Research map / Video / Video generation
Deep Compositional Phase Diffusion for Long Motion Sequence Generation
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.
Proposes Compositional Phase Diffusion, which denoises phase latents from a pretrained ACT-PAE with semantic (SPDM) and transition-aware (TPDM) diffusion modules to generate smooth composite motion sequences, motion inbetweening, and long motion.
Problem
Generating composite motion sequences from multiple semantic text instructions often yields awkward, discontinuous transitions between motion clips, and existing long-motion models operate in raw pose space with limited transition modeling.
Approach
Operate diffusion in a variable-length periodic phase latent space (ACT-PAE) and denoise each segment jointly with a semantic module (SPDM, text-conditioned) and forward/backward transitional modules (TPDM) that condition on clean phase latents of adjacent segments, combined via a cubic-schedule phase mixing equation.
Key results
- Compositional motion pair generation: overall FID 0.782 and overall MMD 4.711, best among compared methods (priorMDM overall FID 0.839, TEACH 1.041).
- Long-term generation over 302,298 frames (168 min): overall FID 0.847 and overall MMD 4.849, best MMD and second-best FID behind PCMDM (0.876).
- Unconditional motion inbetweening at 60/120/180-frame lengths: best L2-Vel (0.0101/0.0102/0.0125) and RMS-Jerk (0.0963/0.1054/0.1213) across all settings.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Compositional Phase Diffusion instead of the strongest listed baseline | Compositional motion pair generation · BABEL-TEACH test | Overall FID: 0.837 -> 0.782 | Helped |
| Use Compositional Phase Diffusion instead of the strongest listed baseline | Compositional motion pair generation · BABEL-TEACH test | Overall MMD: 4.821 -> 4.711 | Helped |
| Use Compositional Phase Diffusion instead of the strongest listed baseline | Long-term motion generation · BABEL-TEACH test | Overall FID: 0.876 -> 0.847 | Helped |
| Use Compositional Phase Diffusion instead of the strongest listed baseline | Long-term motion generation · BABEL-TEACH test | Overall MMD: 4.984 -> 4.849 | Helped |
| Use Compositional Phase Diffusion instead of the strongest listed baseline | Unconditional motion inbetweening · BABEL-TEACH test, 60, 120, and 180 frames | L2-Vel, L2-Rot6D, NPSS, and RMS-Jerk | Helped |
| Use Compositional Phase Diffusion instead of the strongest listed baseline | Conditional motion inbetweening · BABEL-TEACH test, 60, 120, and 180 frames | Smt. FID and MMD | Helped |
| Add frame-level tokens to SPDM and TPDM (ablation) | SPDM and TPDM · Compositional motion generation and motion inbetweening | Denoising performance | Helped |
| Set phase mixing to r = (k/K)^3 for semantic-conditioned segments (ablation) | Compositional Phase Diffusion · Semantic conditional scenarios | Performance | Helped |
| Set phase mixing to r = 1 for unconditional segments (ablation) | Compositional Phase Diffusion · Unconditional scenarios | Performance | Helped |
Surprising findings
- Single-text MDM-30 with linear blending is competitive on transitional metrics, and the authors note TEACH's low overall FID hides transition artifacts such as root roll changes.
Limitations
- Phase mixing uses a basic fixed schedule rather than a learned or adaptive mechanism.
- Evaluated only on BABEL-TEACH, the only dataset with subsequence pair annotations.
Experimental setup
Datasets: BABEL-TEACH BABEL HumanML3D
Benchmarks: BABEL-TEACH
Metrics: FID MMD L2-Vel L2-Rot6D NPSS RMS-Jerk
Baselines: MDM-30 MLD-30 TEACH PCMDM priorMDM RMIB RSMT CMB
Builds on: DeepPhase ACTOR priorMDM PCMDM MLD
Closest papers by meaning
- Human Motion Diffusion as a Generative Prior (ICLR 2024 · similarity 0.89)
- StoryDiffusion: Consistent Self-Attention for Long-Range Image and Video Generation (NeurIPS 2024 · similarity 0.88)
- MotionMix: Weakly-Supervised Diffusion for Controllable Motion Generation (AAAI 2024 · similarity 0.87)
- Seamless Human Motion Composition with Blended Positional Encodings (CVPR 2024 · similarity 0.87)
- Towards Detailed Text-to-Motion Synthesis via Basic-to-Advanced Hierarchical Diffusion Model (AAAI 2024 · similarity 0.86)
- LaMo: Self-Supervised Latent Motion Priors for Physical Realism in Video Generation (NeurIPS 2026 · 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.