_official/requirements.txt.
Set WEATHERAI_CACHE to a large local directory outside synchronized folders.
Author source is fetched at commit 0beb5d20a5b31969e12069ba7d86fc0537e08528
and checked for modifications. Published OCDBT tensors are fetched directly
from the public genfocal GCS bucket via HTTP range reads, with shape, dtype
and SHA-256 records. Optimizer state and duplicate checkpoints are excluded.
Flax module names containing literal dots are restored from Orbax key metadata.
| Stage | Released checkpoint (step 300000) | Parameter tree |
|---|---|---|
| ReFlow debiasing | debiasing/checkpoints/reflow_uvit |
raw params, as in author's sampling code |
| CONUS super-resolution | super_resolution/checkpoints/conus_7d_t2m_w10m_q1000_msl |
ema_state.ema, as in author's inference |
Correction to the earlier documentation: the released debiasing checkpoint
uses channels 768, 768, 768, 1024, 1280, 1536, not the 128-channel example
configuration. Its published run config is loaded from the checkpoint directory.
The raw debiasing tree contains 2,656,553,626 parameters (10.63 GB).
SR uses the author's conus_sample.yaml (128/256/384/512 channels, 84 two-hourly
steps, four output variables). This distinction is essential for weight compatibility.
This entry loads/evaluates each official network. Full climate trajectories also require the author's input normalization, ReFlow integration, diffusion sampling, regional baseline and residual statistics, overlap handling and physical-unit reconstruction. A single denoiser/vector-field call is not a completed forecast. Both complete pretrained networks were executed on HF ZeroGPU: ReFlow 146.13 s (synthetic normalized input), CONUS SR 52.99 s (real ERA5 conditioning, one denoiser call). All outputs were finite. See current validation evidence and limitations.
Official code is Apache-2.0; released checkpoints are CC-BY-4.0. Original weights
remain with the authors and are not redistributed in Git. Previous lite loss/RK4
parity results remain in docs/results/genfocal_parity.json; they do not establish
full pretrained model parity or climate skill.