cs.CVFeb 26, 2026

Causal Motion Diffusion Models for Autoregressive Motion Generation

Authors: Qing Yu, Akihisa Watanabe, Kent Fujiwara

Organizations: LY Corporation · Waseda University

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Recent advances in motion diffusion models have substantially improved the realism of human motion synthesis. However, existing approaches either rely on full-sequence diffusion models with bidirectional generation, which limits temporal causality and real-time applicability, or autoregressive models that suffer from instability and cumulative errors. In this work, we present Causal Motion Diffusion Models (CMDM), a unified framework for autoregressive motion generation based on a causal diffusion transformer that operates in a semantically aligned latent space. CMDM builds upon a Motion-Language-Aligned Causal VAE (MAC-VAE), which encodes motion sequences into temporally causal latent representations. On top of this latent representation, an autoregressive diffusion transformer is trained using causal diffusion forcing to perform temporally ordered denoising across motion frames. To achieve fast inference, we introduce a frame-wise sampling schedule with causal uncertainty, where each subsequent frame is predicted from partially denoised previous frames. The resulting framework supports high-quality text-to-motion generation, streaming synthesis, and long-horizon motion generation at interactive rates. Experiments on HumanML3D and SnapMoGen demonstrate that CMDM outperforms existing diffusion and autoregressive models in both semantic fidelity and temporal smoothness, while substantially reducing inference latency.

Figures & tables

Figure 1: Overview of the existing methods and the proposed method. Existing diffusion-based methods (left) perform full-sequence denoising using the same noise level across all frames. In contrast, our proposed CMDM (right) introduces a causal diffusion forcing mechanism that operates on semantic causal latent features with frame-wise noise levels.
Figure 2: Overview of the proposed CMDM framework. CMDM consists of three key components: (a) MAC-VAE, which encodes motion sequences into motion–language–aligned and temporally causal latent features using a causal encoder–decoder structure supervised by motion-language model alignment; (b) Causal-DiT, which performs diffusion denoising with causal self-attention and cross-attention to text embeddings, ensuring temporally ordered and semantically consistent frame refinement; and (c) Causal Diffusion Forcing, which applies independent frame-level noise during training and a causal uncertainty schedule during inference, where the redness intensity represents the noise level. This design enables CMDM to achieve temporally consistent, semantically aligned, and efficient text-to-motion generation suitable for streaming and long-horizon synthesis.
MethodsFrameworkR-Precision ↑\uparrowFID ↓\downarrowMM-Dist ↓\downarrowMModality ↑\uparrowCLIP-score ↑\uparrow
 Top 1Top 2Top 3
GT−-0.511±.0030.511^{\pm.003}0.703±.0030.703^{\pm.003}0.797±.0020.797^{\pm.002}0.002±.0000.002^{\pm.000}2.974±.0082.974^{\pm.008}−-0.639±.0010.639^{\pm.001}
T2M-GPT [ 51 ]VQ0.492±.0030.492^{\pm.003}0.679±.0020.679^{\pm.002}0.775±.0020.775^{\pm.002}0.141±.0050.141^{\pm.005}3.121±.0093.121^{\pm.009}1.831±.0481.831^{\pm.048}0.607±.0050.607^{\pm.005}
MMM [ 33 ]0.515±.0020.515^{\pm.002}0.708±.0020.708^{\pm.002}0.804±.0020.804^{\pm.002}0.089±.0050.089^{\pm.005}2.926±.0072.926^{\pm.007}1.226±.0351.226^{\pm.035}0.635±.0030.635^{\pm.003}
MoMask [ 10 ]0.521±.0020.521^{\pm.002}0.713±.0020.713^{\pm.002}0.807±.0020.807^{\pm.002}0.045±.002\mathbf{0.045^{\pm.002}}2.958±.0082.958^{\pm.008}1.241±.0401.241^{\pm.040}0.637±.003{0.637^{\pm.003}}
MDM-50Steps [ 42 ]Diffusion0.455±.0060.455^{\pm.006}0.645±.0070.645^{\pm.007}0.749±.0020.749^{\pm.002}0.489±.0250.489^{\pm.025}3.330±.0253.330^{\pm.025}2.290±.070\mathbf{2.290^{\pm.070}}0.481±.0010.481^{\pm.001}
Table 1: Results of text-to-motion generation on HumanML3D. The average is reported over 10 runs with 95% confidence intervals. Methods marked with †\dagger were originally implemented with different motion representations and have been re-trained using our codebase to ensure a fair comparison. Bold indicates the best result, while underline denotes the second-best result.
MethodsFrameworkR-Precision ↑\uparrowFID ↓\downarrowMModality ↑\uparrowCLIP-score ↑\uparrow
Top 1Top 2Top 3
GT−-0.940±.0010.940^{\pm.001}0.976±.0010.976^{\pm.001}0.985±.0010.985^{\pm.001}0.001±.0000.001^{\pm.000}−-0.837±.0000.837^{\pm.000}
T2M-GPT [ 51 ]VQ0.618±.0020.618^{\pm.002}0.773±.0020.773^{\pm.002}0.812±.0020.812^{\pm.002}32.629±.08732.629^{\pm.087}9.172±.1819.172^{\pm.181}0.573±.0010.573^{\pm.001}
MoMask [ 10 ]0.777±.0020.777^{\pm.002}0.888±.0020.888^{\pm.002}0.927±.0020.927^{\pm.002}17.404±.05117.404^{\pm.051}8.183±.1848.183^{\pm.184}0.664±.0010.664^{\pm.001}
MoMask ++ [ 9 ]0.802±.0010.802^{\pm.001}0.905±.002{0.905^{\pm.002}}0.938±.001{0.938^{\pm.001}}15.061±.065{15.061^{\pm.065}}7.259±.1807.259^{\pm.180}0.685±.001{0.685^{\pm.001}}
MDM [ 42 ]Diffusion0.503±.0020.503^{\pm.002}0.653±.0020.653^{\pm.002}0.727±.0020.727^{\pm.002}57.783±.09257.783^{\pm.092}13.412±.231\mathbf{13.412^{\pm.231}}0.481±.0010.481^{\pm.001}
Table 2: Results of text-to-motion generation on SnapMoGen. The average is reported over 10 runs with 95% confidence intervals.
 MethodsSubsequenceTransition
 R-Top3 ↑\uparrowFID ↓\downarrowDiv →\rightarrowMM-Dist ↓\downarrowFID ↓\downarrowDiv →\rightarrowPJ →\rightarrowAUJ ↓\downarrow
HML3DGT0.796±0.0040.796^{\pm 0.004}0.00±0.000.00^{\pm 0.00}9.34±0.089.34^{\pm 0.08}2.97±0.012.97^{\pm 0.01}0.00±0.000.00^{\pm 0.00}9.54±0.159.54^{\pm 0.15}0.04±0.000.04^{\pm 0.00}0.07±0.000.07^{\pm 0.00}
FlowMDM [ 3 ]0.685±0.004{0.685^{\pm 0.004}}0.29±0.01{0.29^{\pm 0.01}}9.58±0.12{9.58^{\pm 0.12}}3.61±0.01{3.61^{\pm 0.01}}1.38±0.05\mathbf{1.38^{\pm 0.05}}8.79±0.09\mathbf{8.79^{\pm 0.09}}0.06±0.00‾\underline{0.06^{\pm 0.00}}0.51±0.01‾\underline{0.51^{\pm 0.01}}
MARDM [ 24 ]0.741±0.004‾\underline{{0.741^{\pm 0.004}}}0.24±0.01‾\underline{{0.24^{\pm 0.01}}}9.29±0.12‾\underline{{9.29^{\pm 0.12}}}3.38±0.01‾\underline{3.38^{\pm 0.01}}2.72±0.09{2.72^{\pm 0.09}}8.28±0.118.28^{\pm 0.11}0.07±0.00{0.07^{\pm 0.00}}0.57±0.01{0.57^{\pm 0.01}}
CMDM0.782±0.003\mathbf{0.782^{\pm 0.003}}0.12±0.04\mathbf{0.12^{\pm 0.04}}9.44±0.13\mathbf{9.44^{\pm 0.13}}3.04±0.01\mathbf{3.04^{\pm 0.01}}1.66±0.06‾\underline{1.66^{\pm 0.06}}8.72±0.10‾\underline{8.72^{\pm 0.10}}0.04±0.00\mathbf{0.04^{\pm 0.00}}0.42±0.01\mathbf{0.42^{\pm 0.01}}
SnapMGGT0.997±0.0060.997^{\pm 0.006}0.00±0.000.00^{\pm 0.00}19.74±0.0519.74^{\pm 0.05}14.83±0.0114.83^{\pm 0.01}0.00±0.000.00^{\pm 0.00}19.21±0.0819.21^{\pm 0.08}1.11±0.011.11^{\pm 0.01}45.16±0.3345.16^{\pm 0.33}
Table 3: Results of long-horizon motion generation on HumanML3D and SnapMoGen. The motion quality of each subsequence and the smoothness of each transition are evaluated.
Figure 3: Qualitative results of long-horizon motion generation. Comparison between our CMDM and previous methods. The generated motion is continuous and seamless; for visualization purposes, we split each long sequence into shorter segments corresponding to their captions. Please refer to the videos in the supplementary materials for the complete motion sequences.
MethodsT2M GenerationMotion Transition
R-Top1 ↑\uparrowFID ↓\downarrowFID ↓\downarrowAUJ ↓\downarrow
CMDM0.588±.004‾\underline{0.588^{\pm.004}}0.068±.003\mathbf{0.068^{\pm.003}}1.66±0.06‾\underline{1.66^{\pm 0.06}}0.42±0.01‾\underline{0.42^{\pm 0.01}}
VAE w/o MA0.561±.005{0.561^{\pm.005}}0.107±.006{0.107^{\pm.006}}2.04±0.06{2.04^{\pm 0.06}}0.52±0.01{0.52^{\pm 0.01}}
C-VAE w/o MA0.575±.004{0.575^{\pm.004}}0.070±.004{0.070^{\pm.004}}1.72±0.06{1.72^{\pm 0.06}}0.44±0.01{0.44^{\pm 0.01}}
w/ Full-Seq. Diff.0.591±.004\mathbf{0.591^{\pm.004}}0.071±.003{0.071^{\pm.003}}1.96±0.07{1.96^{\pm 0.07}}0.72±0.01{0.72^{\pm 0.01}}
w/o adaLN0.583±.002{0.583^{\pm.002}}0.076±.004{0.076^{\pm.004}}1.78±0.06{1.78^{\pm 0.06}}0.47±0.01{0.47^{\pm 0.01}}
Table 4: Ablation studies of CMDM.
Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

MethodsSubsequenceTransition
R-prec ↑\uparrowFID ↓\downarrowDiv →\rightarrowMM-Dist ↓\downarrowFID ↓\downarrowDiv →\rightarrowPJ →\rightarrowAUJ ↓\downarrow
GT0.715±0.0030.715^{\pm 0.003}0.00±0.000.00^{\pm 0.00}8.42±0.158.42^{\pm 0.15}3.36±0.063.36^{\pm 0.06}0.00±0.000.00^{\pm 0.00}6.20±0.066.20^{\pm 0.06}0.02±0.000.02^{\pm 0.00}0.00±0.000.00^{\pm 0.00}
FlowMDM0.702±0.004{0.702^{\pm 0.004}}0.99±0.04{0.99^{\pm 0.04}}8.36±0.13{8.36^{\pm 0.13}}3.45±0.02{3.45^{\pm 0.02}}2.61±0.06{2.61^{\pm 0.06}}6.47±0.05\mathbf{6.47^{\pm 0.05}}0.06±0.00{0.06^{\pm 0.00}}0.13±0.00{0.13^{\pm 0.00}}
Ours0.711±0.005\mathbf{0.711^{\pm 0.005}}0.90±0.06\mathbf{0.90^{\pm 0.06}}8.47±0.20\mathbf{8.47^{\pm 0.20}}3.39±0.05\mathbf{3.39^{\pm 0.05}}2.45±0.05\mathbf{2.45^{\pm 0.05}}6.73±0.05{6.73^{\pm 0.05}}0.05±0.00\mathbf{0.05^{\pm 0.00}}0.11±0.01\mathbf{0.11^{\pm 0.01}}
Appendix
Table 5: Comparison of long-horizon motion generation on BABEL. Subsequence metrics evaluate motion quality and diversity within segments, while transition metrics assess temporal continuity and smoothness between segments.
MethodsFrameworkR-Precision ↑\uparrowFID ↓\downarrowMM-Dist ↓\downarrowMModality ↑\uparrowCLIP-score ↑\uparrow
Top 1Top 2Top 3
T2M-GPT [ 51 ]VQ0.470±.0030.470^{\pm.003}0.659±.0020.659^{\pm.002}0.758±.0020.758^{\pm.002}0.335±.0030.335^{\pm.003}3.505±.0173.505^{\pm.017}2.018±.0532.018^{\pm.053}0.607±.0050.607^{\pm.005}
MMM [ 33 ]0.487±.0030.487^{\pm.003}0.683±.0020.683^{\pm.002}0.782±.0020.782^{\pm.002}0.132±.0040.132^{\pm.004}3.359±.0193.359^{\pm.019}2.241±.0732.241^{\pm.073}0.635±.0030.635^{\pm.003}
MoMask [ 10 ]0.490±.0040.490^{\pm.004}0.687±.0030.687^{\pm.003}0.786±.0030.786^{\pm.003}0.116±.006{0.116^{\pm.006}}3.353±.0103.353^{\pm.010}1.263±.0791.263^{\pm.079}0.637±.003{0.637^{\pm.003}}
MDM-50Step [ 42 ]Diffusion0.440±.0070.440^{\pm.007}0.636±.0060.636^{\pm.006}0.742±.0040.742^{\pm.004}0.518±.0320.518^{\pm.032}3.640±.0283.640^{\pm.028}3.604±.031\mathbf{3.604^{\pm.031}}0.578±.0030.578^{\pm.003}
MotionDiffuse [ 42 ]0.450±.0060.450^{\pm.006}0.641±.0050.641^{\pm.005}0.753±.0050.753^{\pm.005}0.778±.0350.778^{\pm.035}3.490±.0233.490^{\pm.023}3.179±.0463.179^{\pm.046}0.606±.0040.606^{\pm.004}
Appendix
Table 6: Results of text-to-motion generation on HumanML3D without redundant features. The average is reported over 10 runs with 95% confidence intervals. Bold indicates the best result, and underline denotes the second-best result.
MethodInput typePer-crop semantic correctnessRealism
#tracks#cropsR@1 ↑\uparrowR@3 ↑\uparrowM2T ↑\uparrowM2M ↑\uparrowFID ↓\downarrowTransition ↓\downarrow
GT--55.055.073.373.30.7480.7481.0001.0000.0000.0001.51.5
EnergyMoGen [ 50 ]SingleSingle15.915.928.028.00.5910.5910.5670.5670.6040.6041.61.6
MDM-SMPL [ 32 ]SingleSingle12.112.123.523.50.5730.5730.5780.5780.4840.4841.81.8
w/ DiffCollage [ 32 ]SingleMulti29.129.149.749.70.6750.6750.6560.6560.4460.4461.21.2
w/ STMC [ 32 ]MultiMulti30.530.550.950.90.6750.6750.6650.6650.4590.4590.9\mathbf{0.9}
Appendix
Table 7: Comparison with prior compositional motion generation methods on the Multi-track timeline (MTT) dataset [ 32 ] .
ModelConfig.ReconstructionGeneration
FID ↓\downarrowMPJPE ↓\downarrowR-Top1 ↑\uparrowFID ↓\downarrowMM-Dist ↓\downarrow
VAE64, 1/40.0010.0010.0160.0160.5610.5610.1070.1072.7062.706
C-VAE64, 1/40.0000.0000.0120.0120.5750.5750.0700.0702.6502.650
MAC-VAE64, 1/40.000\mathbf{0.000}0.012\mathbf{0.012}0.588\mathbf{0.588}0.0680.0682.620\mathbf{2.620}
