Causal Motion Diffusion Models for Autoregressive Motion Generation

Qing Yu; Akihisa Watanabe; Kent Fujiwara

cs.CVFeb 26, 2026

Causal Motion Diffusion Models for Autoregressive Motion Generation

Authors: Qing Yu, Akihisa Watanabe, Kent Fujiwara

Organizations: LY Corporation · Waseda University

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.

Methods	Framework	R-Precision $\uparrow$			FID $\downarrow$	MM-Dist $\downarrow$	MModality $\uparrow$	CLIP-score $\uparrow$
	Framework	Top 1	Top 2	Top 3	FID $\downarrow$	MM-Dist $\downarrow$	MModality $\uparrow$	CLIP-score $\uparrow$
GT	$-$	$0.511^{\pm.003}$	$0.703^{\pm.003}$	$0.797^{\pm.002}$	$0.002^{\pm.000}$	$2.974^{\pm.008}$	$-$	$0.639^{\pm.001}$
T2M-GPT [ 51 ]	VQ	$0.492^{\pm.003}$	$0.679^{\pm.002}$	$0.775^{\pm.002}$	$0.141^{\pm.005}$	$3.121^{\pm.009}$	$1.831^{\pm.048}$	$0.607^{\pm.005}$
MMM [ 33 ]		$0.515^{\pm.002}$	$0.708^{\pm.002}$	$0.804^{\pm.002}$	$0.089^{\pm.005}$	$2.926^{\pm.007}$	$1.226^{\pm.035}$	$0.635^{\pm.003}$
MoMask [ 10 ]		$0.521^{\pm.002}$	$0.713^{\pm.002}$	$0.807^{\pm.002}$	$\mathbf{0.045^{\pm.002}}$	$2.958^{\pm.008}$	$1.241^{\pm.040}$	${0.637^{\pm.003}}$
MDM-50Steps [ 42 ]	Diffusion	$0.455^{\pm.006}$	$0.645^{\pm.007}$	$0.749^{\pm.002}$	$0.489^{\pm.025}$	$3.330^{\pm.025}$	$\mathbf{2.290^{\pm.070}}$	$0.481^{\pm.001}$

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.

Methods	Framework	R-Precision $\uparrow$			FID $\downarrow$	MModality $\uparrow$	CLIP-score $\uparrow$
Methods	Framework	Top 1	Top 2	Top 3	FID $\downarrow$	MModality $\uparrow$	CLIP-score $\uparrow$
GT	$-$	$0.940^{\pm.001}$	$0.976^{\pm.001}$	$0.985^{\pm.001}$	$0.001^{\pm.000}$	$-$	$0.837^{\pm.000}$
T2M-GPT [ 51 ]	VQ	$0.618^{\pm.002}$	$0.773^{\pm.002}$	$0.812^{\pm.002}$	$32.629^{\pm.087}$	$9.172^{\pm.181}$	$0.573^{\pm.001}$
MoMask [ 10 ]		$0.777^{\pm.002}$	$0.888^{\pm.002}$	$0.927^{\pm.002}$	$17.404^{\pm.051}$	$8.183^{\pm.184}$	$0.664^{\pm.001}$
MoMask ++ [ 9 ]		$0.802^{\pm.001}$	${0.905^{\pm.002}}$	${0.938^{\pm.001}}$	${15.061^{\pm.065}}$	$7.259^{\pm.180}$	${0.685^{\pm.001}}$
MDM [ 42 ]	Diffusion	$0.503^{\pm.002}$	$0.653^{\pm.002}$	$0.727^{\pm.002}$	$57.783^{\pm.092}$	$\mathbf{13.412^{\pm.231}}$	$0.481^{\pm.001}$

Table 2: Results of text-to-motion generation on SnapMoGen. The average is reported over 10 runs with 95% confidence intervals.

