ShotStream: Real-Time Multi-Shot Video Generation

jiqizhixin · x · 2026-07-19

Researchers from CUHK and Kuaishou introduced **ShotStream**, a streaming multi-shot video generation model designed for interactive storytelling. The core approach uses **causal generation** to predict the next shot based on historical footage, paired with a **dual-memory cache** to maintain coherence. It also employs distillation techniques to suppress error accumulation in long sequence generation. The authors claim this method achieves **sub-second latency** and **16 FPS** on a **single GPU**, reaching a real-time performance balance that long-range autoregressive and bidirectional multi-shot models struggle to achieve simultaneously.

Original post →

More from Multimodal

Multimodal channel →