Just Weather Scoring: Efficient End-to-end Nowcasting with Distributional Diffusion

arXiv:2610.12189 · cs.LG, cs.CV · Submitted 2026-10-08 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "Just Weather Scoring".

Jane: Just Weather Scoring introduces Just Weather Scoring (JWS), a single-stage, end-to-end diffusion model that forecasts directly in radar space,

Tom: First, who's behind it and why it matters.

Paper summary: Tom: So we touched on what Just Weather Scoring is, and now let's look at how they actually set up this system. How do they handle the data flow?

Jane: The methodology starts by conditioning on observations, where they sample a set C of k frame indices randomly without replacement to form clean context where tau equals one.

Meng: And for frames not in that clean context, individual diffusion times are sampled from p(tau). That’s how they adapt the corruption levels to the uncompressed radar data.

Lu: The timestep sampling is pretty smart; they sample centers uniformly tau bar over zero to one, and then draw framewise timesteps astilde tau t like a normal distribution with s bar squared.

Tom: They use that spatiotemporal resolution-aware timestep sampling in equation four to adapt the corruption levels specifically for uncompressed radar data.

Jane: Then they predict the clean sequence x hat one as f theta of x tau t, and they compute the velocity-space loss through conversion using v theta t equal to (x hat one t minus x tau t) over (one minus tau t).

Meng: It’s a fairly standard setup for diffusion models, but they've baked in the radar space directly into the prediction target.

Lu: The architecture uses a standard transformer backbone but employs an HDiT-style backbone with spatial attention in high-resolution layers and factorized spatiotemporal attention only in the lower-resolution bottleneck after an additional downsampling step.

Tom: That means they're using spatial attention where it matters most, and then doing some spatiotemporal work later in the network.

Jane: The simple distributional objective they use is defined as L simple of tau t equals one minus tau t CRPS plus tau t MSE. It’s a hyperparameter-free objective that linearly interpolates distribution CRPS and squared error.

Meng: That specific loss function is what allows them to train the model without needing to manually tune tons of parameters for the objective itself.

Lu: This whole setup is more efficient and capable than pipelines with multiple trained components, like those that use latent space methods first training a compression model and then learning denoising on compressed representations.

Tom: So, they’re proving that this end-to-end approach avoids the complexity and uncertainty introduced by those multi-stage training pipelines.

Jane: It really shows how moving directly to radar space modeling greatly simplifies both the training process and the inference process.

Conclusion: Tom: So wrapping up, we're talking about Just Weather Scoring: Efficient End-to-end Nowcasting with Distributional Diffusion. The title itself suggests efficiency is a big focus here.

Jane: And the authors, Jannik Wiese and Johannes Schusterbauer are the ones who put this end-to-end model together. They tackle the problem of making these forecasts reliable by representing all possible outcomes instead of just picking one guess.

Lu: The implication is that for meteorological practitioners, this could be very accessible because it simplifies training and inference significantly.

Meng: It means they can get probabilistic nowcasting with a small footprint, which is something that matters when you have limited computational budget at inference time.

Tom: And the fact that even the smallest model remains competitive using substantially fewer parameters and over seventeen times faster inference speed is pretty compelling.

Jane: So, this paper gives us a simple generative model that doesn't rely on external compression models, and by training with a distributional objective they unlock few-step sampling at inference time to drastically reduce latency.

Lu: Achieving state-of-the-art probabilistic nowcasting performance on the SEVIR and MeteoNet benchmarks is the main result they claim in this work.

Tom: It’s a scalable approach with respect to model size and available computational budget at inference time, and efficient settings perform competitively.

Jane: I think the real impact here is how much it simplifies things for people who deal with actual weather forecasting every day.

Lu: It really makes probabilistic nowcasting more practical by making it easier to adopt because of the efficiency gains in both training and deployment.

Jannik Wiese, Johannes Schusterbauer, Tommaso Martorella

CompVis @ LMU Munich, Munich Center for Machine Learning (MCML)

cs.LG, cs.CV

Submitted: 2026-10-08

Updated: 2026-10-08

Comments: Project Page: https://compvis.github.io/jws

Code: https://github.com/CompVis/jws

Project page: https://compvis.github.io/jws

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 92/100

The gist: Just Weather Scoring introduces Just Weather Scoring (JWS), a single-stage, end-to-end diffusion model that forecasts directly in radar space, addressing issues with separate components and lossy

Key concepts

End-to-end training
JWS is trained directly on raw sensor data without needing separate models for compression or deterministic forecasting. This simplifies the overall system, making both the training process and the final prediction much more straightforward compared to older methods that required multiple specialized components.
Masked Asynchronous Diffusion (MAD)
MAD is a sampling technique used within JWS that keeps important, clean parts of the radar data intact while adapting how noise is added. It uses spatiotemporal awareness to adjust the level of corruption based on high-dimensional radar data, preserving context effectively during training.
Simple distributional objective
This is a specific loss function used during training that combines Distributional Cross-Entropy (CRPS) and Mean Squared Error (MSE). This objective guides the model to learn a probabilistic forecast distribution, which is crucial for accurate weather prediction and enables efficient few-step generation at inference.

