Research map / 3D vision / 3D generation
POPDG: Popular 3D Dance Generation with PopDanceSet
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.
Introduces PopDanceSet, a popularity-filtered 3D dance dataset from Bilibili, and POPDG, an iDDPM-based music-to-dance diffusion model with a Space Augmentation attention and an Alignment Module, plus PBC metric.
Problem
Music-driven 3D dance generation suffers from scarce public datasets (AIST++ limited in genre and aesthetics), unstable/complex training, weak diversity, and weak joint-level spatial modeling and music-dance alignment.
Approach
Builds a popularity-filtered dataset of dance videos for young audiences and a diffusion model (iDDPM/DDIM) with spatial attention whose weights are boosted along the SMPL joint hierarchy toward the hip root, plus an alignment module that fuses temporal features of music and dance via DenseFiLM-style conditioning.
Key results
- On PopDanceSet, POPDG reaches PFC 4.3697 and PBC 5.3863 versus EDGE's 5.9701 and 5.8535, with Divk 4.8641 vs 3.6065.
- On AIST++, POPDG achieves PFC 0.8014 vs EDGE 0.9201 and Beat Align 0.469 vs 0.456 (Divk and Divg differ by mixed margins).
- In a 20-person user study, 70% of participants preferred dances generated from PopDanceSet-trained models over AIST++-trained ones.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use POPDG instead of EDGE | Music-driven dance generation · PopDanceSet test | PFC: 5.9701 -> 4.3697 | Helped |
| Use POPDG instead of EDGE | Music-driven dance generation · AIST++ test | PFC: 0.9201 -> 0.8014 | Helped |
| Use POPDG instead of Bailando | Music-driven dance generation · PopDanceSet test | Beat Align Score: 0.480 -> 0.482 | Helped |
| Use POPDG instead of FACT | Music-driven dance generation · AIST++ test | Divg: 6.6993 -> 3.6707 | Hurt |
| Add MF-Attention (ablation) | POPDG baseline · Ablation study | PFC: 0.9600 -> 0.9290 | Helped |
| Add MF-Attention (ablation) | POPDG baseline · Ablation study | PBC: 5.5712 -> 5.3363 | Helped |
| Add DS-Attention after MF-Attention (ablation) | POPDG · Ablation study | PFC: 0.9290 -> 0.9116 | Helped |
| Add DS-Attention after MF-Attention (ablation) | POPDG · Ablation study | PBC: 5.3363 -> 5.0391 | Helped |
| Add AM to the MF-Attention and DS-Attention configuration (ablation) | POPDG · Ablation study | PFC: 0.9116 -> 0.8487 | Helped |
| Add AM to the MF-Attention and DS-Attention configuration (ablation) | POPDG · Ablation study | Beat Align Score: 0.431 -> 0.448 | Helped |
| Use iDDPM instead of DDPM (ablation) | POPDG · AIST++ test | PFC: 1.1077 -> 0.8014 | Helped |
| Use iDDPM instead of DDPM (ablation) | POPDG · AIST++ test | Divk: 8.3189 -> 7.5374 | Hurt |
Surprising findings
- Removing iDDPM lowers PFC quality (0.8014 to 1.1077) but raises Divk, indicating a quality-diversity trade-off in the ablation.
- FACT's high Divg scores are attributed to excessive meaningless swaying, so the authors caution against trusting Divg alone.
Limitations
- Training cost remains relatively high; end-to-end model is heavy.
- Small PopDanceSet test split (24 segments) and user study with only 20 participants.
- Ablation experiments run on half-dimension model on a single 3090, under constrained conditions.
Experimental setup
Datasets: PopDanceSet AIST++
Benchmarks: PopDanceSet AIST++
Metrics: PFC PBC Divk Divg Beat Align Score FID
Baselines: FACT Bailando EDGE
Builds on: iDDPM DDIM EDGE Bailando FACT
Closest papers by meaning
- MEGADance: Mixture-of-Experts Architecture for Genre-Aware 3D Dance Generation (NeurIPS 2025 · similarity 0.87)
- Bidirectional Autoregessive Diffusion Model for Dance Generation (CVPR 2024 · similarity 0.86)
- DanceAnyWay: Synthesizing Beat-Guided 3D Dances with Randomized Temporal Contrastive Learning (AAAI 2024 · similarity 0.86)
- DanceCamera3D: 3D Camera Movement Synthesis with Music and Dance (CVPR 2024 · similarity 0.85)
- Harmonious Music-driven Group Choreography with Trajectory-Controllable Diffusion (AAAI 2025 · similarity 0.84)
- Breaking the Group Size Barrier: Parameter-Efficient Group Dance Generation with Chain-of-Dancers (NeurIPS 2026 · similarity 0.84)
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.