TCDiff++: Breaking Barriers in Multi-Person AI Dance Generation

---

Overview

Xin Zhiyuan Report

When the metaverse’s digital humans need advanced “group dance skills,” traditional music-driven generation technologies hit serious bottlenecks:

• Dancer collisions
• Stiff, unnatural movements
• Failures in long-sequence choreography

Researchers from NUST, Tsinghua University, and Nanjing University have jointly developed TCDiff++, an end-to-end model that achieves high-quality, long-duration, multi-dancer automatic choreography without collisions or foot sliding.

TCDiff++ supports cross-modal choreography — one click generates harmonious group dances, offering a complete AIGC workflow for virtual concerts, theatre, and metaverse events.

This is an upgrade from the AAAI 2025 open-source TCDiff model and has been accepted by IJCV 2025.

---

Challenges in AI Group Dance Generation

1. High Similarity in Motion Data

• In typical datasets, 80%+ of movements are nearly identical.
• Each dancer’s motion data: 100+ dimensions, but position data is only 3D.
• Result: AI confuses dancer identities → leads to collisions.

2. Foot Sliding Phenomenon

• AI struggles to synchronize upper-body motion and foot placement.
• Foot sliding breaks realism and immersion.

3. Long-Sequence Instability

• Short clips (few seconds) are easy.
• Long sequences (minutes): mutations, stutters, positional drift.
• Real-world performances last much longer — current models fail here.

---

From TCDiff to TCDiff++

In TCDiff (AAAI 2025), researchers introduced the trajectory-controllable concept:

• Two-stage framework: separate trajectory prediction from action generation.
• Worked for collision prevention, but caused rigid transitions and jitter in long sequences.

TCDiff++ solves these with a fully end-to-end architecture.

Core Components:

• Group Dance Decoder
• Generates collision-free coordinated movements from music.
• Footwork Adaptor
• Refines foot trajectories, eliminates sliding, ensures grounded footwork.

---

Links:

📄 Paper: arxiv.org/pdf/2506.18671

🌐 Project: da1yuqin.github.io/TCDiffpp.website

💻 Code: github.com/Da1yuqin/TCDiffpp

---

System Workflow

• Step 1: Group Dance Decoder → initial movements, no collisions.
• Step 2: Footwork Adaptor → optimizes feet, removes sliding.
• Step 3: Integration → stable, realistic dance sequences.

---

Key Innovations

1. Collision Prevention

• Dance Positioning Embedding — encodes dancer positions (left/right) in formation.
• Fusion Projection Module — boosts feature dimensions, improves dancer distinction.
• Global Distance Constraint — maintains safe spacing between dancers.

---

2. Precise Footwork

• Swap Mode Conditioning — guides realistic foot movement from start.
• Footwork Adaptor — uses heel/toe contact & root bone velocity to refine steps.

---

3. Long-Sequence Optimization

• Long Group Diffusion Sampling — segmented generation with half-segment overlap.
• Preserves positional continuity and motion smoothness.

---

Performance Comparisons

Model Benchmarks

Why TCDiff++ Excels:

• Best multi-dancer coordination
• High solo realism & diversity
• Strong long-duration stability

---

Other Models’ Limitations:

• EDGE: Can’t distinguish dancers → sliding & collisions.
• GCD: Ignores coordinate modeling → severe sliding.
• CoDancers: Improves ID clarity but loses formation harmony.
• TCDiff: Better formation but mismatched actions & positions.

---

Long-Sequence Capability Test

• EDGE / GCD: Abrupt position swaps.
• CoDancers: Poor formations.
• TCDiff: Cumulative position–motion errors.
• TCDiff++: Maintains position–movement consistency for 720-frame sequences.

---

Ablation Study

• All modules reduce collisions & sliding.
• Maximum performance achieved when all modules combined.

---

User Preference Survey

Criteria:

• Movement realism
• Music–movement correlation
• Formation aesthetics
• Dancer harmony

Result: TCDiff++ most visually preferred.

---

Limitations & Future Directions

• Currently music-only input
• No support for text input, keyframes, or style parameters.
• Future plans: rich multimodal controls for interactive use.
• Formation change learning (e.g., swaps)
• Dataset scarcity of swap motions & annotations.
• Need larger, richer datasets for dynamic formation choreography.

---

Integration with AI Content Ecosystems

TCDiff++ can be integrated with open-source platforms like AiToEarn, enabling:

• AI-generated choreography
• Cross-platform publishing (Douyin, Bilibili, YouTube, Instagram, etc.)
• Analytics & monetization

---

References:

📄 https://arxiv.org/pdf/2506.18671

---

If you want, I can also prepare a short, visually-optimized summary version of this TCDiff++ feature set for quick reading, which would be ideal for GitHub README or project landing pages. Would you like me to do that?