Research map / Video / Video generation
MotionLCM: Real-time Controllable Motion Generation via Latent Consistency Model
Wenxun Dai, Ling-Hao Chen, Jingbo Wang, Jinpeng Liu, Bo Dai, Yansong Tang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| latent consistency distillation of MLD to 1-step inference | MLD · HumanML3D text-to-motion | AITS (s): 0.225 -> 0.030 | Improved |
| latent consistency distillation of MLD to 1-step inference | MLD · HumanML3D text-to-motion | FID: 0.450 -> 0.467 | No clear effect |
| increase MotionLCM inference steps from 1 to 4 | MotionLCM · HumanML3D text-to-motion | FID: 0.467 -> 0.304 | Improved |
| dynamic training guidance scale [5,15] instead of static 7.5 (ablation) | MotionLCM · HumanML3D text-to-motion (1-step) | FID: 0.479 -> 0.467 | Improved |
| widen training guidance range to [2,18] (ablation) | MotionLCM · HumanML3D text-to-motion (1-step) | FID: 0.467 -> 0.481 | Hurt |
| use EMA rate 0.95 instead of 0 (ablation) | MotionLCM · HumanML3D text-to-motion (1-step) | FID: 0.505 -> 0.467 | Improved |
| use skipping interval k=20 instead of k=1 (ablation) | MotionLCM · HumanML3D text-to-motion (1-step) | FID: 0.635 -> 0.467 | Improved |
| increase skipping interval to k=50 (ablation) | MotionLCM · HumanML3D text-to-motion (1-step) | FID: 0.467 -> 0.547 | Hurt |
| use Huber loss instead of L2 loss (ablation) | MotionLCM · HumanML3D text-to-motion (1-step) | FID: 0.622 -> 0.467 | Improved |
| add motion ControlNet (ablation) | MotionLCM · HumanML3D motion control (1-step) | Loc. err.: 0.2302 -> 0.0344 | Improved |
| add motion-space control loss (lambda=1) (ablation) | MotionLCM ControlNet · HumanML3D motion control (1-step) | Loc. err.: 0.0344 -> 0.0147 | Improved |
| increase control loss weight to 10 (ablation) | MotionLCM ControlNet · HumanML3D motion control (1-step) | FID: FID 0.419 -> 0.636; Loc. err. 0.0147 -> 0.0097 | Mixed |
Closest papers by meaning
- MotionMix: Weakly-Supervised Diffusion for Controllable Motion Generation (AAAI 2024 · similarity 0.87)
- DartControl: A Diffusion-Based Autoregressive Motion Model for Real-Time Text-Driven Motion Control (ICLR 2025 · similarity 0.87)
- MotionGPT: Finetuned LLMs Are General-Purpose Motion Generators (AAAI 2024 · similarity 0.87)
- Motion Prompting: Controlling Video Generation with Motion Trajectories (CVPR 2025 · similarity 0.87)
- Human Motion Diffusion as a Generative Prior (ICLR 2024 · similarity 0.86)
- AMD: Autoregressive Motion Diffusion (AAAI 2024 · 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.
