Warp as History

Warp-as-History is a research method from Yifan Wang and Tong He (Shanghai Jiao Tong University, Shanghai AI Laboratory, Shanghai Innovation Institute) for camera-controlled video generation. It uses a pretrained history-conditioned video generator and makes it follow a target camera trajectory without a new camera encoder, control branch, or test-time optimization.

The method renders the available observation under the target camera path, packs it as camera-warped pseudo-history, and aligns those history tokens to the target frames being denoised. Tokens with no valid source observation are dropped, so newly visible or unreliable regions are completed by the generator. A lightweight offline LoRA update on a single separate camera-annotated video stabilizes the behavior across unseen scenes and trajectories.

The project page describes interactive camera trajectory following and viewpoint manipulation comparable to Genie 3 and HappyOyster, using one training example. A YouTube creator reported that code, models and training scripts were released, but no repository or weights links appear in the provided material. It is a research project, not a hosted product.