StreamHOI: Interaction-aware Temporal Memory Adaptation for Streaming HOI Video Generation

StreamHOI is a low-latency framework for long-duration streaming human-object interaction video generation. Rather than turning a heavily conditioned offline HOI pipeline into a streaming system, it studies how an image-to-video generator should organize bounded historical memory. The method profiles transformer blocks offline with respect to HOI and surrounding regions, then uses bias-guided memory-specialized training to match block-specific memory layouts. A memory distance scaling module improves access to early interaction states. The paper reports 17.6 FPS and 0.75-second first-chunk latency, alongside comparisons with long-video and recent HOI baselines.