GCD: GCM-consistent Diffusion for Zero-shot Downscaling across Heterogeneous GCMs
Authors: Ruian Tie, Wenbo Xiong, Zhengyu Shi, Xinyu Su, Chenyu jiang, Libo Wu, Hao Li
Organizations: Artificial Intelligence Innovation and Incubation Institute, Fudan University, Shanghai, China · Shanghai Innovation Institute, Shanghai, China · Shanghai Academy of AI for Science, Shanghai, China · School of Data Science, Fudan University, Shanghai, China · Institution for Big Data, Fudan University, Shanghai, China
Climate downscaling of General Circulation Models (GCMs) is an ill-posed inverse problem, as GCM outputs and reanalysis fields arise from different dynamical systems and lack fine-scale temporal pairing. Supervised methods thus rely on synthetic degradations or model-specific adaptation, limiting generalization across heterogeneous GCMs. Zero-shot diffusion posterior sampling (DPS) avoids paired supervision by combining a high-resolution climate prior with inference-time constraints from coarse GCM inputs, but faces consistency and efficiency challenges. We propose GCM-consistent Diffusion for Zero-shot Downscaling (GCD), which addresses these challenges with geographic conditioning, spectrally filtered measurement guidance, and distillation-accelerated Conjugate Gradient guidance. GCD reduces inference from 1,000 diffusion steps to 6 while maintaining consistency with the coarse GCM inputs. Across five heterogeneous GCMs, GCD achieves competitive 99th-percentile extreme-event accuracy and reconstructs plausible high-frequency tropical cyclone structures without model-specific retraining. Github: https://github.com/shawnvonbohr/GCD-Climate-Downscaling.git
Figures & tables
Figure 1: Schematic comparison of different downscaling paradigms during training and inference.
Figure 2: Overview of the GCD framework. (a) Training Stage: a geographically-constrained climate diffusion prior conditioned on static terrain and temporal embeddings. (b) Inference Stage: a filter measurement operator is applied during reverse diffusion sampling to reduce the spectral bias between the prior and GCMs. (c) Inference Acceleration: we use distillation and conjugate gradient guidance to compress the teacher into a fast student model.
Figure 3: Schematic illustration of the geometries of the two guidance processes (not measured data).
Method
IPSL
MIROC6
AWI
MPI-LR
MPI-HR
Mean Inf. Time per Sample (s)
per GPU Hours (h)
Bilinear
1.06 / 1.78
1.84 / 2.43
1.50 / 2.19
1.48 / 2.12
1.32 / 1.90
1.80×10−5
–
Cubic
1.03 / 1.79
1.82 / 2.42
1.50 / 2.18
1.47 / 2.12
1.30 / 1.87
1.37×10−4
–
BCSD [ 30 ]
0.98 / 1.67
1.72 / 2.33
1.36 / 2.07
1.35 / 1.98
1.24 / 1.76
0.27
–
SwinIR [ 31 ]
1.01 / 1.67
1.52 / 2.05
1.37 / 2.02
1.35 / 1.95
1.18 / 1.68
0.25
11X
ESRGAN [ 32 ]
1.00 / 1.66
1.50 / 2.03
1.36 / 2.00
1.34 / 1.93
1.17 / 1.67
0.04
22X
CorrDiff [ 33 ]
0.98 / 1.62
1.45 / 1.95
1.33 / 1.96
1.31 / 1.89
1.14 / 1.62
2.33
35X
Table 1: Quantitative comparison of downscaling methods on different GCMs. The first five columns report the spatial MAE/RMSE averaged over the test period (2001-2009) for each respective GCM. Mean Inf. Time denotes the average inference time per sample across all five GCMs. Per GPU Hours reports the computational cost on a single GPU in training stage. Notably, the ‘X’ multiplier indicates that the cost must be multiplied by the number of target GCMs (e.g., requiring separate training for each specific GCM). Best results are highlighted in bold and underline denotes the second-best.
