Wan-Dancer-14B

Wan-Dancer-14B, published on 13 July 2026 under Apache 2.0, generates dance video from music. The interesting problem is not making a figure move to a beat, which shorter models already do; it is keeping a performance coherent across a whole song. A clip generated a few seconds at a time drifts - the dancer changes position, the energy no longer matches the section of the music, and the result reads as a loop rather than a routine. The published approach splits the work in two. A global stage reads the entire music track first and plans keyframes across its full length, deciding the shape of the performance before any detail exists. A local stage then takes that skeleton video and refines it into the final high-resolution result. The team's own name for this is a hierarchical framework, and the paper's claim is minute-scale coherence, which is the number that separates this from the previous generation of music-driven video. Input is a reference photograph plus an audio file. The repository ships prompt presets for five genres - Chinese classical, K-pop, street, tap and Latin - which is a useful admission in itself: the model is steered towards a style in words, not inferred purely from the music. Its place in the catalogue is worth naming. Alibaba stopped publishing weights for its main video line after Wan2.2 in mid-2025, and everything from 2.5 to Wan3.0 is API-only. Specialised models like this one are the exception: the flagship stays closed, while the research side of the same team keeps releasing openly. Anyone counting open Alibaba video models by generation number would miss it.

▸Videos

No videos yet.

▸Reviews