MAC-VAE32, 1/40.0020.0020.0330.0330.5830.5830.0650.0652.6282.628
MAC-VAE16, 1/40.0110.0110.0770.0770.5730.5730.0710.0712.6472.647
Appendix
Table 8: Comparison of reconstruction and generation performance on HumanML3D. MPJPE is measured in millimeters. The notation (d,r)(d,r) denotes the latent dimension dd and the temporal downsampling rate rr .
ModelReconstructionGeneration
FID ↓\downarrowMPJPE ↓\downarrowR-Top1 ↑\uparrowFID ↓\downarrowMM-Dist ↓\downarrow
VAE0.0010.0010.0160.0160.5610.5610.1070.1072.7062.706
C-VAE0.0010.0010.0120.0120.5750.5750.0700.0702.6502.650
Part-TMR [ 49 ]0.000\mathbf{0.000}0.012\mathbf{0.012}0.588\mathbf{0.588}0.068\mathbf{0.068}2.620\mathbf{2.620}
MotionPatches [ 48 ]0.0000.0000.0130.0130.5860.5860.0700.0702.6222.622
TMR [ 31 ]0.0010.0010.0130.0130.5800.5800.0700.0702.6382.638
Appendix
Table 9: Comparison of motion-language models in MAC-VAE on HumanML3D. MPJPE is measured in millimeters.
Model SizeConfig.R-Precision ↑\uparrowFID ↓\downarrowMM-Dist ↓\downarrow
Top1Top2Top3
S (19M)H2, L4, D5120.5430.5430.7380.7380.8340.8340.2470.2472.8452.845
M (38M)H4, L8, D5120.5880.5880.7780.7780.8600.8600.0680.0682.6202.620
L (129M)H6, L12, D7680.5850.5850.7790.7790.8590.8590.0440.0442.6212.621
XL (304M)H8, L16, D10240.590\mathbf{0.590}0.779\mathbf{0.779}0.861\mathbf{0.861}0.042\mathbf{0.042}2.610\mathbf{2.610}
Appendix
Table 10: Comparison of model sizes on HumanML3D. The notation (H,L,D)(H,L,D) denotes the number of attention heads HH , layers LL , and hidden dimension DD .
Text EncoderEmbeddingR-Precision ↑\uparrowFID ↓\downarrowMM-Dist ↓\downarrow
Top1Top2Top3
DistilBERT [ 37 ]Word0.588\mathbf{0.588}0.778\mathbf{0.778}0.860\mathbf{0.860}0.068\mathbf{0.068}2.620\mathbf{2.620}
CLIP [ 35 ]Word0.5560.5560.7510.7510.8430.8430.0860.0862.7172.717
CLIP [ 35 ]Sentence0.5270.5270.7170.7170.8090.8090.1450.1452.9412.941
Sentence-T5 [ 25 ]Sentence0.5640.5640.7540.7540.8410.8410.1260.1262.7372.737
Appendix
Table 11: Comparison of text encoders on HumanML3D.
Figure 4: Qualitative results of long-horizon motion generation on HumanML3D. Comparison between our CMDM and previous methods. The generated motion is continuous and seamless; for visualization purposes, we split each long sequence into shorter segments corresponding to their captions. Please refer to the videos in the supplementary materials for the complete motion sequences.
Figure 5: Qualitative results of long-horizon motion generation on SnapMoGen. Comparison between our CMDM and previous methods. The generated motion is continuous and seamless; for visualization purposes, we split each long sequence into shorter segments corresponding to their captions. Please refer to the videos in the supplementary materials for the complete motion sequences.
Figure 6: Qualitative results of text-to-motion generation on HumanML3D. CMDM produces motions that better capture fine-grained textual semantics and maintain natural body articulation compared to previous methods. Please refer to the supplementary videos for clearer visualization.
Figure 7: Qualitative results of text-to-motion generation on SnapMoGen. Comparison between our CMDM and previous methods. We directly use the raw text prompts without any LLM-based augmentation and CMDM still achieves strong generation quality. Please refer to the supplementary videos for clearer visualization.

