Content area
INTRODUCTION: The rapid proliferation of Generative AI (AIGC) in new media art has intensified the need for real-time, distributed video generation with stable performance and low latency. Conventional centralized rendering and static scheduling frameworks often encounter load imbalance and communication bottlenecks in heterogeneous environments, resulting in degraded visual coherence and responsiveness. To address these challenges, this study develops a unified and adaptive distributed framework, termed H-RLSCO (Heterogeneity-aware Reinforcement Learning and Scheduling Co-Optimization), designed to enhance both computational efficiency and artistic consistency in large-scale AI video generation. The framework integrates three complementary modules: a Heterogeneity Perception Module (HPM) for node profiling and adaptive task partitioning, a Reinforcement Learning Scheduling Controller (RLSC) for dynamic task migration, and a Generation-Scheduling Co-Optimization (GSCO) mechanism that incorporates content-complexity feedback into scheduling decisions to maintain multimodal synchronization. Experiments on the ArtScene-4K and StageSyn-Real datasets demonstrate that H-RLSCO reduces average latency by 14.4% and decreases Fréchet Video Distance by approximately 12.5% compared with the RL-Scheduler baseline, while limiting performance fluctuation to within 3% under multi-noise conditions (p < 0.01). These gains remain consistent across varying bandwidths and node capabilities on a five-node heterogeneous cluster, confirming robust real-time behavior and balanced utilization. Nevertheless, the scalability of H-RLSCO remains constrained when applied to large-scale node clusters, suggesting future work should explore multi-agent reinforcement learning and lightweight diffusion-Transformer architectures to enhance efficiency and expand applicability.