Advancing Narrative Long Video Generation via Training-Free Identity-Aware Memory
Researchers have developed IAMFlow, a novel framework designed to improve the consistency and identity tracking in long video generation. This training-free method explicitly models and follows persistent entities across evolving prompts, preventing issues like identity drift and attribute loss. IAMFlow utilizes an LLM to extract entities and assign IDs, with a VLM refining attributes from rendered frames for precise tracking. The framework also includes an inference acceleration pipeline and a new benchmark, NarraStream-Bench, for evaluating narrative streaming video generation. AI
IMPACT Improves consistency in long-form AI video generation, potentially enabling more coherent and narrative-driven content.