Significance: End-to-end semantic communication leverages deep learning models to extract semantic features from data, enabling intent-driven communication processes that significantly enhance transmission efficiency. However, existing semantic communication paradigms based on discriminative models employ symbol-level rate-distortion optimization and perform maximum likelihood estimation solely based on received signals, failing to satisfy the perceptual requirements of users. To ensure the visual quality of transmitted data, a generative visual semantic communication paradigm has emerged, which adopts a rate-distortion-perception optimization framework to achieve alignment between data transmission and human perception through maximum a posteriori estimation. Diffusion models are advantageous for controlling visual generation and have thus become essential tools for this generative paradigm. Nevertheless, systematic organization of the technical roadmaps for empowering semantic communication using diffusion models is lacking in current research. Progress: This study addresses this gap by modeling the communication process as a mathematical inverse problem and elucidating the general methodology by which diffusion models solve data compression and transmission challenges through posterior sampling. The fundamental concepts, mathematical formulations, and sampling strategies underpinning diffusion models are systematically introduced. In addition, the general methods and key technologies employed for diffusion model-enabled generative compression and transmission are comprehensively reviewed from an inverse problem-solving perspective. Moreover, the performance metrics commonly used for objective assessment of the visual quality of transmitted data are summarized to provide a comprehensive evaluation framework. The core methodology demonstrates that generalized communication processes can be effectively modeled as inverse problems. The approach involves inferring the source data distribution using maximum a posteriori estimation based on channel measurements and forward operators composed of various signal processing operations. Through diffusion posterior sampling, diffusion models solve these communication inverse problems via a three-step process: first, pre-training diffusion models from large-scale datasets are used to obtain diffusion priors; second, joint source-channel codecs are used to mitigate channel distortions in visual data transmission and construct proximal regularization terms; finally, measurement regularization terms are constructed based on channel measurements. By integrating these regularization terms for posterior estimation and distribution sampling, diffusion models can implicitly reconstruct source data through gradient descent, effectively overcoming transmission challenges caused by strong channel noise, nonlinear operators, and time-varying channel conditions. Conclusions and Prospects: The analysis reveals that compared to visual semantic communication approaches based on discriminative deep learning models, the generative visual semantic communication paradigm based on diffusion models can significantly improve transmission efficiency and resilience while ensuring perceptual quality and semantic consistency of visual information. This advancement represents a fundamental shift toward communication systems that prioritize human perceptual requirements alongside traditional distortion metrics. Open issues, including image realism modeling and acceleration of diffusion model sampling, are discussed. The report highlights the effectiveness of conditional diffusion models for enabling existing semantic communication architectures to recover sources at the receiver based on minimal tokens and highly degraded measurements, offering an intelligent and concise design philosophy for future generative visual semantic communication systems.