<p>Pest control is crucial for ensuring the stability of agricultural production and increasing crop yields. Despite the prominent success of deep learning-based real-time object detection technology, Obtaining high-quality training data with various types and forms of pests presents significant challenges. The complex agricultural environment also requires detection models to have high generalization capabilities. To address these issues, we propose a Semantic Integration Guidance, a control network within the diffusion model, which integrates multi-source semantic features to control image generation, ensuring that synthesized samples adhere to real-world attributes. To enhance the generalization performance of real-time models, we designed a Multi-Level Alignment Distillation (MLAD) framework. The MLAD framework optimizes the training process for downstream detection tasks by aligning the intermediate feature representations and softened logits outputs of the teacher and student models, as well as constructing relational graphs for feature channels. Unlike existing methods that modify detection model architectures by introducing additional modules, our approach directly improves data diversity and knowledge distillation, without altering the structure of the detection models. Our experiments on the IP102 dataset show that our method achieves an average FID score of 8.3, an IS of 2.1, and a 17% improvement in AP, highlighting both high-quality sample generation and significant performance gains in detection tasks.</p>