Most video-generation research treats explainer videos—a worked geometry problem, why the sky is blue, how a dynasty fell—as just another target for pixel-level synthesis. A new arXiv preprint, Reels-RSI, argues this is the wrong frame. The authors contend that explainer videos are a different kind of artifact: a program, composed of a script, a storyboard, and—as the title suggests—the code that renders the final video.
On this view, the generative task is not to produce pixels but to produce and refine the program itself. The paper introduces Reels-RSI, a method for program-level self-improvement of such code-rendered explainer videos. Because the abstract is brief, the precise mechanics are not fully spelled out, but the central claim is that improving the underlying program is a more natural and effective route than iterating on raw video output.