	Methods	Subsequence				Transition
	Methods	R-Top3 $\uparrow$	FID $\downarrow$	Div $\rightarrow$	MM-Dist $\downarrow$	FID $\downarrow$	Div $\rightarrow$	PJ $\rightarrow$	AUJ $\downarrow$
HML3D	GT	$0.796^{\pm 0.004}$	$0.00^{\pm 0.00}$	$9.34^{\pm 0.08}$	$2.97^{\pm 0.01}$	$0.00^{\pm 0.00}$	$9.54^{\pm 0.15}$	$0.04^{\pm 0.00}$	$0.07^{\pm 0.00}$
	FlowMDM [ 3 ]	${0.685^{\pm 0.004}}$	${0.29^{\pm 0.01}}$	${9.58^{\pm 0.12}}$	${3.61^{\pm 0.01}}$	$\mathbf{1.38^{\pm 0.05}}$	$\mathbf{8.79^{\pm 0.09}}$	$\underline{0.06^{\pm 0.00}}$	$\underline{0.51^{\pm 0.01}}$
	MARDM [ 24 ]	$\underline{{0.741^{\pm 0.004}}}$	$\underline{{0.24^{\pm 0.01}}}$	$\underline{{9.29^{\pm 0.12}}}$	$\underline{3.38^{\pm 0.01}}$	${2.72^{\pm 0.09}}$	$8.28^{\pm 0.11}$	${0.07^{\pm 0.00}}$	${0.57^{\pm 0.01}}$
	CMDM	$\mathbf{0.782^{\pm 0.003}}$	$\mathbf{0.12^{\pm 0.04}}$	$\mathbf{9.44^{\pm 0.13}}$	$\mathbf{3.04^{\pm 0.01}}$	$\underline{1.66^{\pm 0.06}}$	$\underline{8.72^{\pm 0.10}}$	$\mathbf{0.04^{\pm 0.00}}$	$\mathbf{0.42^{\pm 0.01}}$
SnapMG	GT	$0.997^{\pm 0.006}$	$0.00^{\pm 0.00}$	$19.74^{\pm 0.05}$	$14.83^{\pm 0.01}$	$0.00^{\pm 0.00}$	$19.21^{\pm 0.08}$	$1.11^{\pm 0.01}$	$45.16^{\pm 0.33}$

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.

Methods	T2M Generation		Motion Transition
Methods	R-Top1 $\uparrow$	FID $\downarrow$	FID $\downarrow$	AUJ $\downarrow$
CMDM	$\underline{0.588^{\pm.004}}$	$\mathbf{0.068^{\pm.003}}$	$\underline{1.66^{\pm 0.06}}$	$\underline{0.42^{\pm 0.01}}$
VAE w/o MA	${0.561^{\pm.005}}$	${0.107^{\pm.006}}$	${2.04^{\pm 0.06}}$	${0.52^{\pm 0.01}}$
C-VAE w/o MA	${0.575^{\pm.004}}$	${0.070^{\pm.004}}$	${1.72^{\pm 0.06}}$	${0.44^{\pm 0.01}}$
w/ Full-Seq. Diff.	$\mathbf{0.591^{\pm.004}}$	${0.071^{\pm.003}}$	${1.96^{\pm 0.07}}$	${0.72^{\pm 0.01}}$
w/o adaLN	${0.583^{\pm.002}}$	${0.076^{\pm.004}}$	${1.78^{\pm 0.06}}$	${0.47^{\pm 0.01}}$

