We introduce SkyReels-A3, an end-to-end multimodality-conditioned framework for synthesizing high-fidelity and temporally coherent human videos. Building on pretrained video diffusion transformers, our framework enables long-form video generation with diverse and controllable conditioning via multimodal inputs. It accepts image inputs of any aspect ratio – including portraits, half-body, or full-body shots – delivering highly realistic and superior quality results across diverse scenarios. We employ a learning-based interpolation strategy to support minute-level controlable video generation, while introducing reinforcement learning to enhance inter-action naturalness. Comprehensive benchmarks demonstrate that SkyReels-A3 excels in lip-sync accuracy, identity consistency, and realistic facial dynamics, particularly under complex and challenging conditions.