Paper deep dive
MoRAE: Flow-Friendly Self-Supervised Latents for Text-to-Motion Generation
Yifei Zhu, Mingyi Shi, Yangyang Cai, Miao Cheng, Yoshifumi Kitamura, Taku Komura
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%
Last extracted: 8/3/2026, 2:44:06 AM
Summary
The paper introduces MoRAE (Motion-optimized Representation Autoencoder), a framework for text-to-motion generation that addresses geometric bottlenecks in using self-supervised encoders like Motion-JEPA. It identifies two main issues: spectral ill-conditioning of the feature space and flow residuals aligning with decoder-sensitive directions. MoRAE resolves these by compressing features into a compact 32-dimensional latent space via a variational encoder and employing motion-coupled training to align latent geometry with the decoder, achieving state-of-the-art performance on HumanML3D and KIT-ML datasets.
Entities (8)
Relation Signals (7)
MoRAE → achievesstateofthearton → HumanML3D
confidence 96% · With this flow-friendly latent, a standard non-autoregressive Flow-Matching DiT achieves state-of-the-art performance on HumanML3D
MoRAE → uses → Motion-JEPA
confidence 95% · We directly transfer the RAE recipe to human motion using a frozen Motion-JEPA... as our frozen feature extractor.
MoRAE → achievesstateofthearton → KIT-ML
confidence 94% · and KIT-ML (Plappert et al. 2016), outperforming prior diffusion and VQ-based methods
MoRAE → solves → spectral ill-conditioning
confidence 93% · A compact bottleneck distills the structured JEPA representation... bringing the latent spectrum into a transport-stable regime.
MoRAE → solves → decoder-sensitive directions
confidence 91% · Motion-coupled training then aligns the retained latent geometry with the decoder, making characteristic flow errors less costly after decoding.
Motion-JEPA → exhibits → spectral ill-conditioning
confidence 90% · the JEPA feature space is spectrally ill-conditioned, making the Gaussian-to-data transport unstable
Flow-Matching DiT → operatesin → MoRAE latent space
confidence 88% · With this flow-friendly latent, a standard non-autoregressive Flow-Matching DiT achieves state-of-the-art performance.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Text-to-motion generation must produce motions that are semantically correct, temporally coherent, and physically plausible. A natural approach is to first project motion data into a structured semantic space and then train a generative model within that space. Such a paradigm has been highly successful in image generation through Representation Autoencoders (RAEs), where a frozen self-supervised encoder provides semantic features for diffusion or flow models to learn from. However, direct transfer of such a paradigm to motion space using Motion-JEPA as the frozen encoder fails dramatically. We diagnose this failure geometrically and identify two motion-specific bottlenecks: (1) the JEPA feature space is spectrally ill-conditioned, making the Gaussian-to-data transport unstable; and (2) even with a well-conditioned spectrum, flow residuals tend to align with decoder-sensitive directions, where small latent errors are amplified into large motion artifacts after decoding. Based on these insights, we propose MoRAE. MoRAE addresses the two bottlenecks separately. A compact bottleneck distills the structured JEPA representation while removing weak and redundant directions, bringing the latent spectrum into a transport-stable regime. Motion-coupled training then aligns the retained latent geometry with the decoder, making characteristic flow errors less costly after decoding. With this flow-friendly latent, a standard non-autoregressive Flow-Matching DiT achieves state-of-the-art performance.
Tags
Links
- Source: https://arxiv.org/abs/2607.29180v1
- Canonical: https://arxiv.org/abs/2607.29180v1
Trouble viewing inline? Open PDF directly →
Full Text
52,565 characters extracted from source content.
