Warp-as-History learns camera control from a single training video

Warp-as-History feeds camera-motion warps into a video model as pseudo-history frames with aligned positional encoding. A lightweight LoRA trained on one video generalizes to unseen clips and supports zero-shot camera trajectories without changing the model architecture or optimizing at inference time. The authors reported gains in camera-path adherence, visual quality, and motion.

Visit original source