Research map / Video / Video generation
Executing Your Commands via Motion Diffusion in Latent Space
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use latent-space diffusion rather than diffusion on raw motion sequences | Motion generation · HumanML3D (text-to-motion) | FID: 0.544^{\pm.044} -> 0.473^{\pm.013} | Helped |
| Use latent-space diffusion rather than diffusion on raw motion sequences | Motion generation · HumanML3D (text-to-motion, DDIM 50 steps) | Total inference time: 225.28 -> 10.24 | Helped |
| Use a 7-token latent instead of 1 token in the VAE (ablation) | Motion VAE · HumanML3D (motion VAE) | MPJPE: 54.4 -> 14.7 | Helped |
| Add long skip connections to the VAE encoder and decoder (ablation) | Motion VAE · HumanML3D (motion VAE) | MPJPE: 18.5 -> 14.7 | Helped |
| Use 9 VAE transformer layers instead of 7 (ablation) | Motion VAE · HumanML3D (motion VAE) | MPJPE: 16.0 -> 14.7 | Helped |
| Concatenate condition embeddings instead of using cross-attention (ablation) | MLD-1 denoiser · HumanML3D (text-to-motion) | FID: 1.922{^{\pm.041}} -> 0.473{^{\pm.013}} | Helped |
| Add long skip connections to the denoiser (ablation) | MLD-1 denoiser · HumanML3D (text-to-motion) | FID: 0.784{^{\pm.015}} -> 0.473{^{\pm.013}} | Helped |
| Use a VAE instead of an unregularized autoencoder (ablation) | MLD motion representation · HumanML3D (text-to-motion) | FID: 5.033^{\pm.061} -> 0.473^{\pm.013} | Helped |
| Predict noise rather than the clean latent directly (ablation) | MLD-1 denoiser · HumanML3D (text-to-motion) | FID: 0.513^{\pm.011} -> 0.473^{\pm.013} | Helped |
| Use CLIP rather than BERT as text encoder (ablation) | MLD-1 · HumanML3D (text-to-motion) | R Precision Top 3: 0.725^{\pm.002} -> 0.769^{\pm.002} | Helped |
| Increase classifier-free guidance dropout from p=0.10 to p=0.25 (ablation) | MLD-1 · HumanML3D (text-to-motion) | FID: 0.473^{\pm.013} -> 0.229^{\pm.010} | Helped |
| Increase classifier-free guidance dropout from p=0.10 to p=0.25 (ablation) | MLD-1 · HumanML3D (text-to-motion) | R Precision Top 3: 0.772^{\pm.002} -> 0.757^{\pm.002} | Hurt |
Closest papers by meaning
- AMD: Autoregressive Motion Diffusion (AAAI 2024 · similarity 0.92)
- Rethinking Diffusion for Text-Driven Human Motion Generation: Redundant Representations, Evaluation, and Masked Autoregression (CVPR 2025 · similarity 0.91)
- Human Motion Diffusion as a Generative Prior (ICLR 2024 · similarity 0.90)
- MotionMix: Weakly-Supervised Diffusion for Controllable Motion Generation (AAAI 2024 · similarity 0.90)
- Seamless Human Motion Composition with Blended Positional Encodings (CVPR 2024 · similarity 0.89)
- Move-in-2D: 2D-Conditioned Human Motion Generation (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.