Expand or collapse full text
MoRAE: Flow-Friendly Self-Supervised Latents for Text-to-Motion Generation Yifei Zhu,1 Mingyi Shi,2 Yangyang Cai,1 Miao Cheng,1 Yoshifumi Kitamura,1 Taku Komura2 Abstract Text-to-motion generation must produce motions that are semantically correct, temporally coherent, and physically plausible. A natural approach is to first project motion data into a structured semantic space and then train a generative model within that space. Such a paradigm has been highly successful in image generation through Representation Autoencoders (RAEs) , where a frozen self-supervised encoder provides semantic features for diffusion or flow models to learn from. However, direct transfer of such a paradigm to motion space using Motion-JEPA as the frozen encoder fails dramatically. We diagnose this failure geometrically and identify two motion-specific bottlenecks: (1) the JEPA feature space is spectrally ill-conditioned, making the Gaussian-to-data transport unstable; and (2) even with a well-conditioned spectrum, flow residuals tend to align with decoder-sensitive directions, where small latent errors are amplified into large motion artifacts after decoding. Based on these insights, we propose MoRAE. MoRAE addresses the two bottlenecks separately. A compact bottleneck distills the structured JEPA representation while removing weak and redundant directions, bringing the latent spectrum into a transport-stable regime. Motion-coupled training then aligns the retained latent geometry with the decoder, making characteristic flow errors less costly after decoding. With this flow-friendly latent, a standard non-autoregressive Flow-Matching DiT achieves state-of-the-art performance. 1 Introduction Valid human motions constitute an exceptionally thin manifold within the ambient feature space, which fundamentally complicates generative transport. For example, text-to-motion models must satisfy not only semantic alignment with the language prompt but also strict physical constraints: bone lengths must remain fixed, joint rotations must remain valid, and foot contacts must be consistent with body trajectories. These constraints are largely implicit in the text, yet violating any of them immediately breaks the generated motion. From a geometric perspective, the valid region occupies only a negligible fraction of the high-dimensional space (e.g., 263 dimensions in HumanML3D). Consequently, diffusion or flow models must transport full-dimensional Gaussian noise toward this thin region before they can model text-controlled semantic variation. Near the manifold boundary, invalid directions demand sharp corrective velocities, making the transport field difficult to learn (De Bortoli 2022; Pidstrigach 2022) and consuming much of the generator’s capacity on enforcing physical plausibility rather than following the user’s semantic command. A natural remedy to this geometric difficulty is to project data into a semantically structured feature space, where the valid distribution becomes more isotropic and easier for generative models to learn. Instead of training directly in the raw physical space, one can map motion data into a higher-level representation that captures temporal dynamics, inter-joint dependencies, and action semantics. In image generation, this paradigm has been established through Representation Autoencoders (RAEs) (Zheng et al. 2025; Singh et al. 2026): a strong self-supervised visual encoder (e.g., DINOv2 (Oquab et al. 2024)) is pretrained and then frozen to extract high-dimensional semantic features, a decoder reconstructs the original pixels from these features, and a diffusion or flow model operates directly in this frozen feature space. This recipe has proven remarkably successful in images, yielding stable training and high-quality generations without requiring the generative model to handle low-level pixel correlations. Encouraged by the image-domain success, we directly transfer the RAE recipe to human motion using a frozen Motion-JEPA (Assran et al. 2023), a self-supervised joint-embedding predictive encoder pretrained on motion data, as our frozen feature extractor. We then learn a decoder to reconstruct the original motion from these features and train a standard Flow-Matching DiT (Lipman et al. 2023; Ma et al. 2024) in this feature space. Although reconstruction accuracy is competitive with existing methods, generation quality