Researchers from the College of Hong Kong and Kuaishou’s Kling group have collectively proposed MemFlow, a novel method designed to deal with the long-standing challenges of reminiscence decay and narrative inconsistency in AI-generated lengthy movies.
MemFlow introduces a dynamic, adaptive streaming long-term reminiscence mechanism that considerably improves narrative coherence and visible consistency throughout prolonged video sequences. Conventional strategies usually depend on inflexible reminiscence methods, leading to identification drift or character confusion over time.
The answer options two core parts: Narrative-Adaptive Reminiscence (NAM), which retrieves essentially the most related historic visible context primarily based on the present immediate, and Sparse Reminiscence Activation (SMA), which selectively prompts key data to keep up computational effectivity.
In benchmark checks, MemFlow achieved a VBench-Lengthy total high quality rating of 85.02 and an aesthetic rating of 61.07, whereas sustaining secure long-range semantic consistency. Topic consistency reached 96.60, and real-time inference achieved 18.7 FPS on a single NVIDIA H100 GPU, highlighting each high quality and effectivity good points.
Supply : liangziwei
Elevate your perspective with NextTech Information, the place innovation meets perception.
Uncover the most recent breakthroughs, get unique updates, and join with a worldwide community of future-focused thinkers.
Unlock tomorrow’s developments right this moment: learn extra, subscribe to our e-newsletter, and change into a part of the NextTech neighborhood at NextTech-news.com
