DomainShuttle is a new method for open domain subject-driven text-to-video (S2V) generation that balances subject fidelity with stylistic flexibility. Researchers Nan Chen and colleagues published the study on June 24, 2026, addressing the challenge of maintaining reference subject features while enabling cross-domain edits such as novel styles or semantic shifts. Unlike previous methods focused primarily on in-domain fidelity, DomainShuttle allows for precise subject-level modeling that adapts across different application domains.

To achieve this, the method uses Domain-MoT, an approach that decouples videos and reference features using domain-aware AdaLN (a normalization technique adjusting layer parameters based on domain attributes) for specific modeling of reference images. It further incorporates a Video-Reference DualRoPE scheme, which assigns reference image tokens and video tokens into separate RoPE (rotational positional embedding used to handle sequence order) spaces. This separation ensures precise spatial modeling of the subject. Additionally, a Cross-Pair Consistent Loss function extracts intrinsic subject characteristics while ignoring irrelevant features, resulting in improved performance compared to existing generation techniques across diverse scenarios.

DomainShuttle is a new method for open domain subject-driven text-to-video (S2V) generation that balances subject fidelity with stylistic flexibility. Researchers Nan Chen and colleagues published the study on June 24, 2026, addressing the challenge of maintaining reference subject features while enabling cross-domain edits such as novel styles or semantic shifts. Unlike previous methods focused primarily on in-domain fidelity, DomainShuttle allows for precise subject-level modeling that adapts across different application domains.

To achieve this, the method uses Domain-MoT, an approach that decouples videos and reference features using domain-aware AdaLN (a normalization technique adjusting layer parameters based on domain attributes) for specific modeling of reference images. It further incorporates a Video-Reference DualRoPE scheme, which assigns reference image tokens and video tokens into separate RoPE (rotational positional embedding used to handle sequence order) spaces. This separation ensures precise spatial modeling of the subject. Additionally, a Cross-Pair Consistent Loss function extracts intrinsic subject characteristics while ignoring irrelevant features, resulting in improved performance compared to existing generation techniques across diverse scenarios.