degrades substantially—the flow model produces physically implausible and semantically mismatched motions, performing far worse than when trained in a compact autoencoder latent space. To explain why the motion RAE fails, we perform a diagnostic analysis that uncovers two motion-specific geometric bottlenecks. First, the JEPA feature space remains spectrally ill-conditioned: despite per-coordinate normalization, its covariance spectrum has a large condition number (κflow≈1.3×104 _flow≈ 1.3× 10^4), destabilizing the Gaussian-to-data transport. Second, spectral conditioning alone is insufficient—even when we artificially correct the spectrum, flow prediction residuals tend to align with decoder-sensitive directions, where the motion decoder has high Jacobian gain. A small latent error in these directions is amplified into large motion artifacts after decoding, severely hurting generation quality even when reconstruction error is minimal. Guided by these findings, we propose MoRAE (Motion-optimized Representation Autoencoder), a framework that reshapes the latent space to be explicitly flow-friendly. Instead of generating directly in the raw JEPA feature space, we first compress its frozen features into a compact 32-dimensional latent space using a variational encoder. Crucially, we introduce coupled learning: the motion reconstruction loss is backpropagated through the entire tokenizer—including both the feature decoder and the encoder—so that the latent space is explicitly shaped to be both transport-stable and robust to decoder amplification. With this flow-friendly latent space, a standard non-autoregressive Flow-Matching DiT achieves state-of-the-art performance on HumanML3D (Guo et al. 2022) and KIT-ML (Plappert et al. 2016), outperforming prior diffusion and VQ-based methods in both semantic metrics (FID, R-Precision) and physical plausibility (foot skating, jerk, bone variation), all without relying on vector quantization or autoregressive factorization. 2 Related Work Text-to-Motion Generation. Text-to-Motion Generation is a topic that is widely attracting researchers for synthesizing single human motion (Tevet et al. 2023), human-object interaction (Xu et al. 2023), human-scene interaction (Wang et al. 2022) and multi-person interaction (Liang et al. 2024). In addition to simply using text prompts as inputs, methods to use trajectories (Xie et al. 2024) or style motion (Guo et al. 2024b) are useful for controlling the details of the motion. In terms of the representation of motion, it has been approached through three main paradigms. The first family generates motions directly in the raw feature space using diffusion or flow models (e.g., MDM (Tevet et al. 2023), MotionDiffuse (Zhang et al. 2022), FLAME (Kim et al. 2023)). While straightforward, these methods must learn both high-level semantics and low-level physical constraints simultaneously, often leading to unstable training or physical artifacts. The second family quantizes motion into discrete tokens and generates sequences autoregressively or via masked prediction (e.g., T2M-GPT (Zhang et al. 2023), MoMask (Guo et al. 2024a), M (Pinyoanuntapong et al. 2024)). Although effective at capturing multimodal distributions, these methods suffer from quantization errors, restricted continuous variation, and typically require sequential prediction. The third family operates in a continuous latent space learned by autoencoders or VAEs (e.g., MLD (Chen et al. 2023), MARDM (Meng et al. 2024)), offering a compromise between raw-space flexibility and discrete-token efficiency. In this work, we adopt the continuous-latent paradigm based on a semantically structured space and then learn a compact, flow-friendly latent representation through a coupled autoencoder. From Reconstruction to Structured Representations. Conventional continuous motion latents—whether deterministic autoencoders or VAEs—are optimized primarily for reconstruction fidelity, with VAEs additionally matching a Gaussian prior. While this yields compact representations, reconstruction alone does not guarantee that the latent space is well-suited for generative modeling; as we will show in Sec. 3, such spaces often exhibit poor spectral conditioning and decoder sensitivity. In contrast, self-supervised learning (SSL) offers an alternative: instead of reconstructing raw inputs, SSL methods learn structured features through