Answers are grounded in this paper and cite their sources.
Cross-modal remote sensing image generation, the task of translating among Red, Green, Blue (RGB), Synthetic Aperture Radar (SAR), and PANchromatic (PAN) imagery, is essential for filling sensor gaps in time-critical Earth observation tasks. A fundamental challenge is that different land-cover types follow mutually contradictory physical mapping rules. While the smoothing needed for water bodies conflicts with the edge preservation required for urban structures, existing generators apply identical parameters to every spatial location, forcing a single network to reconcile these conflicting objectives. We introduce RS-MoEGAN, which integrates sparse Mixture-of-Experts (MoE) into a Vision Transformer generator. A learned router assigns each image token to a specialized expert, so that incompatible generation rules are handled by disjoint parameter subsets. Only approximately 10% of the total parameters are activated per forward pass, keeping FLOPs on par with a standard dense ViT. Experiments on two public multimodal datasets show that RS-MoEGAN achieves competitive FID scores and visual quality across multiple translation directions. Expert routing visualizations suggest that the router separates tokens according to semantic content, implicitly grouping semantically distinct regions into different expert subsets.
No takes yet. Share an insight, caveat, or question.
Fan et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: