Apppublic

CNcreator0331/domainshuttle-subject-video

sourceHugging Faceupdated 3mo agoView on Hugging Face

2likes

App README

DomainShuttle — Subject-driven Text-to-Video

Freeform, open-domain subject-driven video generation. Provide one to three reference images of subjects (people, objects, characters) plus a text prompt, and the model renders those subjects into a new video — including cross-domain restyling (e.g. a real photo re-imagined in Pixar style).

—Model: `CNcreator0331/DomainShuttle_weight` — a two-expert (high/low noise) MoE finetune of Wan2.2-A14B.
—Code: HKUST-C4G/DomainShuttle
—Project page: cn-makers.github.io/DomainShuttle

Runs on ZeroGPU with accelerate model-CPU-offload: the two ~14B experts are streamed to the GPU one at a time (only one is active per diffusion step), keeping peak VRAM within the ZeroGPU budget.