Planetary gearboxes are widely used as essential transmission devices in industrial applications, yet they are prone to failures under complex operating conditions and prolonged loads. Traditional fault diagnosis methods heavily rely on expert knowledge and expensive equipment, facing challenges such as data scarcity and low diagnostic efficiency. To address these limitations, the development of generative adversarial networks (GANs) has provided innovative solutions for image generation and data augmentation in recent years. However, existing GAN models often encounter issues such as semantic misalignment and artifacts when processing small-sample datasets, limiting their potential in intelligent fault diagnosis. In this context, this paper proposes a multi-scale attention and progressive feature fusion GAN (MSA-PF-GAN) model, which integrates a progressive decoder structure with multi-scale attention mechanisms to significantly improve image generation quality and fault diagnosis accuracy under small-sample conditions. Experiments conducted on two independent planetary gearbox fault datasets validate the proposed method, showing that it substantially reduces the FID score and enhances diagnostic accuracy (by 35% and 20%, respectively). Across multiple evaluation metrics, the MSA-PF-GAN outperforms other state-of-the-art methods. Further analysis demonstrates that the model, through progressive feature fusion and multi-scale attention mechanisms, excels in generating diverse and realistic images while effectively capturing complex fault features. Therefore, this technique shows promising potential and practical value in the field of planetary gearbox fault diagnosis.