Research map / Image generation / Autoregressive and latent generation
Emotional Speech-driven 3D Body Animation via Disentangled Latent Diffusion
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 AMUSE, a latent-diffusion model that disentangles speech into content, emotion, and style latents to generate emotional 3D body gestures from speech with user-controllable emotion and style.
Problem
Speech-driven 3D body gesture synthesis ignores explicit control over the emotion expressed by the generated gestures.
Approach
Disentangle driving speech into separate content, emotion, and personal-style latent vectors via a DeiT-based audio autoencoder, then condition a latent diffusion denoiser over a VAE motion prior on them, so emotion and style can be swapped by recombining latents from different audio clips.
Key results
- AMUSE achieves FGD 388.63 versus 762.15 for TalkSHOW and best scores on BA (0.81), Div (25.06), and GA (46.76) in Table 1.
- Speech disentanglement model reaches 91.53% emotion accuracy and 96.06% style accuracy on the test set.
- In a perceptual study on Amazon Mechanical Turk, AMUSE is preferred over baselines for both emotion appropriateness and speech synchronization by a considerable margin, though a gap to ground truth remains.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use AMUSE instead of TalkSHOW-BEAT | speech-driven gesture generation · BEAT test | SRGR: 0.31 -> 0.36 | Helped |
| Use AMUSE instead of TalkSHOW-BEAT | speech-driven gesture generation · BEAT test | beat align (BA): 0.64 -> 0.81 | Helped |
| Use AMUSE instead of TalkSHOW | speech-driven gesture generation · BEAT test | FGD: 762.15 -> 388.63 | Helped |
| Use AMUSE instead of TalkSHOW-BEAT | speech-driven gesture generation · BEAT test | Div: 24.16 -> 25.06 | Helped |
| Use AMUSE instead of TalkSHOW | speech-driven gesture generation · BEAT test | gesture emotion accuracy (GA): 29.41 -> 46.76 | Helped |
| Edit emotion using a different audio's emotion latent | AMUSE · BEAT test, emotion editing | beat align (BA): baseline methods: 0.23–0.31; AMUSE-EmoEdit: 0.79 | Helped |
| Edit emotion using a different audio's emotion latent | AMUSE · BEAT test, emotion editing | gesture emotion accuracy (GA): baseline methods: 14.17–29.41; AMUSE-EmoEdit: 34.18 | Helped |
Surprising findings
- Ground truth motions have higher diversity and beat alignment than AMUSE, and are easier to classify, highlighting the difficulty of the task.
Limitations
- Only upper-body motion is modeled; lower-body and locomotion are excluded.
- Does not model semantic gestures (deictic, metaphoric) since no text modality is used.
- Face expressions are omitted; the model is trained separately from facial animation and not trained end-to-end.
Experimental setup
Datasets: BEAT
Benchmarks: BEAT
Metrics: FGD Diversity Gesture emotion accuracy Beat align SRGR emotion classification accuracy F1 perceptual preference
Baselines: TalkSHOW TalkSHOW-BEAT DiffuseStyleGesture MoGlow CaMN Habibie et al.
Builds on: Latent diffusion models DeiT TalkSHOW BEAT SMPL-X
Closest papers by meaning
- Weakly-Supervised Emotion Transition Learning for Diverse 3D Co-speech Gesture Generation (CVPR 2024 · similarity 0.90)
- Chain of Generation: Multi-Modal Gesture Synthesis via Cascaded Conditional Control (AAAI 2024 · similarity 0.89)
- DiffSHEG: A Diffusion-Based Approach for Real-Time Speech-driven Holistic 3D Expression and Gesture Generation (CVPR 2024 · similarity 0.89)
- Co-Speech Gesture Video Generation via Motion-Decoupled Diffusion Model (CVPR 2024 · similarity 0.88)
- DEEPTalk: Dynamic Emotion Embedding for Probabilistic Speech-Driven 3D Face Animation (AAAI 2025 · similarity 0.88)
- The Language of Motion: Unifying Verbal and Non-verbal Language of 3D Human Motion (CVPR 2025 · similarity 0.88)
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.