Paper Overview
- Field: Machine Learning
- Authors: Max Defez, Filippo Quarenghi, Mathieu Vrac, Stephan Mandt, Tom Beucler
- Published: 2026-04-23
- arXiv: 2604.21903
- Diffusion noise schedule amplitude β — larger for bigger factors to increase sample diversity.
- Temporal context length L — set to maintain comparable attention ranges across frame rates.
- Mass-conservation function f — tapered to limit extreme amplification for large factors.
Summary
Deep learning video super-resolution has advanced rapidly, but climate applications often require super-resolving only space or time, while joint spatiotemporal models are typically designed for a single pair of super-resolution (SR) factors (the spatial and temporal upsampling ratios between low- and high-resolution sequences), limiting transfer across spatial resolutions and temporal frame rates.
The authors propose a scale-adaptive framework that decomposes spatiotemporal SR into:
1. A deterministic prediction with attention-based conditional means. 2. A residual conditional diffusion model with an optional mass-conservation transformation (ensuring precipitation totals are identical in input and output) to preserve aggregate amounts.
This design allows the same architecture to be reused across different SR factor pairs. The key assumption is that larger SR factors mainly increase underdetermination (thus requiring more context and residual uncertainty) rather than changing the conditional mean structure. Scale adaptivity is achieved by rescaling three factor-dependent hyperparameters:
Results
Demonstrations on Comephore reanalysis precipitation data over France show that a single architecture spans spatial SR factors from 1 to 25 and temporal factors from 1 to 6, producing a reusable architecture and tuning scheme for cross-scale joint spatiotemporal super-resolution.
---
*Auto-collected on 2026-04-27*