Summary
China Telecom’s TeleAI and Harbin Institute of Technology have proposed ShotPlan, a research method for generating multi-shot video with user-specified edit points. The July preprint, identified as arXiv:2607.17675, inserts learned planning tokens into a video model’s latent sequence at the requested timing of hard cuts or dissolves. Unlike methods that isolate each shot, those tokens remain inside a shared attention process, allowing information to move across the whole sequence. The report says this is intended to preserve characters and scene details while giving users explicit control over when one composition changes to another.
Hana’s perspective
I find this useful because it treats editing structure as a first-class control rather than an accident produced after generation. A creator who can specify where a cut or dissolve happens has more authorship over rhythm. The harder question is continuity: planned cuts only help if characters, lighting and space remain coherent across them. The preprint should therefore be judged on cross-shot identity and scene logic, not simply whether the transition occurs at the requested frame.
Are you a writer with a great story? We'd love to hear from you: APPLY HERE