Wan-Dancer: Minute-Level Music-to-Dance Generation
pmttyji · reddit · 2026-07-13
Wan-Dancer: Minute-Level Music-to-Dance Generation
The authors introduce Wan-Dancer, a model designed to generate long, high-resolution, rhythm-synchronized dance videos from full music tracks. It aims to overcome the instability that current diffusion video models face around the 20-second mark.
Core Approach
- Decouples generation into two tiers: global keyframe planning + local temporal refinement to ensure long-range consistency.
- Utilizes the full music context to maintain structural coherence over long durations.
- Introduces time-mapped RoPE for dynamic frame rate adaptation, improving audio-visual alignment.
- Employs optical flow-based loss to enhance motion continuity.
- Uses motion-speed control to preserve details during fast movements.
Results & Availability
- The paper claims it outperforms traditional duration-limited methods in experiments.
- Capable of generating stable dance videos exceeding 1 minute at 720p/30fps.
- Covers 5 dance styles and supports both audio and text conditions.
- Model weights and inference code are open-sourced, with demos available on ModelScope Studio and Hugging Face Space.
Related event: Wan-Dancer-14B Open-Sourced for Music-Driven Long Dance Video Generation(6 posts)→
More from Multimodal
- CapCut plugin is coming to ChatGPT for AI-powered video editing workflows — HeyZoyaKhan · 2026-10-27
- Magnific teases October 28 release: 'We're making it move' — cuenca · 2026-10-27
- Midjourney --sref demo: retro Americana Batman art series — gizakdag · 2026-10-27
- AI video is a new medium, not cinema or animation — and it's perfect for depicting dreams — Kyrannio · 2026-10-27
- Dad turns his toddler's 31-year-old quote into a song via SUNO kids' choir — NathanpmYoung · 2026-10-27
- Band's local-AI music video hits 200k views: 900 generations kept 64, 3-second clips beat face drift — Hwang0500 · 2026-10-27