Answers are grounded in this paper and cite their sources.
Text-guided remote sensing image (RSI) generation often struggles to generate layouts that strictly conform to text prompts. While existing layout-guided methods introduce additional images, such as geometric boundary maps (GBM) or real images, as reference layouts, their suboptimal layout constraint mechanisms frequently interfere with the semantic guidance of the text prompts. To address this, we propose a novel RSI generation framework that incorporates a dynamic self-attention guidance strategy. Motivated by the insight that self-attention maps inherently capture the spatial relationships between visual units during the generation process, we leverage the self-attention maps extracted from the reconstruction process of a real image as robust layout constraints. Unlike existing methods that apply constraints throughout the entire generation process, often disturbing the semantic guidance of the text prompt, we introduce a dynamic guiding strategy. This strategy injects these self-attention maps only during critical intermediate generation steps, which effectively decouples the layout guidance of the self-attention maps from the semantic guidance of the text prompt. Our method ensures that the layout of the newly generated image is constrained by the real image, while the semantic content is freely determined by the text prompt through sufficient semantic interaction. Experiments demonstrate that our method outperforms baselines in terms of layout consistency and text-guided scene diversity, and effectively improves the performance of downstream tasks, such as scene classification.
No takes yet. Share an insight, caveat, or question.
Shi et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: