Wan-Dancer

Feed it a song and a reference image of a dancer and Wan-Dancer choreographs a full performance: minute-plus 720p videos at 30fps whose movement stays locked to the beat well past the roughly 20-second wall where ordinary video diffusion loses coherence. The 14B model from Alibaba's Wan team plans hierarchically, first laying out global keyframes against the entire music track, then filling the motion between them with local temporal refinement, and it covers five trained styles: Chinese Classical, K-Pop, Street, Tap, and Latin. Code is Apache-2.0 with weights on Hugging Face and ModelScope plus sample images, music clips, and prompt files, and generation runs through provided shell scripts. Hardware expectations are serious; the reference configuration is eight 80GB A800 GPUs with CUDA 12.4, so this is datacenter or rented-GPU territory rather than a desktop toy. Released July 2026 with an accompanying paper, it stands as the strongest open option for long-form music-conditioned human animation.