Rethinking Fixed Temporal Grids: Frequency-Disentangled Motion Generation

Yunjiao Zhou; Junlang Qian; Gen Li; Xinying Guo; Lihua Xie; Jianfei Yang

cs.CVOct 2, 2026

Rethinking Fixed Temporal Grids: Frequency-Disentangled Motion Generation

Authors: Yunjiao Zhou, Junlang Qian, Gen Li, Xinying Guo, Lihua Xie, Jianfei Yang

Organizations: Nanyang Technological University, Singapore

Abstract

Most human motion generation methods encode motion as tokens on a uniform temporal grid, where every token spans the same fixed time window. Human motion, however, is temporally heterogeneous: slowly evolving global trajectories coexist with rapid transient events such as foot contacts and joint impulses. Forcing such multi-scale dynamics onto tokens of identical temporal resolution entangles motion frequencies, leaving slow regions redundant while smoothing out the rapid details that distinguish realistic motion. We propose FreqMo, a scale-adaptive motion representation that decomposes motion into wavelet frequency bands, separating dynamics across temporal scales while preserving temporal localization and exact reconstruction. Unified Frequency Residual Quantization (UFRQ) then encodes all bands within a single shared codebook, compressing the token sequence threefold and enabling stable single-stage generation. Experiments show FreqMo attains SOTA fidelity with substantially improved high-frequency preservation, and the same decomposition transfers to continuous diffusion backbones.

Figures & tables

Figure 1: Left: Uniform grids induce temporal over-smoothing or redundancy. Right: FreqMo resolves it by adapting window sizes to intrinsic frequency scales.

Figure 2: Model structure for FreqMo during training.

Figure 3: Model structure for FreqMo during inference.

On the HumanML3D dataset
Methods	Type	FID $\downarrow$	R-Precision $\uparrow$			MM. Dist. $\downarrow$	MM. $\uparrow$
Methods	Type	FID $\downarrow$	Top1	Top2	Top3	MM. Dist. $\downarrow$	MM. $\uparrow$
MDM ( Tevet et al. 2022 )	Diff.	0.544 ±.044	-	-	0.611 ±.007	5.566 ±.027	2.799 ±.072
MLD ( Chen et al. 2023 )	Diff.	0.473 ±.013	0.481 ±.003	0.673 ±.003	0.772 ±.002	3.196 ±.010	2.413 ±.079
MotionDiffuse ( Zhang et al. 2024 )	Diff.	0.630 ±.001	0.491 ±.001	0.681 ±.001	0.782 ±.001	3.113 ±.001	1.553 ±.042
ReMoDiffuse ( Zhang et al. 2023b )	Diff.	0.103 ±.004	0.510 ±.005	0.698 ±.006	0.795 ±.004	2.974 ±.016	1.795 ±.043

Table 1: Performance comparison on HumanML3D and KIT-ML datasets. Bold: best result, underline: second best.

Figure 4: Visual comparisons between the different methods given the distinct text descriptions from HumanML3D testset.

	Reconstruction			High-Frequency Fidelity
Method	MPJPE $\downarrow$	#Token $\downarrow$	#Codebook $\downarrow$	Accel $\downarrow$	Jerk $\downarrow$	HF-RMSE $\downarrow$
VQ-VAE	58.7	49	1	3.42	854	0.045
FSQ	51.5	49	–	3.88	911	0.038
RQ-VAE 2	46.8	98	2	2.85	629	0.032
RQ-VAE 6	28.7	294	6	2.41	586	0.024
FreqMo	29.5	106	1	2.15	532	0.012

Table 2: Quantization representation quality comparison.

Figure 5: Side-by-side comparison on test-set prompts.

Figure 6: Power spectral density comparison.

Figure 7: Comparison of time-frequency distribution. (a) RQ-VAE suffers from frequency entanglement, encoding aliased and noisy signals. (b) Our approach successfully disentangles motion into structured, hierarchical frequency bands.

