Text-to-image generation techniques have made significant progress driven by the development of diffusion model, but there are still challenges in achieving diverse image generation and semantic consistency under few-shot conditions. In this paper, we propose a lightweight multi-strategy fusion method for few-shot image generation, called LMSGen, which significantly improves the diversity and semantic consistency performance of the generated images through implicit scene and perspective Token learning, local detail enhancement mechanism, and Token fusion strategy in the inference stage. In addition, we design a regularization dataset covering multiple scenes and perspectives for each image class, which provides a rich visual context for the model. Experimental results demonstrate that our method achieves favorable performance across key metrics compared to representative baselines. This work offers a new perspective and solution for controllable image generation under low-resource conditions.