Real-world single-image super-resolution (SISR) is challenging due to the complex and diverse degradations present in real-world images, which existing degradation models fail to fully capture. This results in training data lacking necessary diversity and richness. Our work addresses this issue by proposing a Multivariate Stochastic Degradation Model (MSDM). We introduce a new type of degradation called color distortion, along with two methods: randomized operation sequence strategy and short circuit strategy. These methods better simulate complex real-world degradation. To further improve the perceptual quality of super-resolution images, we developed a Multi-level Feature Fusion Network (MFFNet) that uses dual attention mechanism and gated feature fusion network. This network achieves better reconstructed image quality. Additionally, we applied MFFNet to a generative adversarial network to build the MFFGAN model. Extensive comparisons show that MFFGAN achieves superior visual compared to previous work on real-world datasets, average 6.86% lower NIQE indicator compare to Real-ESRGAN.