Human-Centric Content Generation with Diffusion Models: A Survey

Human-centric content generation, where faces, bodies, and their behaviors are the central subjects, has become a pivotal direction in the era of AI-generated content (AIGC). It encompasses a broad spectrum of tasks, ranging from appearance-level editing to motion-driven synthesis, enabling diverse real-world applications. Compared to generic content generation, human-centric generation poses stricter requirements, demanding not only high visual fidelity but also accurate identity preservation, fine-grained controllability, temporal and cross-view consistency, and strong generalization across scenarios. Diffusion models provide a powerful foundation for addressing these challenges. Their probabilistic denoising formulation enables stable optimization, broad mode coverage, and flexible conditioning mechanisms, making them particularly well-suited for human-centric modeling. In this survey, we organize diffusion-based human-centric generation through a two-axis perspective that captures both task-level categorization and technical analysis. From the task axis, we categorize generation problems by the nature of the human subject and its dynamics, encompassing facial and full-body content, appearance-oriented and motion-driven synthesis, as well as the driving modalities that condition generation. From the technical axis, we analyze how these tasks are addressed by diffusion models, examining core requirements-accuracy, controllability, consistency, and generalization-and contrasting models trained from scratch with approaches adapted from pretrained models. Beyond this taxonomy, we review the foundational components underlying these tasks and conclude by summarizing open challenges and future research directions. Overall, this survey provides a unified and task-grounded roadmap for diffusion-based human-centric generation, offering a structured lens to understand existing methods, identify design trade-offs, and guide future research in this rapidly evolving field. To support the community, a curated list of literature is maintained at Awesome-Diffusion-Human-Centric.