PixelFlowCast: Latent-Free Precipitation Nowcasting via Pixel Mean Flows

Abstract: Precipitation nowcasting aims to forecast short-term radar echo sequences for extreme weather warning, where both prediction fidelity and inference efficiency are critical for real-world deployment. However, diffusion-based models, despite their strong generative capability, suffer from slow inference due to multi-step sampling trajectories, limiting their practical usability. Conditional Flow Matching (CFM) improves efficiency via straightened trajectories, but relies on latent space compression, which inevitably discards high-frequency physical details and degrades fine-grained prediction quality. To address these limitations, we propose PixelFlowCast, a two-stage probabilistic forecasting framework that achieves both high-efficiency and high-fidelity prediction without latent compression. Specifically, in the first stage, a deterministic model first produces coarse forecasts to capture global evolution trends. In the subsequent stage, the proposed KANCondNet extracts deep spatiotemporal evolution features to provide accurate conditional guidance. Based on this, a latent-free, few-step Pixel Mean Flows (PMF) predictor employs an $x$-prediction mechanism to generate high-quality predictions, effectively preserving fine-grained structures while maintaining fast inference. Experiments on the publicly available SEVIR dataset demonstrate that PixelFlowCast outperforms existing mainstream methods in both prediction accuracy and inference efficiency, particularly for long sequence forecasting, highlighting its strong potential for real-world operational deployment.

• The paper introduces a two-stage autoregressive model that combines a deterministic coarse prediction with a conditional pixel mean flows refinement to enhance forecasting accuracy.
• The methodology eschews latent space compression, thereby preserving high-frequency meteorological details while achieving significant improvements in global and perceptual metrics on the SEVIR VIL benchmark.
• The model achieves operational efficiency with inference times under 0.6 seconds per sample, outperforming existing deterministic and probabilistic baselines in both accuracy and real-time performance.

PixelFlowCast: A Latent-Free, Pixel Mean Flows Approach for Precipitation Nowcasting

Context and Motivation

Short-term precipitation nowcasting is a critical computational task in meteorology, directly impacting real-world early warning systems for extreme weather phenomena. Conventional deterministic deep learning models for radar echo prediction generally optimize for global spatiotemporal coherence but tend to yield overly smoothed outputs under MSE-driven objectives, leading to significant loss of fine-grained physical details, particularly over extended forecast horizons. Recent advances in probabilistic generative modeling—especially diffusion models—have elevated fidelity in nowcasting but remain hampered by slow, multi-step sampling and, when combined with latent space compression, suffer degradation in capturing high-frequency meteorological signals. Conditional Flow Matching (CFM) accelerates sampling via straightened trajectories, yet its dependency on latent compression introduces further detail loss, and direct pixel-space flow regression becomes computationally intractable.

Methodology

PixelFlowCast introduces a two-stage, autoregressive generative framework that eschews all forms of latent space compression, operating exclusively in the pixel space for both conditioning and generation. The architecture consists of:

1. Deterministic Coarse Prediction: The first stage leverages a deterministic backbone (instantiated as SimVP) to rapidly forecast macroscopic spatiotemporal trends, providing a physically plausible baseline ($X_{\text{coarse}}$).
2. Conditional Pixel Mean Flows Refinement: The second stage refines the coarse baseline by modeling the high-frequency residual ($X_{\text{res}}$), crucial for realistic extreme event reproduction. This is accomplished by:

• KANCondNet: A conditional encoder built on Kolmogorov-Arnold Networks (KANs) with learnable spline activations, extracting deep spatiotemporal features from $(X_{\text{past}}, X_{\text{coarse}})$. KANResNetBlocks are confined to deep feature levels to avoid overfitting on high-frequency noise and retain generalization.
• Pixel Mean Flows (PMF) Predictor: A latent-free, x-prediction generative module outputs residuals directly in pixel space via a few-step accelerated trajectory, bypassing the need to regress chaotic velocity fields. By structuring the forward process as a simple linear path between noisy and denoised states and utilizing an auxiliary, trainable time-step parameter, the model circumvents optimization collapse in high dimensions and achieves both detail preservation and operational efficiency.

1. Autoregressive Extension: Coarse and refined predictions are rolled out via a sliding temporal window to enable long-range sequence prediction, with the model's receptive field consistently updated.

Optimization is performed by joint minimization of the MSE loss on the coarse stage and an adapted PMF velocity matching loss in the generative refinement stage. Inference is executed with few-step Euler integration, continually extracting the noise-free x-prediction to evade accumulation of numerical errors that plague high-dimensional integration in pixel space.

Empirical Evaluation

Experiments utilize the SEVIR VIL benchmark, covering spatiotemporal radar echo sequences over a 384 km × 384 km domain. Evaluation focuses on both global metrics (CSI, HSS across multiple VIL thresholds and pooled scales) and perceptual metrics (LPIPS, SSIM), with thorough measurement of inference time for real-time deployment feasibility.

Key Results:

• Accuracy and High-Threshold Sensitivity: PixelFlowCast surpasses all deterministic (SimVP, Earthformer, U-Net) and probabilistic (PreDiff, DiffCast, FlowCast) baselines in critical indices. Notably, it achieves an overall CSI of 0.2246 and HSS of 0.2867 on 36-frame, 3-hour forecasts—delivering 8–14% relative gains over the strongest deterministic contenders and much wider margins against pure generative baselines, particularly under high precipitation thresholds (e.g., for VIL ≥ 133, CSI 0.1541 vs. DiffCast 0.1320).
• Visual Quality and Structure Preservation: The model attains best-in-class LPIPS (0.3241) and SSIM (0.6106), maintaining sharp boundaries and realistic spatial structures at extended lead times, as corroborated by qualitative visualizations over 180-minute prediction horizons.
• Inference Efficiency: PixelFlowCast achieves a per-sample inference time of 0.589s for a full 3-hour forecast, outperforming PreDiff (112.34s), DiffCast (11.56s), and FlowCast (0.788s), and setting a new Pareto frontier in the fidelity-vs-efficiency space for operational nowcasting.

Ablation and Analysis

• Generative Paradigm: Direct CFM in pixel space without latent compression is suboptimal due to optimization collapse; latent-space CFM recovers some performance at the cost of detail. The PMF x-prediction approach in original pixel space achieves the highest conditional skill, confirming that bypassing latent dimensionality is superior when paired with an appropriate generative formulation.
• KANCondNet Design: KAN-based conditional encoding is shown to be critical for extreme event forecasting. Deploying KANResNetBlocks solely at deep layers (Deep-KAN) outperforms both full-KAN deployments (risk of capacity overfitting) and traditional ResNet- or GlobalNet-based conditioning, especially at high VIL thresholds.
• Few-Step Sampling and Noise-Free Extraction: Optimal fidelity–efficiency trade-off is reached with 10 integration steps; further increases yield diminishing returns due to numerical smoothing. Noise-free extraction at the terminal step avoids residual integration errors, enhancing extreme precipitation core preservation.

Implications and Future Developments

PixelFlowCast demonstrates that high-fidelity, real-time meteorological nowcasting is achievable via latent-free flow-based generative modeling, provided that (i) coarse physical evolution is separated from fine stochastic detail, and (ii) conditional guidance is sufficiently nonlinear and well-regularized. By fully dispensing with latent compression, it provides a pathway for domain-specific, large-scale sequence generation tasks that require both statistical sharpness and operational tractability.

Future progress could harness explicit integration of atmospheric physics (e.g., physics-informed neural priors, constraints from conservation laws or governing PDEs, as in [38, 39, 40]) into the generative refinement step. Hybridization of KAN capacity with physical regularization stands to further close the gap between data-driven fidelity and physical consistency, potentially enabling robust generalization to even rarer or more chaotic meteorological phenomena.

Conclusion

PixelFlowCast establishes a new state-of-the-art in precipitation nowcasting by combining deterministic backbone forecasting, KAN-based conditional encoding, and accelerated, latent-free pixel mean flows. It achieves a stringent balance between predictive accuracy, high-resolution structure, and real-time inference, paving the way for both scalable operational deployment and further research in conditional generative modeling of high-dimensional, multiscale spatiotemporal processes.

Reference: "PixelFlowCast: Latent-Free Precipitation Nowcasting via Pixel Mean Flows" (2605.10046)