Table 4: Ablation studies of CMDM.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Methods	Subsequence				Transition
Methods	R-prec $\uparrow$	FID $\downarrow$	Div $\rightarrow$	MM-Dist $\downarrow$	FID $\downarrow$	Div $\rightarrow$	PJ $\rightarrow$	AUJ $\downarrow$
GT	$0.715^{\pm 0.003}$	$0.00^{\pm 0.00}$	$8.42^{\pm 0.15}$	$3.36^{\pm 0.06}$	$0.00^{\pm 0.00}$	$6.20^{\pm 0.06}$	$0.02^{\pm 0.00}$	$0.00^{\pm 0.00}$
FlowMDM	${0.702^{\pm 0.004}}$	${0.99^{\pm 0.04}}$	${8.36^{\pm 0.13}}$	${3.45^{\pm 0.02}}$	${2.61^{\pm 0.06}}$	$\mathbf{6.47^{\pm 0.05}}$	${0.06^{\pm 0.00}}$	${0.13^{\pm 0.00}}$
Ours	$\mathbf{0.711^{\pm 0.005}}$	$\mathbf{0.90^{\pm 0.06}}$	$\mathbf{8.47^{\pm 0.20}}$	$\mathbf{3.39^{\pm 0.05}}$	$\mathbf{2.45^{\pm 0.05}}$	${6.73^{\pm 0.05}}$	$\mathbf{0.05^{\pm 0.00}}$	$\mathbf{0.11^{\pm 0.01}}$

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.

Methods	Framework	R-Precision $\uparrow$			FID $\downarrow$	MM-Dist $\downarrow$	MModality $\uparrow$	CLIP-score $\uparrow$
Methods	Framework	Top 1	Top 2	Top 3	FID $\downarrow$	MM-Dist $\downarrow$	MModality $\uparrow$	CLIP-score $\uparrow$
T2M-GPT [ 51 ]	VQ	$0.470^{\pm.003}$	$0.659^{\pm.002}$	$0.758^{\pm.002}$	$0.335^{\pm.003}$	$3.505^{\pm.017}$	$2.018^{\pm.053}$	$0.607^{\pm.005}$
MMM [ 33 ]		$0.487^{\pm.003}$	$0.683^{\pm.002}$	$0.782^{\pm.002}$	$0.132^{\pm.004}$	$3.359^{\pm.019}$	$2.241^{\pm.073}$	$0.635^{\pm.003}$
MoMask [ 10 ]		$0.490^{\pm.004}$	$0.687^{\pm.003}$	$0.786^{\pm.003}$	${0.116^{\pm.006}}$	$3.353^{\pm.010}$	$1.263^{\pm.079}$	${0.637^{\pm.003}}$
MDM-50Step [ 42 ]	Diffusion	$0.440^{\pm.007}$	$0.636^{\pm.006}$	$0.742^{\pm.004}$	$0.518^{\pm.032}$	$3.640^{\pm.028}$	$\mathbf{3.604^{\pm.031}}$	$0.578^{\pm.003}$
MotionDiffuse [ 42 ]		$0.450^{\pm.006}$	$0.641^{\pm.005}$	$0.753^{\pm.005}$	$0.778^{\pm.035}$	$3.490^{\pm.023}$	$3.179^{\pm.046}$	$0.606^{\pm.004}$

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.

Method	Input type		Per-crop semantic correctness				Realism
Method	#tracks	#crops	R@1 $\uparrow$	R@3 $\uparrow$	M2T $\uparrow$	M2M $\uparrow$	FID $\downarrow$	Transition $\downarrow$
GT	-	-	$55.0$	$73.3$	$0.748$	$1.000$	$0.000$	$1.5$
EnergyMoGen [ 50 ]	Single	Single	$15.9$	$28.0$	$0.591$	$0.567$	$0.604$	$1.6$
MDM-SMPL [ 32 ]	Single	Single	$12.1$	$23.5$	$0.573$	$0.578$	$0.484$	$1.8$
w/ DiffCollage [ 32 ]	Single	Multi	$29.1$	$49.7$	$0.675$	$0.656$	$0.446$	$1.2$
w/ STMC [ 32 ]	Multi	Multi	$30.5$	$50.9$	$0.675$	$0.665$	$0.459$	$\mathbf{0.9}$

Appendix

Table 7: Comparison with prior compositional motion generation methods on the Multi-track timeline (MTT) dataset [ 32 ] .

