Domain Shuttle keeps video characters consistent across mixed art styles

Anyone who has experimented with AI video generation knows the struggle: render a new shot, and your main character suddenly has an entirely new face. Try to mix a cartoon mascot with live-action actors, and the scene usually melts into surreal visual artifacts.

Researchers from the Hong Kong University of Science and Technology open-sourced Domain Shuttle, a new subject-driven video generation project designed to solve that headache. By feeding the model reference images, creators can generate videos that strictly preserve subject identities, even across jarringly different artistic styles.

Live-action, 2D anime, and 3D in one frame

Most existing subject-driven video models only work well within their own comfort zone. If you feed in a realistic human portrait, the tool expects to generate a realistic video. Ask it to turn everything into a Saturday-morning cartoon, and the subject identity quickly falls apart.

Domain Shuttle attacks this limitation directly. In side-by-side demonstrations, the model takes a real person, a flat 2D anime figure, and a 3D character, and renders all three interacting naturally in the exact same scene.

The tool also handles complex spatial details. Prompts can specify a man talking to a 3D holographic figure, or viewing a 2D graffiti mural on a street wall. In one demanding test, the model successfully printed a reference 2D character onto a coffee mug held by a moving woman without texture warping.

Under the hood, the system separates identity from style. The HKUST team introduced a Domain-MoT architecture and a dual RoPE scheme, decoupling subject attributes from domain styles so the model remembers who the character is without being handcuffed to the original medium.

Running it locally will test your hardware. Domain Shuttle is built on top of the Wan 2.2 video foundation model, and the model weights weigh in at 35GB. You will need a high-end enterprise GPU to spin it up, though the code and setup instructions are already live on GitHub.

Why it matters

Domain Shuttle tackles what has long been the primary blocker for narrative AI filmmaking: consistent identity across flexible stylistic domains.

For commercial creators, indie animators, and social video producers, this opens up massive practical shortcuts. Dropping a virtual mascot into real-world product demos or producing multi-shot ads typically demands dedicated 3D tracking, rigging, and manual compositing. Being able to direct mixed-reality scenes entirely through text prompts slashes production time.

It also highlights the rapid evolution of the open-source video ecosystem. By building directly on Wan 2.2, the team proved that specialized architectures can outperform monolithic models on niche control tasks without requiring millions of dollars in base training.

The 35GB footprint means everyday hobbyists will probably have to wait for distilled checkpoints or cloud interfaces. But for studio pipelines looking for consistent multi-subject generation, this is one of the most capable tools yet.