A Scale-Adaptive Framework for Joint Spatiotemporal Super-Resolution with Diffusion Models

2026-04-23 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Environmental Science & Earth Systems · Depth: Advanced, quick

Summary

A new scale-adaptive framework addresses limitations in deep-learning video super-resolution for climate applications, which typically handle either spatial or temporal upscaling but struggle with joint spatiotemporal models across varying super-resolution (SR) factors. This framework reuses a single architecture by decomposing spatiotemporal SR into a deterministic conditional mean prediction using attention and a residual conditional diffusion model. It includes an optional mass-conservation transform to preserve aggregated totals. Scale adaptivity is achieved by retuning three factor-dependent hyperparameters: the diffusion noise schedule amplitude beta, the temporal context length L, and optionally a mass-conservation function f. Demonstrated on reanalysis precipitation over France (Comephore), the architecture supports SR factors from 1 to 25 in space and 1 to 6 in time.

Key takeaway

For AI Scientists developing climate models, this framework offers a reusable architecture for joint spatiotemporal super-resolution across diverse scales. You should consider implementing this scale-adaptive approach to reduce the need for designing distinct models for each SR factor, potentially streamlining development and improving model transferability for precipitation reanalysis.

Key insights

A scale-adaptive framework reuses one architecture for spatiotemporal super-resolution across diverse scaling factors.

Principles

• Larger SR factors increase underdetermination.
• Conditional mean structure remains consistent across factors.

Method

Decomposes spatiotemporal SR into a deterministic conditional mean prediction with attention and a residual conditional diffusion model, optionally applying a mass-conservation transform. Adaptivity is achieved by retuning three hyperparameters.

In practice

• Apply to reanalysis precipitation data.
• Adjust diffusion noise schedule for diversity.
• Maintain comparable attention horizons.

Topics

• Spatiotemporal Super-Resolution
• Diffusion Models
• Scale-Adaptive Framework
• Mass Conservation Transform
• Climate Applications

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Scale Space Diffusion — Highly noisy diffusion states can be processed at lower resolutions without information loss.
• Scalable Deep Learning Framework for Global High-Resolution Land Use Reconstruction — AI4Land uses U-Net on HPC to reconstruct global 1 km land use from coarse data, improving climate model inputs.
• TUDSR: Twice Upsampling-Diffusion for Higher Super-Resolution — TUDSR uses a two-stage, chunk-based diffusion framework to achieve high-quality super-resolution at resolutions up to 2048^2.
• LatticeVision: Image to Image Networks for Modeling Non-Stationary Spatial Data — Image-to-image networks globally estimate non-stationary spatial parameters faster and more accurately than local methods.
• TOC-SR: Task-Optimal Compact diffusion for Image Super Resolution — TOC-SR creates efficient one-step super-resolution models via compact diffusion backbone discovery and distillation.
• AnyBand-Diff: A Unified Remote Sensing Image Generation and Band Repair Framework with Spectral Priors — AnyBand-Diff generates physically consistent remote sensing images by integrating spectral priors and multi-scale physical constraints into a diffusion model.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.