Level	Freq. Range	#Tokens	Cumulation FID	Ablation FID
L0 (Base)	0.0–0.1 Hz	4	6.44 –	1.30 1.26 $\uparrow$
L1	0.0–0.2 Hz	5	3.54 2.90 $\downarrow$	0.16 0.12 $\uparrow$
L2	0.0–0.4 Hz	8	1.49 2.05 $\downarrow$	0.15 0.11 $\uparrow$
L3	0.1–0.8 Hz	14	0.47 1.02 $\downarrow$	0.16 0.12 $\uparrow$
L4	0.1–1.5 Hz	26	0.11 0.36 $\downarrow$	0.19 0.15 $\uparrow$
L5 (Fine)	0.1–2.5 Hz	49	0.04 0.07 $\downarrow$	0.12 0.08 $\uparrow$

Table 3: Hierarchical frequency contribution analysis.

Group. 1: Tokenizer Design
Configuration	Multi-scale	Wavelet	Res.	FID $\downarrow$
(a) Single-scale	✗	✗	✗	0.141
(b) Err. expansion	✗	✗	✓	0.082
(c) Time domain	✓	✗	✓	0.064
(d) Freq. domain	✓	✓	✗	0.092
Full (FreqMo)	✓	✓	✓	0.040

Table 4: Ablation study. Tokenizer components (Group. 1), generator architecture (Group. 2), decomposition depth (Group. 3), and wavelet basis (Group. 4).

Human Motion Generation

Human motion generation creates articulated movement from language, audio, video, spatial constraints, or interaction goals.

See all 291 papers

MotionVLA: Vision-Language-Action Model for Humanoid Motion
Jun 13, 2026Nonghai Zhang, Siyu Zhai, Yanjun Li +5Text-to-Motion Generation Human Motion Generation
School of Computer Science, Peking University · AI2 Robotics
ScaleMoGen: Autoregressive Next-Scale Prediction for Human Motion Generation
May 12, 2026Inwoo Hwang, Hojun Jang, Bing Zhou +3Text-to-Motion Generation Human Motion Generation
Seoul National University · Snap Inc. · Meta Reality Labs
DC-Motion: Decoupling Structure and Details via Discrete-Continuous Tokens for Human Motion Generation
May 28, 2026Hequan Wang, Xuean Chen, Jiaxu Zhang +2Text-to-Motion Generation Human Motion Generation
State Key Laboratory of Information Engineering in Surveying, Mapping and Remote Sensing, Wuhan University, Wuhan 430072, China · Geophysical Exploration Brigade of Hubei Provincial Geological Bureau, Wuhan 430058, China · Information Systems Technology and Design Pillar, Singapore University of Technology and Design, 487372, Singapore

Jun 13, 2026J/K move · Enter open · S save

MotionVLA: Vision-Language-Action Model for Humanoid Motion

Nonghai Zhang, Siyu Zhai, Yanjun Li, Zeyu Zhang+4

School of Computer Science, Peking University · AI2 Robotics

Text-to-Motion Generation Human Motion Generation Vision-Language-Action Models Motion Representation Learning

Open paper AskPDF arXiv

Generating realistic humanoid motion from scene images and text involves both low-frequency pose semantics and high-frequency physical dynamics. However, many existing methods tokenize motion with a single shared codebook, forcing heterogeneous motion signals into the same quantization space. Our frequency-domain analysis of human motion data reveals a clear mismatch between single-codebook quantization and motion statistics: five DCT coefficients capture 93% of joint-position energy but only 37% of joint-velocity energy, which can bias quantization toward pose statistics and under-represent high-frequency velocity components. A second challenge lies in adapting a standard autoregressive model to effectively model high-frequency physical signals in motion sequences. Therefore, we propose DSFT, a dual-stream frequency tokenizer that separates motion into Base and physical streams and compresses them independently with DCT truncation and BPE. Furthermore, we present MotionVLA, a Qwen3.5-based model that arranges Base and physical tokens in a unified sequence, where Phys tokens are predicted after Base tokens. Experiments on HumanML3D and MBench show that, despite using a lightweight 2B backbone, MotionVLA reduces the Diversity gap to real data by over 50% on HumanML3D and improves Motion-Condition Consistency by 3.8% on MBench, supporting frequency-aware dual-stream decoupling as an effective formulation for autoregressive motion generation. Code: https://github.com/AIGeeksGroup/MotionVLA. Website: https://aigeeksgroup.github.io/MotionVLA.

Abstract

Human Motion Generation

Explore similar work

MotionVLA: Vision-Language-Action Model for Humanoid Motion

ScaleMoGen: Autoregressive Next-Scale Prediction for Human Motion Generation

DC-Motion: Decoupling Structure and Details via Discrete-Continuous Tokens for Human Motion Generation