Research map / Video / Video generation
Lodge: A Coarse to Fine Diffusion Network for Long Dance Generation Guided by the Characteristic Dance Primitives
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Lodge generates extremely long music-conditioned 3D dance sequences in parallel using a two-stage coarse-to-fine diffusion model whose intermediate representation is sparse 8-frame characteristic dance primitives, plus a Foot Refine Block for foot-skating reduction.
Problem
Generating long (minutes-scale) music-conditioned 3D dance sequences that have global choreographic structure while keeping local motion quality; autoregressive and sliding-window methods accumulate errors and freeze, and EDGE-style splicing lacks overall structure.
Approach
A global diffusion model generates sparse, high-kinematic-energy 8-frame "characteristic dance primitives" aligned to music structure; these guide parallel local diffusion segments via inpainting-style hard constraints at segment boundaries and soft guidance elsewhere, enabling parallel generation of long dances. A Foot Refine Block with cross-attention on foot keypoints further reduces skating.
Key results
- On FineDance, Lodge (DDPM) achieves FIDk 45.56 and FIDg 34.29 versus EDGE's 94.34 and 50.38, and the best BAS of 0.2397.
- Lodge (DDPM) reaches the highest Divg of 5.64 on FineDance, while Bailando has the best Run Time at 5.46 s versus Lodge's 30.93 s for DDPM sampling (1000 steps).
- On AIST++, Lodge improves over its EDGE baseline on several metrics (FIDk 37.09 vs 42.16, BAS 0.2423 vs 0.2334) but does not reach the best overall scores.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add characteristic dance primitives, mirroring, and beat alignment (ablation) | Lodge · FineDance test set | FIDk: 60.91 -> 45.56 | Helped |
| Add characteristic dance primitive guidance with mirroring (ablation) | Lodge · FineDance test set | Divk: 5.16 -> 5.54 | Helped |
| Add characteristic dance primitive guidance with beat alignment (ablation) | Lodge · FineDance test set | BAS: 0.2090 -> 0.2139 | Helped |
| Increase soft-cue guidance strength from s=0 to s=1 (ablation) | Lodge · FineDance test set | FIDk: 60.91 -> 45.56 | Helped |
| Increase soft-cue guidance strength from s=0 to s=1 (ablation) | Lodge · FineDance test set | BAS: 0.2090 -> 0.2397 | Helped |
| Add the Foot Refine Block (ablation) | Lodge · FineDance test set | FIDk: 53.48 -> 45.56 | Helped |
| Add the Foot Refine Block (ablation) | Lodge · FineDance test set | Foot Skating Ratio: 5.94% -> 5.01% | Helped |
Surprising findings
- Lodge's parallel architecture yields no clear speedup over Bailando in the reported run times, and its DDIM and DDPM run-time labels appear swapped in Table 1 relative to the text.
Limitations
- Cannot generate hand gestures or facial expressions.
- Long-dance evaluation relies on FineDance; AIST++ clips are too short to show long-range benefits.
- DDPM sampling with 1000 steps is slow relative to baselines.
Experimental setup
Datasets: FineDance AIST++
Benchmarks: FineDance AIST++
Metrics: FIDk FIDg Foot Skating Ratio Divk Divg BAS Run Time user study wins
Baselines: FACT MNET Bailando EDGE Li et al. DanceNet DanceRevolution
Builds on: EDGE DDPM Bailando LEMO MNET
Closest papers by meaning
- Bidirectional Autoregessive Diffusion Model for Dance Generation (CVPR 2024 · similarity 0.88)
- Controllable 3D Dance Generation Using Diffusion-Based Transformer U-Net (AAAI 2025 · similarity 0.87)
- MEGADance: Mixture-of-Experts Architecture for Genre-Aware 3D Dance Generation (NeurIPS 2025 · similarity 0.87)
- Breaking the Group Size Barrier: Parameter-Efficient Group Dance Generation with Chain-of-Dancers (NeurIPS 2026 · similarity 0.86)
- Harmonious Music-driven Group Choreography with Trajectory-Controllable Diffusion (AAAI 2025 · similarity 0.86)
- MoMu-Diffusion: On Learning Long-Term Motion-Music Synchronization and Correspondence (NeurIPS 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.