The existence of thin clouds within remote sensing images results in the loss of image information. The removal of thin clouds is crucial for enhancing data quality and increasing the application scope of remote sensing imagery. A method that combines model-driven homomorphic filtering and data-driven diffusion models for thin cloud removal is proposed in this paper. Homomorphic filtering guidance and spatial domain guidance are employed to transform the generation process of a pre-trained unconditional cloud-free remote sensing images generative model from random to directed, thereby generating a cloud-free image that corresponds to the provided thin-cloud image. Unlike most existing data-driven methods, this approach requires only cloud-free images for model training, thus avoiding the difficulties associated with dataset construction. Additionally, the prior knowledge obtained from the diffusion model is used to compensate for the inherent color loss in homomorphic filtering, addressing the limitations of traditional model-based methods. Comparative experiments were conducted by training on 655 cloud-free GaoFen-2 satellite scenes and testing on 68 simulated and 20 real thin-cloud scenes. On the simulated set, the proposed method achieved an average Peak Signal to Noise Ratio (PSNR) of 26.0 dB and Structural Similarity Index (SSIM) of 0.873. On real scenes, it raised image sharpness and color saturation by 22% and 40%, respectively, while preserving ground features more effectively than competing methods. These results demonstrate the effective removal of thin clouds and the superior generalization capability of the proposed approach.