Figure 4: Spectral and spatial demonstration of GCD (1000 NFE) capabilities. (a) Annual mean power spectra of 10m Wind Speed (WS10m). (b) Visual comparison of WS10m and TCs. The top row displays raw outputs from various GCMs. The bottom row shows the corresponding high-resolution fields reconstructed by our GCD framework, revealing detailed vortex structures. Columns correspond to samples from (left to right): 2001-02-24 00:00, 2002-01-26 00:00, 2001-01-01 00:00, 2001-01-03 00:00, and 2001-03-16 00:00.
Figure 5: Impact of geographical constraints on generated climate states. The Top Row (without constraints) fails to capture topography-driven phenomena, lacking both the terrain blocking effect over the Andes (red box) and the expected high-pressure system over the Tibetan Plateau (red triangle), while also exhibiting spurious high-pressure artifacts (red ellipses). In contrast, the Bottom Row (with constraints) produces more physically plausible states, showing correct topography-induced wind deceleration, restoring missing pressure features, and eliminating unphysical anomalies.
Figure 6: Impact of cutoff scale of the filter measurement operator ( Afilter ). All evaluations are based on data from the test year 2001. (a) The 99% quantile MAE evaluated at different cutoff scales: raw (representing the native resolution of the GCMs), 2.5°, and 5.0°. (b) PSD comparison among ERA5, MIROC6, and IPSL. The vertical dashed lines indicate the wavenumbers corresponding to the 5.0° and 2.5° scales.
Figure 7: Performance evaluation on the MIROC6 (2001) dataset. The left panel evaluates the impact of DDIM NFE to identify the optimal sweet spot (NFE=25). Based on this optimal trajectory, the right table presents an ablation study of student models distilled down to 6 steps, comparing different distillation losses and guidance strategies (GD vs. CG).
Appendix figures & tables14 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 9
Model Name
Institution
Nom. Res.
Grid Size
IPSL-CM6A-LR
IPSL
∼ 250 km
144 × 143
AWI-ESM-1-1-LR
AWI
∼ 250 km
192 × 96
MPI-ESM1-2-LR
MPI-M
∼ 250 km
192 × 96
MIROC6
JAMSTEC
∼ 140 km
256 × 128
MPI-ESM1-2-HR
MPI-M
∼ 100 km
384 × 192
Appendix
Table 2: Summary of GCMs used in this study.
Figure 8: Visualization of MSL fields across different GCMs before and after downscaling. The time period is consistent with Fig. 4 . Note that unlike surface pressure, which mirrors rugged topography, MSL is a derived variable designed to filter out orographic effects, resulting in a naturally smoother field at synoptic scales, although major features like the Tibetan Plateau and the Andes remain visually discernible. To quantify the variations that are less apparent, we provide a spectral analysis in Fig. 11 to illustrate the differences in small-scale energy.
Figure 9: Visualization of Z250 fields across different GCMs before and after downscaling. The time period is consistent with Fig. 4 . Since upper-level fields are generally smooth, making differences in small-scale details difficult to distinguish visually, we conducted a spectral analysis in Fig. 11 .
Figure 10: Visualization of Z500 fields across different GCMs before and after downscaling. The time period is consistent with Fig. 4 . Since upper-level fields are generally smooth, making differences in small-scale details difficult to distinguish visually, we conducted a spectral analysis in Fig. 11 .
Figure 11: PSD of MSL, Z250, Z500 fields across different GCMs before and after downscaling. The time period is consistent with Fig. 4 .
Figure 12: Ablation study of individual geographical constraints. Both the “Only with DEM” and “Only with LSM” configurations successfully capture the terrain blocking effect over the Andes (red boxes in WS10M) and maintain stable MSL patterns.
Figure 13: Impact of Aup(⋅) in Afilter(⋅) . (a) Bias distributions. (b) Evolution of gradients during diffusion posterior sampling, illustrating the trajectory from timestep t=1000 to 0 for both methods.
