MOVi: Training-Free Text-Conditioned Multi-Object Video Generation

Recent diffusion-based text-to-video (T2V) models show strong progress but still struggle with multi-object generation, often rendering objects as static backgrounds or blending their features. We propose a training-free method that combines diffusion models with large language models (LLMs), where the LLM acts as a "director" to guide object trajectories via noise re-initialization for controllable motion. We further adjust the attention mechanism to capture object-specific features and avoid cross-object interference. Experiments show our approach boosts multi-object accuracy and motion dynamics by 42%, while preserving fidelity and smoothness.