invariance, masked prediction, or joint-embedding objectives (Caron et al. 2021; Oquab et al. 2024; He et al. 2022; Assran et al. 2023). These representations capture high-level content and structural relations—such as temporal dynamics and inter-joint dependencies in motion—but are typically high-dimensional and not directly decodable to the original domain. Recent image-generation methods bridge this gap through representation autoencoders (RAEs) (Zheng et al. 2025; Singh et al. 2026): they retain a pretrained SSL encoder, freeze it, and learn a decoder that maps its features back to the input space, enabling generative models to operate on semantically structured features. Our work adopts this philosophy for motion, using Motion-JEPA (Assran et al. 2023) as the frozen encoder. However, motion introduces additional difficulty: its features are temporally and kinematically coupled, and raw SSL representations remain wide and redundant. Our MoRAE therefore does not merely freeze and decode JEPA features; it compresses them into a compact continuous latent and couples its learning to the motion decoder, ensuring that the resulting space is both transport-stable and decoder-compatible. 3 Diagnosis To understand why accurate motion tokenizers can still result in poor motion synthesis, we dissect three geometric factors: spectral conditioning, residual magnitude, and decoder alignment. We first quantify the spectral anisotropy of motion representations (§3.1), then introduce a framework to decompose decoded error into interpretable components (§3.2). Using this framework, Probe 1 isolates the causal effect of spectral conditioning (§3.3), while Probe 2 reveals that residual–decoder alignment is an independent bottleneck (§3.4). We conclude by synthesizing these findings into design principles that motivate MoRAE (§3.5). 3.1 Spectral Bias in Motion Representations Motion representations are high-dimensional but strongly constrained by kinematics, temporal consistency, and contact. Consequently, valid motions occupy only a thin subset of the ambient space. For example, local PCA on the raw 263263-D HumanML3D representation and its 6767-D relative-position subset gives local-d95=26d_95=26 and 1717, respectively, where local-d95d_95 denotes the number of local principal components explaining 95%95\% of neighborhood variance. A low intrinsic dimension, however, does not necessarily make full-dimensional generative transport easier. When data variance is concentrated in a small dominant subspace, the remaining directions have little variance but must still be contracted from an isotropic Gaussian prior. The resulting transport becomes difficult when the covariance spectrum contains many weak directions. We measure this anisotropy after applying the same per-coordinate normalization used at the generator input: ~ z =diag()−1(−),=Cov(~), =diag( s)^-1( z- μ), C=Cov( z), (1) κflow _flow =cond(+σp2)=λ1+σp2λd+σp2,σp=0.02, =cond ( C+ _p^2 I )= _1+ _p^2 _d+ _p^2, _p=02, where λ1≥⋯≥λd _1≥·s≥ _d are the eigenvalues of C. Per-coordinate normalization removes marginal scale differences but not cross-coordinate correlations. The perturbation floor prevents unresolved near-zero directions from dominating the estimate. Thus, κflow _flow measures the joint spectral anisotropy actually seen by the generator. Table 1 yields three observations. First, SSL improves representation content without guaranteeing transport-ready geometry. Motion-JEPA reduces κflow _flow relative to raw motion (1.3⋅1041.3·10^4 versus 6.8⋅1046.8·10^4), but its 768768-D representation has local-d95=43d_95=43. Predictable motion factors therefore occupy a small dominant subspace, while many remaining directions are weak or correlated. Second, a compact bottleneck can improve conditioning by removing such weak directions. If an information-preserving compression retains k dominant directions, the relevant spectral ratio changes approximately from λ1/λd _1/ _d to λ1/λk _1/ _k. The gain therefore comes from discarding near-null directions, rather than from reducing dimensionality itself. Indeed, reducing the raw representation from 263263 to 6767 dimensions barely improves conditioning, and the 3232-D VAE remains severely ill-conditioned. Third, the retained spectrum depends on what the bottleneck is trained to preserve. MoRAE latent and the standard AE have the same width and local dimension (d=32d=32, local-d95=18d_95=18), yet their condition numbers differ by a factor of five (2727 versus 136136). This suggests that compressing contextualized JEPA features organizes the retained information more evenly than directly reconstructing motion. Thus, semantic pretraining supplies structured content, while compact distillation removes weak directions; neither component alone guarantees a flow-friendly latent. latent d local-d95d_95 d95/d_95/d κflow↓ _flow raw (HumanML3D feature) 263 26 9.9% 6.8⋅1046.8·10^4 essential (relative position) 67 17 25.4% 3.7⋅1043.7·10^4 AE 512 40 7.8% 7.2⋅1047.2·10^4 AE 32 18 56.3% 136136 VAE 256 53 20.7% 1.1⋅1031.1·10^3 VAE 32 14 43.8% 1.2⋅1041.2·10^4 SSL (raw H) 768 43 5.6% 1.3⋅1041.3·10^4 SSL (MoRAE latent) 32 18 56.3% 27 Table 1: Geometry of continuous HumanML3D representations. The AE and VAE architectures follow MARDM (Meng et al. 2024) and MLD (Chen et al. 2023), respectively; the latent width is varied as indicated. 3.2 A Generative-Error View of Latent Quality Spectral conditioning alone does not guarantee accurate decoding, as the decoder may amplify residuals in direction-dependent ways. To separate the effects of residual magnitude, decoder sensitivity, and their directional alignment, we introduce a decomposition of the decoded error. Our generative process consists of two stages: a flow model transports Gaussian noise (,)N(0,I), to a latent variable z, and the decoder D then maps z to motion x. The encoder E is used only during training to provide the target latents z for the flow. Let (,^)( z, z) be a pair of a ground-truth latent encoded from a motion x and its inference produced from the Gaussian noise. The residual can be defined as =^−. δ= z- z. (2) We now analyze the effect the residual δ to the final motion during inference. The effect of δ depends on both its magnitude and direction. The decoder Jacobian induces the local pullback metric ()=D()⊤D(),D()=∂D∂. G( z)= J_D( z)^\! J_D( z), J_D( z)= ∂ D∂ z. (3) For a sufficiently small residual, ‖D(+)−D()‖22=⊤()+O(∥23). D( z+ δ)-D( z) _2^2= δ^\! G( z) δ+O( δ _2^3). (4) Thus, residuals with the same Euclidean norm can produce very different motion errors when they point along decoder directions with different local gains. To separate the generator and decoder contributions, define the conditional residual covariance e()=[⊤∣], C_e( z)=E[ δ δ^\! z], (5) which captures where the flow places residual energy. The expected decoded error is then approximately ‖D(+)−D()‖22≈[Tr(()e())].E D( z+ δ)-D( z) _2^2 _ z [Tr ( G( z) C_e( z) ) ]. (6) This is our central diagnostic: the generator determines e() C_e( z), while the decoder determines () G( z). To make this interaction interpretable, we factor the decoded error into three scalar quantities for tokenizers in a common standardized coordinate system: Mdec⏟decoded residual cost=Mlat⏟residual magnitude⋅giso⏟mean decoder gain⋅ρalign⏟directional alignment, 424.94574pt$ M_dec_decoded residual cost= M_lat_residual magnitude· g_iso_mean decoder gain· _align_directional alignment$, (7) where Mlat M_lat =1d∥22, = 1dE δ _2^2, giso g_iso =1d[Tr()], = 1dE[Tr G( z)], (8) ρalign _align =[⊤()]∥22[Tr()]/d, = E[ δ^\! G( z) δ]E δ _2^2\,E[Tr G( z)]/d, Mdec M_dec =1d[⊤()]. = 1dE[ δ^\! G( z) δ]. Here, MlatM_lat measures the generator error, gisog_iso measures the decoder’s average sensitivity, and ρalign _align measures whether residual energy is concentrated in decoder-sensitive directions. In particular, ρalign>1 _align>1 indicates a larger decoded cost than an isotropically oriented residual with the same average energy. This decomposition allows Probe 2 to attribute the generation gap to residual magnitude, decoder sensitivity, or directional alignment. The individual factors are coordinate-dependent, so we compare them only after placing all tokenizers in the same standardized coordinates. Reconstruction FID is used as an information-preservation control, while generation FID and R@1 evaluate generative quality. Probe 1 isolates the effect of spectral conditioning with information and the decoder fixed; Probe 2 uses Eq. (7) to attribute the remaining generation gap to residual magnitude, decoder sensitivity, or directional alignment. 