Distillation Trajectory
Intermediate Models
MAE
RMSE
25→12→6 (iterative halving)
1
1.2523
1.7133
25→6 (direct, ours)
0
1.2541
1.7135
Appendix
Table 3: 99th-percentile errors on MIROC6 (2001) for two distillation trajectories, both trained with LMSE only and evaluated with CG guidance.
Method
MIROC6
IPSL
AWI
MPI-LR
MPI-HR
Mean field
Bilinear
0.76
0.48
0.68
0.65
0.59
SwinIR
0.69
0.42
0.63
0.58
0.54
ESRGAN
0.67
0.41
0.61
0.56
0.52
CorrDiff
0.62
0.44
0.57
0.53
0.49
CM+NOISE
0.59
0.39
0.55
0.50
0.46
Appendix
Table 4: MAE of the mean field, 50th and 90th percentile. Computed as for the 99th-percentile results in Table 1 .
Method
MIROC6
IPSL
AWI
MPI-LR
MPI-HR
Bilinear (= MAE)
0.76
0.48
0.68
0.65
0.59
SwinIR (= MAE)
0.69
0.42
0.63
0.58
0.54
ESRGAN (= MAE)
0.67
0.41
0.61
0.56
0.52
CorrDiff
0.58
0.42
0.54
0.50
0.46
CM+NOISE
0.54
0.37
0.51
0.47
0.43
GCD (1000 NFE)
0.43
0.34
0.46
0.41
0.37
Appendix
Table 5: Continuous Ranked Probability Score (CRPS) for 10 ensemble members, averaged over variables. ERA5 reanalysis is used as the ground truth reference. For deterministic methods, CRPS is mathematically equivalent to MAE.
Method
MIROC6
IPSL
AWI
MPI-LR
MPI-HR
Bilinear
3.15
3.22
2.95
2.88
2.75
SwinIR
1.85
1.92
1.78
1.72
1.65
ESRGAN
1.82
1.88
1.75
1.68
1.62
CorrDiff
0.95
1.05
0.88
0.85
0.82
CM+NOISE
0.82
0.94
0.75
0.72
0.68
GCD (1000 NFE)
0.18
0.24
0.20
0.17
0.15
Appendix
Table 6: Log-Spectral Distance (LSD) between the radially averaged power spectrum of generated fields and ERA5, computed over the largely unguided high-frequency band ( k≥72 , wavelengths ≲550 km at the equator). Lower is better.
Configuration
MAE
RMSE
Time (s)
Bilinear
1.06
1.78
—
GCD, 6 NFE ( λ=0.1 )
1.49
1.95
4.38
GCD, 6 NFE ( λ=0.5 )
0.98
1.48
4.38
GCD, 6 NFE ( λ=1.0 )
0.92
1.38
4.38
GCD, 15 NFE ( λ=0.1 )
0.89
1.35
10.95
GCD, 1000 NFE
0.87
1.32
589.62
Appendix
Table 7: P99 error on IPSL for different configurations. The 15-step model is obtained by evaluating the continuous-time distilled model with a finer 15-step ODE solver schedule. Time is measured in seconds per sample.
Setting
MIROC6
IPSL
AWI
MPI-LR
MPI-HR
Teacher NFE
15
1.18
1.67
1.25
1.19
1.17
25 (default)
1.05
1.49
1.14
1.07
1.08
50
1.09
1.54
1.17
1.11
1.10
Regularization weight λ
0.01
1.12
1.73
1.21
1.15
1.16
Appendix
Table 8: P99 MAE of the 6-step student for different teacher step budgets and regularization weights λ , on all five GCMs. Defaults are NFE=25 and λ=0.1 .
School of Computer Science and Engineering, University of New South Wales, Sydney, New South Wales, Australia · ARC Centre of Excellence for the Weather of the 21st Century and Climate Change Research Centre, University of New South Wales, Sydney, New South Wales, Australia
Department of Mechanical Engineering, Massachusetts Institute of Technology, Cambridge, MA, USA · Department of Mechanical Engineering, City University of Hong Kong, Kowloon, Hong Kong, China · Google Research, Mountain View, CA, USA +1