Model	Config.	Reconstruction		Generation
Model	Config.	FID $\downarrow$	MPJPE $\downarrow$	R-Top1 $\uparrow$	FID $\downarrow$	MM-Dist $\downarrow$
VAE	64, 1/4	$0.001$	$0.016$	$0.561$	$0.107$	$2.706$
C-VAE	64, 1/4	$0.000$	$0.012$	$0.575$	$0.070$	$2.650$
MAC-VAE	64, 1/4	$\mathbf{0.000}$	$\mathbf{0.012}$	$\mathbf{0.588}$	$0.068$	$\mathbf{2.620}$
MAC-VAE	32, 1/4	$0.002$	$0.033$	$0.583$	$0.065$	$2.628$
MAC-VAE	16, 1/4	$0.011$	$0.077$	$0.573$	$0.071$	$2.647$

Appendix

Table 8: Comparison of reconstruction and generation performance on HumanML3D. MPJPE is measured in millimeters. The notation

(d,r)

denotes the latent dimension

d

and the temporal downsampling rate

r

Model	Reconstruction		Generation
Model	FID $\downarrow$	MPJPE $\downarrow$	R-Top1 $\uparrow$	FID $\downarrow$	MM-Dist $\downarrow$
VAE	$0.001$	$0.016$	$0.561$	$0.107$	$2.706$
C-VAE	$0.001$	$0.012$	$0.575$	$0.070$	$2.650$
Part-TMR [ 49 ]	$\mathbf{0.000}$	$\mathbf{0.012}$	$\mathbf{0.588}$	$\mathbf{0.068}$	$\mathbf{2.620}$
MotionPatches [ 48 ]	$0.000$	$0.013$	$0.586$	$0.070$	$2.622$
TMR [ 31 ]	$0.001$	$0.013$	$0.580$	$0.070$	$2.638$

Appendix

Table 9: Comparison of motion-language models in MAC-VAE on HumanML3D. MPJPE is measured in millimeters.

Model Size	Config.	R-Precision $\uparrow$			FID $\downarrow$	MM-Dist $\downarrow$
Model Size	Config.	Top1	Top2	Top3	FID $\downarrow$	MM-Dist $\downarrow$
S (19M)	H2, L4, D512	$0.543$	$0.738$	$0.834$	$0.247$	$2.845$
M (38M)	H4, L8, D512	$0.588$	$0.778$	$0.860$	$0.068$	$2.620$
L (129M)	H6, L12, D768	$0.585$	$0.779$	$0.859$	$0.044$	$2.621$
XL (304M)	H8, L16, D1024	$\mathbf{0.590}$	$\mathbf{0.779}$	$\mathbf{0.861}$	$\mathbf{0.042}$	$\mathbf{2.610}$

Appendix

Table 10: Comparison of model sizes on HumanML3D. The notation

(H,L,D)

denotes the number of attention heads

H

, layers

L

, and hidden dimension

D

Text Encoder	Embedding	R-Precision $\uparrow$			FID $\downarrow$	MM-Dist $\downarrow$
Text Encoder	Embedding	Top1	Top2	Top3	FID $\downarrow$	MM-Dist $\downarrow$
DistilBERT [ 37 ]	Word	$\mathbf{0.588}$	$\mathbf{0.778}$	$\mathbf{0.860}$	$\mathbf{0.068}$	$\mathbf{2.620}$
CLIP [ 35 ]	Word	$0.556$	$0.751$	$0.843$	$0.086$	$2.717$
CLIP [ 35 ]	Sentence	$0.527$	$0.717$	$0.809$	$0.145$	$2.941$
Sentence-T5 [ 25 ]	Sentence	$0.564$	$0.754$	$0.841$	$0.126$	$2.737$

