DomainShuttle enables subject-consistent video generation

DomainShuttle solves the classic trade-off between style adherence and identity preservation by cleanly separating the spatial embedding spaces of the reference image and the target video.

* The Video-Reference DualRoPE is a clever architecture change that prevents reference image tokens from contaminating the spatial structure of the generated video.

* Leveraging the Wan-2.2 foundation model gives the project a strong baseline, making it much more practical than training a subject-driven model from scratch.

* While the results look promising, actual performance remains heavily dependent on the quality and orientation of the input reference image.