CC-MotionGen Model Architecture Reference

1. [Architecture Overview](#architecture-overview) 2. [GaussianDiffusion](#gaussiandiffusion) 3. [UNet1D](#unet1d) 4. [MotionDecoder](#motiondecoder) 5. [Conditioning System](#conditioning-system) 6. [Motion Representation](#motion-representation) 7. [Inference Pipeline](#inference-pipeline) | Component | Parameters | Location | Purpose | |-----------|------------|----------|---------| | GaussianDiffusion | - | `model/diffusion.py` | Noise scheduling, DDPM/DDIM sampling | | UNet1D | 116M | `model/unet.py` | Denoising network | | MotionDecoder | 2M | `model/decoder.py` | Semantic motion mapping | | AudioConditioner | ~2M | `model/conditioning.py` | FiLM conditioning | | MPMSConditioner | ~0.5M | `model/conditioning.py` | Memory context blending | The diffusion process that defines the forward (noising) and reverse (denoising) processes. | Index | Name | Dimension | Description | Range | |-------|------|-----------|-------------|-------| | 0-2 | position | 3 | World-space position (x, y, z) | [-50, 50] | | 3-5 | velocity | 3 | Linear velocity (derived) | [-50, 50] | | 6-8 | acceleration | 3 | Linear acceleration (derived) | [-100, 100] | | 9-12 | quaternion | 4 | Orientation (w, x, y, z) | Unit sphere | | 13-15 | angular_velocity | 3 | Rotational velocity | [-10, 10] | | 16 | phase | 1 | Beat-aligned phase | [0, 1] per beat | | 17-24 | style | 8 | Learned style embedding | Unit sphere | | Property | Constraint | Threshold | |----------|------------|-----------| | Velocity coherence | \|v - dp/dt\| | < 5.0 | | Acceleration coherence | \|a - dv/dt\| | < 50.0 | | Jerk bound | \|d³p/dt³\| | < 50000 | | Quaternion continuity | q_t · q_{t+1} | > 0.95 | | Phase monotonicity | phase_{t+1} >= phase_t | Always |