Music generation has become a key area in artificial intelligence, achieving significant progress in recent years. However, current research focuses primarily on general music tasks, with limited support for ethnic music. Moreover, the lack of multimodal guidance, such as text and image inputs, restricts generative models in understanding complex semantics and producing high-quality music. To address these limitations, we propose MusDiff, a multimodal music generation framework that combines text and image inputs to enhance music quality and cross-modal consistency. MusDiff is based on a diffusion model architecture, integrating IP-Adapter and KAN (Kolmogorov–Arnold Network) optimizations to improve feature fusion and modality alignment. Additionally, we introduce a new multimodal dataset, MusiTextImg, which includes diverse music categories, such as ethnic and modern styles, with annotations for text, image, and music modalities. We also extend the MusicCaps dataset by adding matched image pairs to text descriptions, further supporting multimodal research. Experimental results demonstrate that MusDiff outperforms existing methods on benchmark datasets (MusiTextImg and MusicCaps), excelling in realism, detail fidelity, and multimodal alignment. MusDiff not only sets a new performance standard for multimodal music generation but also opens new research directions in the field of multimodal generation.