Research map / Image generation / Editing and style transfer
Generative Human Motion Stylization in Latent Space
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 a generative latent-space motion stylization framework that decomposes pretrained autoencoder motion codes into deterministic content codes and probabilistic style codes, supporting motion-, label-, and prior-based stylization.
Problem
Human motion stylization should change the style of a motion while keeping its content, but prior methods work in raw pose space, are deterministic or limited in diversity, and generalize poorly across datasets.
Approach
Perform stylization in the latent space of a pretrained motion autoencoder (VAE or AE), with a probabilistic global style code sampled from a Gaussian prior and a deterministic temporal content code, trained with reconstruction, homo-style alignment, swap-cycle reconstruction, and KL losses; plus a small 1D-conv global motion predictor for root velocity.
Key results
- Ours (A) reaches 0.945 style accuracy and 0.344 geodesic distance on the first test set in motion-based supervised stylization, versus 0.891 and 0.531 for the strongest prior baseline.
- Latent stylization improves generalization: on the unseen CMU Mocap set, Ours (V) reaches 0.918 style accuracy vs 0.674 for a GAN baseline.
- The method runs in 4.76 ms per forward pass, about 14 times faster than the most advanced prior baseline (67.563 ms).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use latent stylization rather than direct pose-space stylization (ablation) | Ours (A), supervised motion-based stylization · [cite] test set | style accuracy: 0.932 ± 0.008 -> 0.945 ± 0.007 | Helped |
| Use a probabilistic style space instead of a deterministic one (ablation) | Ours (A), supervised motion-based stylization · [cite] test set | style accuracy: 0.913 ± 0.007 -> 0.945 ± 0.007 | Helped |
| Add homo-style alignment (ablation) | Ours (A), supervised motion-based stylization · [cite] test set | style accuracy: 0.883 ± 0.012 -> 0.945 ± 0.007 | Helped |
| Add homo-style alignment (ablation) | Ours (A), supervised motion-based stylization · [cite] test set | content accuracy: 0.537 ± 0.016 -> 0.674 ± 0.011 | Helped |
| Add autoencoding (ablation) | Ours (A), supervised motion-based stylization · [cite] test set | style accuracy: 0.900 ± 0.010 -> 0.945 ± 0.007 | Helped |
| Add cycle reconstruction (ablation) | Ours (A), supervised motion-based stylization · [cite] test set | style accuracy: 0.917 ± 0.009 -> 0.945 ± 0.007 | Helped |
| Use probabilistic rather than deterministic style space (ablation) | Ours (A), supervised motion-based stylization · [cite] test set | geodesic distance: 0.509 ± 0.004 -> 0.344 ± 0.002 | Helped |
| Make the content space probabilistic (ablation) | Ours (A), supervised motion-based stylization · [cite] test set | style accuracy: 0.945 ± 0.007 -> 0.947 ± 0.001 | Helped |
| Make the content space probabilistic (ablation) | Ours (A), supervised motion-based stylization · [cite] test set | content accuracy: 0.674 ± 0.011 -> 0.417 ± 0.012 | Hurt |
| Train latent extractor and stylization model end-to-end rather than separately (ablation) | Ours (A), supervised motion-based stylization · [cite] test set | style accuracy: 0.945 ± 0.007 -> 0.125 ± 0.010 | Hurt |
| Add global motion prediction (ablation) | Ours (S), motion-based stylization · [cite] test set | foot skating: 0.141 ± 0.001 -> 0.130 ± 0.001 | Helped |
| Add global motion prediction (ablation) | Ours (U), motion-based stylization · [cite] test set | foot skating: 0.116 ± 0.001 -> 0.102 ± 0.001 | Helped |
Surprising findings
- End-to-end training of the latent extractor and stylization model collapses to about 15% stylization accuracy, versus 92% with separate training.
- A probabilistic content space improves in-domain style accuracy but generalizes worse out-of-domain.
Limitations
- Adversarial losses were found to be extremely unstable and were not used.
- Trade-off: stronger homo-style alignment or KL weight reduces diversity; the paper notes limited diversity of the training dataset.
Experimental setup
Datasets: Xia et al. styled motion dataset (193 min, 16 styles) 8-style action-labeled motion dataset (25 min) CMU Mocap
Benchmarks: Xia et al. styled motion dataset test set CMU Mocap test set 8-style action-labeled dataset test set
Metrics: style accuracy style FID content accuracy content FID geodesic distance diversity foot skating MPJPE user preference
Baselines: Aberman et al. 2020 (unpaired motion style transfer) Dong et al. 2020 (GAN-based style transfer) Ours w/o latent Text2Motion generator (stylized text2motion)
Builds on: AdaIN VAE cycle-consistency
Closest papers by meaning
- MMM: Generative Masked Motion Model (CVPR 2024 · similarity 0.87)
- Shape My Moves: Text-Driven Shape-Aware Synthesis of Human Motions (CVPR 2025 · similarity 0.86)
- SnapMoGen: Human Motion Generation from Expressive Texts (NeurIPS 2025 · similarity 0.86)
- MotionMix: Weakly-Supervised Diffusion for Controllable Motion Generation (AAAI 2024 · similarity 0.86)
- Seamless Human Motion Composition with Blended Positional Encodings (CVPR 2024 · similarity 0.86)
- ReMoGPT: Part-Level Retrieval-Augmented Motion-Language Models (AAAI 2025 · 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.