Explore similar work

CardsList

  1. RDM: Recurrent Diffusion Model for Human Motion Generation

    Jun 11, 2024Mirgahney Mohamed, Harry Jake Cunningham, Riccardo Barbano +2Diffusion Models for Motion GenerationDiffusion Models

    Department of Computer Science, University College London

  2. Coordinate-Based Dual-Constrained Autoregressive Motion Generation

    Apr 9, 2026Kang Ding, Hongsong Wang, Jie Gui +1Text-to-Motion GenerationHuman Motion Generation

    School of Cyber Science and Engineering, Southeast University, Nanjing 210096, China · School of Computer Science and Engineering, Key Laboratory of New Generation Artificial Intelligence Technology and Its Interdisciplinary Applications, Ministry of Education, Southeast University, Nanjing 210096, China · School of Cyber Science and Engineering, Southeast University, also with Purple Mountain Laboratories, Nanjing 211111, China, and also with Engineering Research Center of Blockchain Application, Supervision And Management (Southeast University), Ministry of Education, Southeast University, Nanjing 210096, China +1

  3. Reconstruction-Anchored Diffusion Model for Text-to-Motion Generation

    Jan 21, 2026Yifei Liu, Changxing Ding, Ling Guo +2Diffusion Model GuidanceText-to-Motion Generation

    South China University of Technology · Joy Future Academy