Terminology

Summary

Just Weather Scoring introduces Just Weather Scoring (JWS), a single-stage, end-to-end diffusion model that forecasts directly in radar space, addressing issues with separate components and lossy compression by combining Masked Asynchronous Diffusion with a simple scoring-rule objective The gist is that JWS achieves state-of-the-art probabilistic forecasting performance at reduced training and inference cost by operating directly on uncompressed data

How it works

JWS avoids additional stages and operates directly on uncompressed data, enabling forecasting with a single transformer parametrized to predict clean radar maps (Li & He, 2025) <ref:2610.12189#pg4> The model combines Masked Asynchronous Diffusion (MAD), which preserves clean context while sampling correlated, framewise noise levels for predicted timesteps, with a simple scoring-rule objective that aligns training with probabilistic forecasting and unlocks few-step generation <ref:2610.12189#pg4> This approach is more efficient and capable than pipelines with multiple trained components <ref:2610.12189#pg3>

Key Components and Innovations

The paper introduces several key innovations to achieve its efficiency and performance goals. These include:

  1. End-to-end training, where JWS is trained directly on sensor data without separately trained compression or deterministic forecasting stages, which simplifies training and inference <ref:2610.12189#pg3>

  2. Masked Asynchronous Diffusion (MAD), a timestep-sampling scheme that preserves clean context while adapting diffusion training to high-dimensional spatio-temporal data <ref:2610.12189#pg4> MAD uses spatiotemporal resolution-aware timestep sampling in Eq. (4) to adapt corruption levels to uncompressed radar data <ref:2610.12189#pg6>

  3. Simple distributional objective, which involves simple Distributional Diffusion Models (sDDM) trained with a hyperparameter-free objective, linearly interpolating distributional CRPS and squared error <ref:2610.12189#pg6> This objective is defined as L simple(tau t) = (1-tau t)CRPS + tau t MSE <ref:2610.12189#pg6>

Methodology Details

The methodology involves several specific procedural steps for training and inference.

- Conditioning on observations:

Forecasting inherently depends on available past observations, and the paper samples a set C of k frame indices randomly without replacement to form clean context where tau t = 1 <ref:2610.12189#pg4> With probability pU, the model is trained for unconditional generation with C = ∅ <ref:2610.12189#pg4> For frames not in C, individual diffusion times are sampled from p(τ) <ref:2610.12189#pg4>

- Timestep sampling:

The paper samples centers uniformly τ¯ ∼ U(0, 1), and then draws framewise timesteps astilde tau t∼ N(bar tau,s bar 2), where s bar is calculated based on the sequence-level center <ref:2610.12189#pg6> Dimensionality-aware corruption is applied using Eq. (4) to obtain the final timesteps τ t <ref:2610.12189#pg6>

- Prediction target:

The model predicts the clean sequence xˆ1 = fθ(xτ, τ) and computes the velocity-space loss (Eq. 2) through conversion, v theta t= (xˆ1 t - x tau t)/(1-tau t) <ref:2610.12189#pg6>

- Architecture:

JWS uses standard transformers but employs an HDiT-style backbone (Fig. 4), with spatial attention in high-resolution layers and factorized spatiotemporal attention only in the lower-resolution bottleneck after an additional downsampling step <ref:2610.12189#pg6>

Evaluation and Results

The model is evaluated on SEVIR and MeteoNet benchmarks, achieving state-of-the-art probabilistic forecasting performance at reduced training and inference cost <ref:2610.12189#pg2> The results show that even the smallest model remains competitive using substantially fewer parameters and more than 17× faster inference <ref:2610.12189#pg2> JWS achieves superior CRPS, SSIM, HSS, and RI on SEVIR compared to previous methods without classifier-free guidance (CFG) <ref:2610.12189#pg9> Furthermore, JWS-T performs competitively with much larger models and enables probabilistic nowcasting with a small footprint <ref:2610.12189#pg10> The model demonstrates superior zero-shot extrapolation performance to 180 min into the future on SEVIR-Long <ref:2610.12189#pg11>

Conclusion

JWS is a simple end-to-end generative model that does not rely on compression models, and by training with a distributional objective unlocks few-step sampling at inference time, drastically reducing latency <ref:2610.12189#pg11> Our approach achieves state-of-the-art probabilistic nowcasting performance on the SEVIR and MeteoNet benchmarks <ref:2610.12189#pg2> JWS is scalable with respect to the model size and available computational budget at inference time, while efficient settings perform competitively <ref:2610.12189#pg2> By greatly simplifying training and inference, we hope that JWS can be easily adopted by meteorological practitioners <ref:2610.12189#pg11> The approach is scalable with respect to the model size and available computational budget at inference time, while efficient settings perform competitively <ref:2610.12189#pg2>

