Conditional Visual Content Creation Based on Generative Models

This thesis studies conditional visual content creation using generative models, systematically covering single- and multi-conditional image and video generation. It proposes U2AFN for uncertainty-aware image inpainting, OmniPainter for fast stylized text-to-image generation without fine-tuning, Latte for latent diffusion Transformer for video generation, and Cinemo for instruction-driven image animation via motion residual modeling. Across these tasks, the thesis improves visual quality, semantic coherence, and temporal consistency, providing practical insights and methodologies that support future research and commercial applications in conditional visual generation.