The technique converts desired camera motion into camera-warped pseudo-history and feeds those visual history tokens through the backbone pathway that the video model already uses to continue generation. It aligns history tokens with the target frames being denoised and drops invalid source observations. A lightweight LoRA update from one camera-annotated video stabilizes the zero-shot camera-following behavior across unseen scenes and trajectories.
Warp-as-History is useful for researchers developing camera-controllable generative video, interactive viewpoint editing, and low-resource adaptation methods. Its main value is that it exposes camera-following behavior in a frozen history-conditioned video generator rather than requiring a new encoder, control branch, or test-time optimization. It fits workflows where small amounts of camera-labeled data must unlock stronger controllability.
