Researchers from the Kling team at the University of Hong Kong and Kuaishou have jointly proposed MemFlow, a new approach designed to address the long-standing challenges of memory decline and narrative inconsistency in long AI-generated videos.
MemFlow introduces a dynamic, adaptive streaming long-term memory mechanism that significantly improves narrative coherence and visual consistency across augmented video sequences. Traditional methods often rely on rigid memory strategies, leading to identity slippage and personality confusion over time.
The solution features two core components: Narrative Adaptive Memory (NAM), which retrieves the most relevant historical visual context based on the current prompt, and Sparse Memory Activation (SMA), which selectively activates critical information to maintain computational efficiency. In benchmark tests, MemFlow achieved an overall VBench-Long quality score of 85.02 and an aesthetic score of 61.07, maintaining stable long-range semantic consistency. Subject consistency reached 96.60, and real-time inference achieved 18.7 FPS on one NVIDIA H100 GPU, highlighting improvements in both quality and efficiency.
Source: Liang Ziwei
