Ultra-short-term solar irradiance forecasting from sky images usually follows a two-stage design. First, a video-prediction model generates future sky frames; then, a separately trained regressor maps those frames to an irradiance value. According to a new arXiv preprint, this dominant approach may be throwing away information in the second stage.

The authors propose a single-stage alternative that jointly diffuses sky images and the clear-sky index. Rather than treating frame prediction and irradiance mapping as separate problems, the model learns them together. The paper's title explicitly asks what the frame-to-irradiance stage discards, suggesting that the two-stage pipeline loses details relevant to accurate forecasting.

The abstract is brief and does not include experimental results or comparisons. Still, the conceptual shift is clear: replacing a modular pipeline with a unified generative model could preserve information that would otherwise be lost between stages. The preprint is available on arXiv under identifier 2610.21795.