PreciseVideo: a dual-process framework for zero-shot text-to-video generation with quantitative content control

Abstract

Text-to-video (T2V) generation has recently gained significant attention, yet existing methods primarily focus on global temporal consistency and lack fine-grained, element-wise control over background dynamics and character behaviors. We propose PreciseVideo, a zero-shot T2V framework that enables controllable video synthesis at both the background and foreground levels. PreciseVideo introduces a dual-stage generation paradigm, separating background and character synthesis, and incorporates three novel modules: the Region-Independent Noise Modulator for quantifiable, region-wise temporal dynamics, Sparse Fusion Attention for structured cross-frame coherence, and Optimal-Reference-Frame Attention to preserve full-body character identity and appearance. This modular design ensures high-fidelity, temporally coherent, and behaviorally consistent video generation, even in complex multi-character scenarios. Extensive experiments demonstrate that PreciseVideo excels in element-wise controllability, character quantity accuracy, and multi-character scene synthesis compared with both zero-shot and training-based baselines. Ablation studies validate the effectiveness of each proposed module, while additional evaluations on scene-to-character and inter-character occlusions highlight the framework’s robustness and flexibility. Collectively, our results establish PreciseVideo as a highly controllable and scalable T2V approach, filling a critical gap in fine-grained, element-wise controllable video generation and setting a foundation for future advances in complex scene synthesis. Our code and related experimental results are available at https://github.com/GG-Bond2023/PreciseVideo.