AI video generators are getting better at remembering what happened in a story - this paper is about making them remember the right parts.
Researchers behind Weave Forcing built a training-free framework for interactive long video generation, meaning it adds smarter memory handling without retraining the underlying model. An LLM first splits a prompt into character and background pieces, then points each piece to the specific past shot it should draw from, instead of dumping an entire prior scene into the mix. A technique called masked memory weaving then builds semantic masks from attention maps to expose only the matching visual tokens in that compressed memory, filtering out whatever doesn't belong. The team also built in "coverage adaptive RoPE," which tunes how closely a new shot leans on old references depending on whether that reference is complete, partial, or missing - aimed at visual artifacts that crop up when thin references sit too close to the current frame.
The problem here is narrower than "make AI video longer." Autoregressive generators already handle extending a single continuous shot reasonably well. What they're bad at is cutting back to an earlier character or location without either ignoring it or smearing in unrelated visual baggage - the exact failure mode anyone trying to build actual multi-scene AI stories keeps hitting.
The paper reports improved cross-shot consistency with "competitive" visual quality and text alignment - solid-sounding numbers, but all self-reported, with no outside benchmark or shipped product to check them against yet.