HelioFill. Top: diffusion stage. The masked two-view input, the persistence map P, and the information-age map A are brought to the latent grid and concatenated with the noised latent z_t and the mask (14 channels); a GLA-UNet with constant learned context tokens predicts the noise, and 75 DDIM steps with harmonized reanchoring yield the diffusion estimate. Bottom: refinement stage, a rectified flow that adds texture inside the masked region. Black: inference path; dashed orange: training only; violet: history input.