3.3 Probe 1: Spectral Conditioning In this probe, we show that a poor spectral condition number causally degrades flow-based generation. To isolate this causal effect, we artificially manipulate the condition number of a fixed latent space while keeping its information content and decoder unchanged, then retrain the flow in each manipulated space. To this end, we apply an invertible reparameterization to a fixed centered and standardized latent z. Let W be its ZCA-whitening transform and R a fixed random orthogonal matrix. For a target condition number κ, κ′ z _κ =κ,κ=κ, = A_κ z, A_κ= R S_κ W, (9) [κ]ii [ S_κ]_i =κ−i−12(d−1),i=1,…,d. =κ^- i-12(d-1), i=1,…,d. The resulting covariance eigenvalues are geometrically spaced from 11 to 1/κ1/κ. For each κ, we retrain the same flow in the transformed coordinates, map its endpoints back through κ−1 A_κ^-1, and use the unchanged decoder. Only κ S_κ varies. Round-trip error is below 2⋅10−62·10^-6 and rFID remains 0.0680.068. The rotated-natural control preserves the natural spectrum while changing its basis. Target κ is imposed before flow-input normalization, while κflow _flow is measured afterward using Eq. (1). Following Sec. 3.2, we evaluate the latent residual in the flow coordinates (MtransM_trans) and after mapping it back to the native coordinates (MorigM_orig). We also report their ratio and the native-coordinate SWD. variant κflow _flow rFID MtransM_trans MorigM_orig amp. SWD↓ gen FID↓ R@1↑ whitened 11 0.0680.068 0.0470.047 0.0350.035 0.7×0.7× 0.0740.074 0.1700.170 0.5130.513 natural 2727 0.0680.068 0.0420.042 0.0420.042 1.0×1.0× 0.0820.082 0.2270.227 0.5040.504 rotated natural 2323 0.0680.068 0.0450.045 0.0410.041 0.9×0.9× 0.0830.083 0.2180.218 0.5130.513 κ=27κ\!=\!27 (geom.) 2525 0.0680.068 0.0280.028 0.0860.086 3.1×3.1× 0.0780.078 0.1600.160 0.508¯ 0.508 κ=102κ\!=\!10^2 8989 0.0680.068 0.0220.022 0.1210.121 5.4×5.4× 0.1250.125 0.1740.174 0.5040.504 κ=103κ\!=\!10^3 793793 0.0680.068 0.0200.020 0.3020.302 15.5×15.5× 0.1130.113 0.163¯ 0.163 0.5010.501 κ=103.5κ\!=\!10^3.5 2.2⋅1032.2·10^3 0.0680.068 0.0180.018 0.6010.601 32.8×32.8× 0.1560.156 0.5340.534 0.4960.496 κ=104κ\!=\!10^4 5.4⋅1035.4·10^3 0.0680.068 0.0170.017 1.281.28 73×73× 0.2840.284 8.058.05 0.3560.356 κ=106κ\!=\!10^6 2.3⋅1042.3·10^4 0.0680.068 0.0130.013 82.482.4 6112×6112× 4.964.96 87.487.4 0.0660.066 Table 2: Probe 1: invertible spectral reparameterization with represented information and decoder fixed. Amp. is Morig/MtransM_orig/M_trans. All variants use the same DiT-S protocol for 500500 epochs. Table 2 shows a broad stable regime followed by a sharp collapse. Generation FID remains between 0.1600.160 and 0.2270.227 for κflow≤793 _flow≤ 793, degrades at 2.2⋅1032.2·10^3, and becomes catastrophic under stronger squeezing despite unchanged rFID. Whitening is therefore not required; the spectrum only needs to remain within a transport-stable regime. The natural and rotated-natural controls perform similarly, indicating a small basis effect. However, spectra with similar κflow _flow still differ slightly, so condition number does not fully characterize spectral shape. As squeezing increases, MtransM_trans decreases from 0.0470.047 to 0.0130.013, while MorigM_orig rises from 0.0350.035 to 82.482.4 and SWD rises to 4.964.96. Thus, small residuals in the flow coordinates can conceal large errors along squeezed directions after mapping back to the native latent space. 3.4 Probe 2: Directional Alignment Probe 1 established that poor spectral conditioning can cause catastrophic generation, but not whether good conditioning alone suffices. To localize the remaining gap, we compare two similarly conditioned tokenizers. In coupled MoRAE (Sec. 4.2), the motion loss backpropagates through the feature decoder and encoder, shaping the latent by its motion-space consequences. The decoupled control uses the same architecture and objectives but stops this gradient before the feature decoder and encoder, leaving the latent optimized only for feature reconstruction. Both lie in the stable regime (κflow=27 _flow=27 and 29.529.5), yet the decoupled variant generates worse motion despite better reconstruction and a smaller residual. Spectral conditioning, reconstruction quality, and residual magnitude therefore cannot explain the remaining gap. We apply the decoder-aware decomposition of Eq. (7) to held-out endpoint pairs under the fixed evaluation coupling. All terms are evaluated at guidance scale 11 in the standardized coordinates of Sec. 3.2. At the observed residual scale, the local approximation is accurate: δ⊤Gδ Gδ correlates with realized decoded error