StarWAM makes WAM video-world-model experiments more modular
StarWAM is an early codebase for experimenting with video-generation-based WAM systems. Its goal is to make parts of a easier to swap, so different designs can be tested inside one shared framework. It supports MoT-style designs such as Motus and FastWAM, shared DiT designs such as DreamZero, and feature-conditioned designs inspired by StarVLA WM4A and Mimic Video.
It can use Wan2.2 5B and Cosmos Predict2 as backbones. It also includes support for testing robot-task behavior. The project is still rough because models have complex internal parts, and each backbone tends to organize those parts differently.
Large training runs and tuning have not been done yet because of limited compute and time.
Key points
- StarWAM is an codebase for modular WAM experiments.
- It supports Motus, FastWAM, DreamZero, StarVLA WM4A, and Mimic Video-style ideas.
- It works with Wan2.2 5B and Cosmos Predict2 backbones.
- It includes support for robot-task .
- It is still early, with limited training and tuning so far.