你是否曾被AI视频生成的不连贯性所困扰？ 在交互式创作中，仅仅切换一句提示词，故事就可能瞬间"崩塌"：一个角色暂时离开画面后再次出现，却"面目全非"，仿佛换了演员；或者，当你尝试引 入一个新角色，AI却在后续的剧情中反复"召唤"这个新人，甚至将多个角色的特征混淆在一起。这种"金鱼记忆"式的顽疾，正是长视频生成在叙事上的一 大致命伤。 现在，来自香港大学和快手可灵（Kling）团队的研究者们，联合推出了突破性方案——MemFlow。 3. "各自为政"的流程：还有一些流程试图将任务拆分，先让一个模型制作关键帧脚本，再让另一个模型根据脚本生成视频。这种方式在根据各段脚本分别 生成时是各自独立的，拼接成的完整视频缺乏全局一致性。 这些僵化的、非自适应的记忆策略，无法应对交互式创作中流动的、不可预测的叙事需求，这正是导致交互式长视频生成一致性差的原因。 产生真正的长时记忆与叙事连贯性 这是一种创新的流式自适应记忆机制，它赋予了AI强大的长时记忆与叙事连贯性，有望彻底解决上述难题。 流动的叙事 vs. 僵化的记忆 为了生成长视频，主流模型普遍采用"分块生成"的策略，即像放映幻灯片一样，一段一段地生成视频片段。 然而，如 ...

Core Viewpoint - The article discusses the challenges of AI-generated long videos, particularly issues with narrative coherence and character consistency, and introduces MemFlow, a new memory mechanism designed to address these problems [1][2][3]. Group 1: Challenges in AI Video Generation - AI-generated long videos often suffer from narrative inconsistencies, such as characters appearing different after a scene change or the AI confusing multiple characters [1]. - Traditional models use a "chunk generation" strategy, which leads to difficulties in maintaining continuity across video segments [4][6]. - Existing memory strategies have significant limitations, including only remembering the first segment, fixed-size memory compression, and independent processing of segments, all of which contribute to narrative disjointedness [5][6]. Group 2: Introduction of MemFlow - MemFlow is a novel adaptive memory mechanism that enhances AI's long-term memory and narrative coherence, aiming to resolve the aforementioned issues [3][7]. - It establishes a dynamic memory system that maintains visual consistency and narrative clarity, even in complex scenarios with multiple characters [8][9]. Group 3: Mechanisms of MemFlow - MemFlow employs two core designs: Narrative Adaptive Memory (NAM) and Sparse Memory Activation (SMA), which allow for efficient retrieval of relevant visual memories and reduce computational load [11]. - NAM intelligently retrieves the most relevant memories based on current prompts, while SMA activates only the most critical information, enhancing both speed and quality of video generation [11]. Group 4: Performance Evaluation - MemFlow demonstrated significant improvements in key performance metrics, achieving a quality consistency score of 85.02 and an aesthetic score of 61.07, outperforming other models in long video generation tasks [13][14]. - The model maintained high semantic consistency throughout the video, particularly in the latter segments, which is crucial for narrative coherence [15][17]. - In terms of subject and background consistency, MemFlow achieved scores of 98.01 and 96.70 respectively, showcasing its ability to maintain visual unity amidst complex narrative changes [18][17]. Group 5: Visual Comparisons and Efficiency - Visual comparisons highlighted MemFlow's superiority in maintaining character consistency and avoiding narrative confusion, unlike other models that struggled with character drift and inconsistencies [19][21][23]. - MemFlow operates efficiently on a single NVIDIA H100, achieving a real-time inference speed of 18.7 FPS, with minimal performance loss compared to baseline models [25]. Group 6: Future Implications - MemFlow represents a significant advancement in AI video generation, transitioning from simple video creation to complex narrative storytelling [26][27]. - This innovation indicates a shift towards AI systems capable of understanding, remembering, and coherently narrating stories, marking the dawn of a new era in AI video creation [28].