Calabash Brothers Model
2024/AI Model
Project Lead@Shanghai Film
Revitalizing the 40-year-old "Calabash Brothers" IP through a custom generative video model trained on extreme small data.In 2024, as part of Shanghai Film's "iNEW" AI strategy, I led the R&D initiative to build a custom video generation model for the legendary 40-year-old IP "Calabash Brothers" (葫芦兄弟).Collaborating with the StepFun (阶跃星辰) technical team, we tackled a unique challenge in the GenAI field: how to train a high-fidelity video model for a specific artistic style—Chinese Paper-Cut Animation—using an incredibly limited dataset.
Award
- Selected as a key innovation exhibit at WAIC 2024 and presented to Premier Li Qiang, highlighting AI's role in revitalizing cultural heritage.
- Serves as the foundational case study for Shanghai Film's "iNEW" strategy, validating the commercial viability of Generative Video for specific cultural aesthetics.
1. The Small DataBreakthroughUnlike modern 3D animation, the original Calabash Brothers produced only 190 minutes of total footage in its lifetime. Training a video model on such scarce data is notoriously difficult. I guided the team to achieve a "phase-level breakthrough" in low-resource training, successfully teaching the model to understand and replicate the intricate textures and jerky movement characteristic of stop-motion paper-cut artistry.
2. High Controllability & Narrative Logic The model goes beyond simple style transfer; it achieves precise semantic control. We successfully enabled the model to generate novel plot-driven actions—such as commanding the "Grandpa" character to breathe fire or "Fire Child" to spray water—while maintaining strict character consistency. The model also demonstrates advanced perception of camera movement and cinematic framing.