In detail
- Stores internal image features in 3D space slots instead of colored point clouds
- Projects stored features directly onto target camera views, skipping rendering and re‑encoding
- Generates videos in segments, writing stable geometry back into growing memory while filtering out moving objects and sky
- Built on Alibaba’s Wan2.2 with a small add‑on module teaching the model to use the new memory
Why it matters
Avoiding costly pixel rendering reduces compute and memory use while keeping scene structure consistent during long camera moves — useful for companies building simulations, virtual tours, or long‑take video generation.
For you Evaluate whether your video or simulation pipelines could benefit from feature‑based spatial memory; plan pilots for models that maintain long‑term scene coherence.