ShotStream: Kling AI Ships Streaming Multi-Shot Video Gen at 16 FPS on Single GPU
arXiv / HuggingFace Daily Papers·medium signal
Kling AI Research published ShotStream (arXiv 2603.25746), a causal multi-shot video architecture that enables interactive storytelling with sub-second latency at 16 FPS on a single GPU. Users can dynamically alter content, visual styles, or introduce new characters mid-generation via streaming prompts. The approach fine-tunes a text-to-video model into a bidirectional generator, then distills it into a causal student via Distribution Matching Distillation with dual-cache memory for shot coherence. Trending at 105 upvotes on HuggingFace.