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

Jun 11, 2024

RDM: Recurrent Diffusion Model for Human Motion Generation

Human motion generation is a challenging task due to its high dimensionality and the difficulty of generating fine-grained motions. Diffusion methods have been proposed due to their high sample quality and expressiveness. Early approaches treat the entire sequence as a whole, which is computationally expensive and restricts sequence length. In contrast, autoregressive diffusion models generate longer sequences. However, their reliance on fully denoising previous frames complicates training and inference. Consequently, we propose RDM, a new recurrent diffusion formulation inspired by Recurrent Neural Networks (RNNs). RDM explicitly conditions the diffusion process on preceding noisy frames, avoiding the cost of full denoising. To overcome the challenge of maintaining its probabilistic nature, we employ Normalizing Flows to model recurrent connections. Our evaluations demonstrate RDM's effectiveness: it achieves comparable performance to autoregressive baselines and generates long sequences that remain aligned with the text. RDM also skips diffusion steps during inference, reducing computational cost by up to 18x over autoregressive baselines. Project page: https://www.mirgahney.com/rfdm.io/.

Mirgahney Mohamed, Harry Jake Cunningham, Riccardo Barbano +2

Department of Computer Science, University College London

Diffusion Models for Motion Generation Diffusion Models Text-to-Motion Generation Diffusion Model Inference Acceleration

Ask PDF

Apr 9, 2026

Coordinate-Based Dual-Constrained Autoregressive Motion Generation

Text-to-motion generation has attracted increasing attention in the research community recently, with potential applications in animation, virtual reality, robotics, and human-computer interaction. Diffusion and autoregressive models are two popular and parallel research directions for text-to-motion generation. However, diffusion models often suffer from error amplification during noise prediction, while autoregressive models exhibit mode collapse due to motion discretization. To address these limitations, we propose a flexible, high-fidelity, and semantically faithful text-to-motion framework, named Coordinate-based Dual-constrained Autoregressive Motion Generation (CDAMD). With motion coordinates as input, CDAMD follows the autoregressive paradigm and leverages diffusion-inspired multi-layer perceptrons to enhance the fidelity of predicted motions. Furthermore, a Dual-Constrained Causal Mask is introduced to guide autoregressive generation, where motion tokens act as priors and are concatenated with textual encodings. Since there is limited work on coordinate-based motion synthesis, we establish new benchmarks for both text-to-motion generation and motion editing. Experimental results demonstrate that our approach achieves state-of-the-art performance in terms of both fidelity and semantic consistency on these benchmarks.

Kang Ding, Hongsong Wang, Jie Gui +1

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

Text-to-Motion Generation Human Motion Generation Constrained Generative Modeling

Ask PDF

Jan 21, 2026

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

Diffusion models have seen widespread adoption for text-driven human motion generation and related tasks due to their impressive generative capabilities and flexibility. However, current motion diffusion models face two major limitations: a representational gap caused by pre-trained text encoders that lack motion-specific information, and error propagation during the iterative denoising process. This paper introduces Reconstruction-Anchored Diffusion Model (RAM) to address these challenges. First, RAM leverages a motion latent space as intermediate supervision for text-to-motion generation. To this end, RAM co-trains a motion reconstruction branch with two key objective functions: self-regularization to enhance the discrimination of the motion space and motion-centric latent alignment to enable accurate mapping from text to the motion latent space. Second, we propose Reconstructive Error Guidance (REG), a testing-stage guidance mechanism that exploits the motion diffusion model's inherent self-correction ability to mitigate error propagation. At each denoising step, REG uses the motion reconstruction branch to reconstruct the previous estimate, reproducing the prior error patterns. By amplifying the residual between the current prediction and the reconstructed estimate, REG highlights the improvements in the current prediction. Extensive experiments demonstrate that RAM achieves significant improvements and state-of-the-art performance. Our code will be released.

Yifei Liu, Changxing Ding, Ling Guo +2

South China University of Technology · Joy Future Academy

Diffusion Model Guidance Text-to-Motion Generation Human Motion Generation Diffusion Models for Motion Generation

Ask PDF