Prompt Relay routes prompts by scene

This is a neat paper because it attacks a real failure mode in video generation without asking model builders to retrain or redesign their stack.

• –The core value is temporal prompt isolation: each scene gets its own prompt, so concepts do not smear across cuts
• –Because it works at inference time, it is easier to adopt across existing video models than architecture-heavy fixes
• –The method is especially relevant for story-driven generation, where prompt order, duration, and transitions matter as much as visual quality
• –The reported Wan integration suggests this is already moving from paper idea toward practical workflow tooling
• –If the gains hold up broadly, this could become a standard control layer for multi-scene video prompting