CRS-Diff: A Controllable Generative Remote Sensing Foundation Model CRS-Diff is a new remote sensing (RS) generative foundation model that leverages diffusion models and integrates advanced control mechanisms to enable precise image generation. It supports text, metadata, and image condition inputs, allowing for more accurate and stable generation of RS images. The model introduces a new conditional control mechanism to achieve multi-scale feature fusion, enhancing the guiding effect of control conditions. Experimental results show that CRS-Diff outperforms previous methods in generating RS images both quantitatively and qualitatively. Additionally, CRS-Diff can serve as a data engine to generate high-quality training data for downstream tasks like road extraction. The model is based on the Stable Diffusion (SD) framework, with ControlNet integrated to include additional control signals for RS image generation. These signals adjust global and local condition information, including six additional image control conditions and textual conditions. The optional combination of multiple conditions ensures that the resulting RS images are visually realistic and accurately reflect specific geographic and temporal information. The contributions of CRS-Diff include proposing a new controllable generative RS foundation model that supports multiple types of controllable conditions. It also introduces a new conditional control mechanism to achieve multi-scale feature fusion, enhancing the guiding effect of control conditions. Experimental results demonstrate the superiority of CRS-Diff in generating RS imagery that adheres to specific conditions and surpasses previous methods both quantitatively and qualitatively. Additionally, CRS-Diff can serve as a data engine to generate high-quality training data for downstream tasks. The model is evaluated using four metrics: Inception Score (IS), Fréchet Inception Distance (FID), CLIP Score, and Overall Accuracy (OA). The results show that CRS-Diff outperforms existing methods in three metrics and achieves second place in the Inception Score. The model also excels in controllability and generation quality, meeting the demands of practical applications like urban planning. CRS-Diff is tested under multiple conditions, but conflicts may arise among these conditions. The model is designed to minimize conflicts between conditions, ensuring that the generated images are accurate and realistic. The model is also applied to downstream tasks like road detection, where it demonstrates excellent performance in generating training data that supports the task. The results show that synthetic training datasets can achieve almost the same performance as real datasets, indicating that CRS-Diff can simulate real images. By adding synthetic datasets to real datasets, the detection performance can be significantly improved, demonstrating the benefits of generated RS images for downstream tasks.CRS-Diff: A Controllable Generative Remote Sensing Foundation Model CRS-Diff is a new remote sensing (RS) generative foundation model that leverages diffusion models and integrates advanced control mechanisms to enable precise image generation. It supports text, metadata, and image condition inputs, allowing for more accurate and stable generation of RS images. The model introduces a new conditional control mechanism to achieve multi-scale feature fusion, enhancing the guiding effect of control conditions. Experimental results show that CRS-Diff outperforms previous methods in generating RS images both quantitatively and qualitatively. Additionally, CRS-Diff can serve as a data engine to generate high-quality training data for downstream tasks like road extraction. The model is based on the Stable Diffusion (SD) framework, with ControlNet integrated to include additional control signals for RS image generation. These signals adjust global and local condition information, including six additional image control conditions and textual conditions. The optional combination of multiple conditions ensures that the resulting RS images are visually realistic and accurately reflect specific geographic and temporal information. The contributions of CRS-Diff include proposing a new controllable generative RS foundation model that supports multiple types of controllable conditions. It also introduces a new conditional control mechanism to achieve multi-scale feature fusion, enhancing the guiding effect of control conditions. Experimental results demonstrate the superiority of CRS-Diff in generating RS imagery that adheres to specific conditions and surpasses previous methods both quantitatively and qualitatively. Additionally, CRS-Diff can serve as a data engine to generate high-quality training data for downstream tasks. The model is evaluated using four metrics: Inception Score (IS), Fréchet Inception Distance (FID), CLIP Score, and Overall Accuracy (OA). The results show that CRS-Diff outperforms existing methods in three metrics and achieves second place in the Inception Score. The model also excels in controllability and generation quality, meeting the demands of practical applications like urban planning. CRS-Diff is tested under multiple conditions, but conflicts may arise among these conditions. The model is designed to minimize conflicts between conditions, ensuring that the generated images are accurate and realistic. The model is also applied to downstream tasks like road detection, where it demonstrates excellent performance in generating training data that supports the task. The results show that synthetic training datasets can achieve almost the same performance as real datasets, indicating that CRS-Diff can simulate real images. By adding synthetic datasets to real datasets, the detection performance can be significantly improved, demonstrating the benefits of generated RS images for downstream tasks.