Sliding-Window T2V

Key Insight

Sliding-window generation is the cheapest way to push a text-to-video (T2V) model past the clip length it was trained on: generate a series of short clips that overlap by a few frames, then blend each overlap so the joins are invisible. This project chains 5-second clips into a 30-second video and fuses the shared frames in the model's latent space — averaging there is smoother than on raw pixels and avoids ghosting. Because the only thread tying distant moments together is whatever those short overlaps can carry forward, the picture slowly drifts: colors creep and a character's outfit mutates the further you travel from the opening frame. It is the first long-form trick worth trying, since it needs no retraining — only a clever way to stitch existing outputs.