wtf is wan-dancer?

Wan-Dancer is a Python research project that generates dance videos directly from music using AI. Give it a reference image of a person, a music file, and a chosen dance style, and it produces a video of that person dancing in sync with the rhythm, for durations well beyond what similar tools can normally manage. The core problem the project addresses is that most AI video generators fall apart after about twenty seconds: the dancer's identity drifts, motions start repeating, and the video loses coherence over time. Wan-Dancer works around this with a two stage approach. First it plans the overall structure of the dance across the entire song, producing lower resolution global keyframes that capture the big picture movements. Then a second local refinement stage fills in high resolution detail while staying consistent with that global plan. According to the accompanying research paper, this lets the system produce stable 720p video at 30 frames per second for videos over a minute long, something existing methods struggle to do. The project ships with example prompts and reference material for five distinct dance styles, including Chinese classical dance, K-pop, street dance, tap dance, and Latin dance, each with its own prompt files for both the global and local generation stages. Running Wan-Dancer requires a fairly serious hardware setup. The authors tested it on eight NVIDIA A800 GPUs with 80GB of memory each, running Ubuntu and a specific version of Python, along with a long list of exact dependency versions including PyTorch, diffusers, and flash attention. Usage involves running a shell script to generate the global keyframe video first, then running a second shell script that takes that output and produces the final high resolution version, with parameters controlling things like the random seed, the reference image, the music file, and the number of diffusion steps used during generation. This is research code released alongside a published paper, and the models are also available through Hugging Face and ModelScope for anyone who wants to try the approach without running the full pipeline themselves.