Control video generation by warping the past, not the model

Camera-controlled video generation typically requires expensive post-training on large annotated datasets or test-time optimization. This work introduces Warp-as-History, which converts camera-induced geometric warps into pseudo-historical frames aligned with target positional encodings, enabling zero-shot camera control without modifying the base model. Optional LoRA finetuning on just one annotated video further improves adherence to camera trajectories and generalizes to unseen videos. Experiments across diverse datasets show improvements in camera adherence, visual quality, and motion consistency compared to test-time guidance baselines.