Multiple Conditions-Guided Diffusion Model for Remote Sensing Image Generation

To address the issues of large scenes and detail attributions for generating remote sensing images (RSIs), this study proposes a multiple conditions-guided diffusion (MCGD) model. First, a prior object image control (POIC) module based on multipoint optimization, which models common ground objects, is proposed as a condition of the diffusion model. Then, a caption condition encoding and control (CCEC) module, which mines the caption semantics of the generated image, is designed to construct the semantic space and realize features' cross-modal transformation from text to image. Finally, this study proposes a visual context perception module based on attention, which deeply integrates the conditional features of POIC and CCEC to enhance the fine-grained RSI generation. Experiments show that MCGD can obtain CLIP-scores of 25.23 and 31.09, FID values of 44.26 and 10.42, and IS values of 4.6 and 7.8 on RSICD and NWPU-Captions datasets, respectively, which proves its effectiveness.