PixelFlowCast: Latent-Free Precipitation Nowcasting via Pixel Mean Flows
PixelFlowCast: Latent-Free Precipitation Nowcasting via Pixel Mean Flows
Authors: Yufeng Zhu, Chunlei Shi, Yongchao Feng, Dan Niu Date: 2026-05-11 Paper ID: openalex:2605.10046
Summary
PixelFlowCast is a latent-free, two-stage probabilistic precipitation nowcasting framework designed to overcome the efficiency limitations of diffusion models and the fidelity loss of latent-space models. By employing a deterministic model for global trends and a KANCondNet for spatiotemporal conditioning, it uses a Pixel Mean Flows (PMF) predictor to generate high-quality forecasts in few-step inference. The approach demonstrates superior performance on the SEVIR benchmark, balancing structural fidelity with real-world operational speed requirements.
Key Contributions
- Introduced PixelFlowCast, a two-stage latent-free framework for high-efficiency, high-fidelity precipitation nowcasting.
- Developed Pixel Mean Flows (PMF) to eliminate latent compression, preserving high-frequency details.
- Outperformed state-of-the-art methods on the SEVIR dataset in terms of both prediction accuracy and inference speed.
Open Questions & Future Work
Key Concepts
- pixel-mean-flows-pmf: A latent-free, few-step flow-matching predictor for precipitation nowcasting that directly operates in pixel space.
Archivist Review
I approved the Pixel Mean Flows (PMF) concept because it provides a distinct, high-fidelity alternative to latent-space generative modeling. I rejected the KANCondNet as it functions as a subcomponent, whereas the general KAN architecture is already a concept in the vault. I also included the SEVIR dataset and a pertinent open question regarding physical consistency in latent-free generative models.
Approved Concepts
- Pixel Mean Flows (PMF): Core innovation replacing latent-space compression with a direct pixel-space flow matching approach for high-fidelity precipitation nowcasting.
Approved Open Questions
- Physics-Informed Latent-Free Precipitation Nowcasting: Integrating physical consistency into purely data-driven generative models is essential for improving the reliability of extreme weather nowcasting and ensuring that predictions adhere to the underlying governing dynamics of atmospheric systems.
Rejected Candidates
- [concept] KANCondNet (
kancondnet) - subcomponent_of_broader_mechanism: This is a specific architectural module (a KAN-based encoder) rather than a broadly reusable mechanism; KANs themselves are already in the vault.
Datasets
Links
Metadata & Links
- url
- https://arxiv.org/abs/2605.10046
- paper_id
- 2605.10046
- paper_source
- openalex
- domain
- nlp
- tags
- time-series-forecastingprobabilistic-forecastingspatiotemporal-modelingflow-matchingdiffusion-modelsprecipitation-nowcasting
- architectures
-
- datasets
- sevir
- concept_slugs
- pixel-mean-flows-pmf
- dataset_slugs
- sevir
- skill
- TimeSeriesSkill
- created_at
- 2026-05-14T06:10:20Z