What's this about?
WanPE expands a short user request into a timed plan for actions, shots, cameras, lighting and sound before Wan 3.0 generates the video.
What exactly is reverse construction?
The model learns by watching a finished video and reconstructing the detailed screenplay that could have produced it, rather than only rewriting short prompts into longer prose.
Why does this matter?
author-reported preprint; 36 HF upvotes.
- Training uses 1.05 million real-world videos and a 397B-parameter model.
- Human preference gains grow sharply for 30-second multi-shot videos.
- Comparisons are tied to the authors' generator and evaluation arena, so generalisation needs independent tests.