HelloWorld: Enabling Socially Interactive Characters in Video World Models

HelloWorld adds socially interactive characters to video world models. A self-distillation pipeline synthesizes training clips that combine character responses with camera motion, enabling camera-pose conditioning without sacrificing interaction quality. At inference, a training-free module temporally localizes the response by modifying DiT cross-attention masks so interaction-related text attends only to frames inside the button-press window. The authors also introduce HelloWorldBench, a 400-sample benchmark with three social-interaction metrics and three conventional metrics. According to the abstract, HelloWorld outperforms several baselines on interaction quality while retaining state-of-the-art picture aesthetics and camera-pose following.