GeoRID: Role-Specialized Instruction Decomposition with Compact Visual Tokens for Complex-Instruction Remote Sensing Image Synthesis

Abstract

Environmental monitoring, urban-growth rehearsal, and disaster-scenario simulation often need a satellite or aerial scene drawn from a compound geospatial instruction: object counts, relative bearings, land-cover mosaics, and a stated ground sample distance (GSD). Existing remote sensing text-to-image (T2I) models flatten that instruction into a single CLIP sentence and denoise every latent token uniformly, so quantity words and GSD cues are washed out by scene nouns, and homogeneous farmland or water wastes self-attention. We present GeoRID, a latentdiffusion framework that factors the instruction and the compute budget along geospatial axes. Role-specialized decomposition agents (RSDA) parse layout, cover, and scale slots and gate the three conditions into UNet cross-attention. Geo-aware compact token compression (GCTC) merges only homogeneous land-cover tokens, keeping tanks, ships, vehicles, and runway edges. Object–scene dual guidance (OSDG) applies a time-varying pair of classifier-free scales so coarse partitions form early and countable objects lock in late. On the RSICD protocol used by recent diffusion systems, GeoRID reaches FID 36.82 and zero-shot classification OA 75.34 versus 41.14 / 72.54 for Any2RSI. Gains are larger on a 400-prompt count-and-bearing subset (count match 0.565 vs. 0.428), and the default merge ratio ρ=0.35 cuts UNet FLOPs by about 30% while still improving FID.