36氪AI测评近期以《霸王别姬》为主题对可灵3.0进行了系统压力测试，评估其在短片创作中的表现能力[1]。测试结果显示，该模型在关键镜头、构图、光影氛围和人物一致性方面表现突出，能够稳定输出高质量画面[1]。然而，在涉及跨空间移动、多道具配合和多人物交互的复杂场景中，产出的确定性和稳定性明显下降[1]。
测试涵盖了5轮空间路径测试和数十轮战斗镜头测试，充分验证了模型在复杂交互场景中的控制困难[1]。针对这一局限，36氪建议内容创作者采用分镜策略，将长动作序列拆解为单目标的短镜头单元，各自独立生成后再通过后期剪辑串联为完整叙事段落[1]。这一方案旨在充分发挥可灵3.0在单一视觉目标上的优势，同时规避其在复杂多元素协调中的不稳定性[1]。
A comprehensive evaluation by 36Kr AI conducted stress tests on Keling 3.0 using themes from the classic film "Farewell My Concubine" to assess its performance in creating complex narrative shorts.[1] The model demonstrated notable strengths in key technical areas, excelling at composition, lighting, emotional atmosphere, and character consistency.[1] Short action sequences produced stable, high-quality outputs, and the system proved particularly effective for capturing iconic moments.[1]
However, the evaluation revealed significant limitations when handling intricate storytelling scenarios.[1] Complex interactive scenes involving cross-space movement, multiple props, and multi-character action sequences showed substantially lower output certainty compared to single-objective short clips.[1] Testing encompassed five rounds of spatial pathway tests and dozens of combat scene trials, confirming that controlling such complex interactions remains difficult.[1] To work within these constraints, researchers recommend that creators decompose lengthy action sequences into single-objective short clips and reassemble them into complete narratives through storyboard design and post-production editing.[1]