How it works

The paper introduces several key innovations to achieve its efficiency and performance goals. These include:

  1. End-to-end training, where JWS is trained directly on sensor data without separately trained compression or deterministic forecasting stages, which simplifies training and inference <ref:2610.12189#pg3>

  2. Masked Asynchronous Diffusion (MAD), a timestep-sampling scheme that preserves clean context while sampling correlated, framewise noise levels for predicted timesteps <ref:2610.12189#pg4> MAD uses spatiotemporal resolution-aware timestep sampling in Eq. (4) to adapt corruption levels to uncompressed radar data <ref:2610.12189#pg6>

  3. Simple distributional objective, which involves simple Distributional Diffusion Models (sDDM) trained with a hyperparameter-free objective, linearly interpolating distributional CRPS and squared error <ref:2610.12189#pg6> This objective is defined as L simple(tau t) = (1-tau t)CRPS + tau t MSE <ref:2610.12189#pg6>

Conclusion

JWS is a simple end-to-end generative model that does not rely on compression models, and by training with a distributional objective unlocks few-step sampling at inference time, drastically reducing latency <ref:2610.12189#pg11> Our approach achieves state-of-the-art probabilistic nowcasting performance on the SEVIR and MeteoNet benchmarks <ref:2610.12189#pg2> JWS is scalable with respect to the model size and available computational budget at inference time, while efficient settings perform competitively <ref:2610.12189#pg2> By greatly simplifying training and inference, we hope that JWS can be easily adopted by meteorological practitioners <ref:2610.12189#pg11>

How it works

The paper introduces several key innovations to achieve its efficiency and performance goals. These include:

  1. End-to-end training, where JWS is trained directly on sensor data without separately trained compression or deterministic forecasting stages, which simplifies training and inference <ref:2610.12189#pg3>

Improvements for AI systems

  1. Bold header: End-to-end Pixel-Space Nowcasting

This system can perform forecasting directly in radar space and enabling few-step generation, simplifying training by eliminating separate compression or deterministic forecasting components, as stated in the abstract.

  1. Bold header: Masked Asynchronous Diffusion (MAD) for Flexible Conditioning

The model can be conditioned on observed past observations and simultaneously sample individual diffusion times for unobserved frames, which permits forecasting by conditioning on past observations (C = 1,..., C), infilling by excluding missing frames from C, and conditioning on corrupted data by setting τt ≠ 1 to express confidence in individual frames.

  1. Bold header: Simple Distributional Objective (sDDM) for Efficient Few-Step Generation

The system can enable few-step inference by being trained with a hyperparameter-free objective, specifically the loss function L simple(τt) = (1-τt)CRPS + τtMSE, which interpolates between distributional supervision and regression.

  1. Bold header: Spatiotemporal Resolution-Aware Timestep Sampling

The model can adapt corruption levels to high-dimensional data by using a mechanism that explicitly accounts for both spatial and temporal resolution through the formula in Eq. (4), leading to superior performance compared to rescaled logit-normal sampling.

  1. Bold header: Scalable Architecture via Hierarchical Transformers

The system can maintain scalability with respect to model size by using an HDiT-style (Crowson et al., 2024) backbone, which uses spatial attention in high-resolution layers and factorized spatiotemporal attention only in the lower-resolution bottleneck.

  1. Bold header: Robust Calibration via Ensemble Variance

The system can provide reliable probabilistic forecasts by having ensembling quantify uncertainty directly in radar-space, as evidenced by the fact that JWS maintains better calibration and outperforms FlowCast on CRPS at the more challenging long-lead times.

Abstract

Generative diffusion models are well-suited for probabilistic precipitation nowcasting, but existing approaches often rely on separately trained compression or deterministic forecasting components and remain costly at inference due to iterative denoising. We introduce Just Weather Scoring (JWS), a single-stage, end-to-end diffusion model which addresses both issues by forecasting directly in radar space and enabling few-step generation. Radar-space modeling greatly simplifies training and inference and eliminates uncertainty arising from lossy compression. JWS combines Masked Asynchronous Diffusion, a timestep-sampling scheme that preserves clean context while adapting diffusion training to high-dimensional spatio-temporal data, with a simple scoring-rule objective that aligns training with probabilistic forecasting and unlocks few-step generation. On the SEVIR and MeteoNet benchmarks, JWS achieves state-of-the-art probabilistic forecasting performance at reduced training and inference cost. Even our smallest model remains competitive using substantially fewer parameters and more than 17x faster inference.

Sources

Related papers