at Pearson r≥0.92r≥ 0.92, with an actual-to-predicted ratio of 0.80.8–1.11.1. tokenizer decoder training κflow _flow rFID MlatM_lat gisog_iso ρalign _align MdecM_dec gen FID↓ R@1↑ coupled (MoRAE) joint 2727 0.0680.068 0.0590.059 0.550.55 4.854.85 0.1570.157 0.1600.160 0.5080.508 decoupled detached 29.529.5 0.0570.057 0.0550.055 0.590.59 6.086.08 0.1950.195 0.2430.243 0.5030.503 Table 3: Probe 2: decoder-aware decomposition of held-out transport residuals. Both tokenizers lie in the transport-stable spectral regime and use the same DiT-S protocol for 500 epochs. Table 3 localizes the gap. The decoupled flow has a slightly smaller residual (Mlat=0.055M_lat=0.055 vs. 0.0590.059) and similar mean decoder gain (giso=0.59g_iso=0.59 vs. 0.550.55). The main difference is directional: its residual alignment is about 25%25\% higher (ρalign=6.08 _align=6.08 vs. 4.854.85), indicating that more residual energy falls along decoder-sensitive directions. This raises the decoded cost (Mdec=0.195M_dec=0.195 vs. 0.1570.157) and worsens generation FID (0.2430.243 vs. 0.1600.160), while R@1 remains nearly unchanged. The gap therefore lies primarily in motion-distribution quality rather than semantic retrieval. Coupled training improves generation not by reducing residual magnitude or mean decoder sensitivity, but by making the flow’s characteristic errors less aligned with directions strongly amplified by the decoder. 3.5 Implications for Continuous Motion Latents The two probes provide a geometric interpretation of common representation choices in motion generation. This view also refines the representation diagnosis of MARDM (Meng et al. 2024). MARDM removes the derived channels of the canonical 263263-D HumanML3D format and retains the 6767-D animation-relevant state; Table 1 shows that even this essential representation remains spectrally thin and severely ill-conditioned. Thus, removing explicit channel redundancy helps, but does not eliminate the weak continuous directions that make global Gaussian transport difficult. A wide deterministic autoencoder can therefore reconstruct accurately while leaving many weak latent directions for the generator to model. Autoregressive or masked-autoregressive generation provides a practical workaround by decomposing one global transport problem into a sequence of restricted conditional predictions. Reducing latent width instead can improve spectral conditioning, but does not by itself make the remaining transport errors compatible with the decoder. The same view explains why Gaussian prior matching is insufficient: a VAE may regularize coordinate marginals while leaving the joint spectrum poorly conditioned. VQ tokenizers avoid continuous transport through weak directions by restricting generation to discrete codes. This helps explain their empirical robustness, and complements MARDM’s observation that redundant motion channels can regularize codebook learning, but comes at the cost of quantization, restricted continuous variation, and typically sequential prediction (van den Oord et al. 2017; Meng et al. 2024). MoRAE instead uses a compact continuous latent for transport stability, while coupled decoding makes the generator’s characteristic errors less costly in motion space. 4 Method Figure 1: Overview of MoRAE. (a) Motion-JEPA predicts EMA-teacher features under temporal-span and joint-group masks. (b) Frozen features are compressed into a 3232-D latent and jointly decoded back to features and motion. (c) After freezing the tokenizer, a non-autoregressive DiT learns flow matching in the compact latent space, and the frozen decoder maps generated latents back to motion. As shown in Fig. 1, MoRAE is trained in three stages. We first pretrain a self-supervised Motion-JEPA encoder f. With f frozen, we then train a compact tokenizer consisting of a variational encoder E, a feature decoder DfD_f, and a motion decoder DmD_m. Finally, all tokenizer modules are frozen and a standard flow-matching DiT is trained on the resulting latent sequence. The complete reconstruction path is →frozen→Df^→Dm^. x [frozen]f H E z D_f H D_m x. (10) 4.1 Motion-JEPA Pretraining We pretrain the source representation through masked latent prediction in the joint-embedding predictive framework (Assran et al. 2023). An online encoder fθf_θ predicts the clean-motion features of an exponential-moving-average teacher fθ¯f_ θ at masked positions: ℒJEPA=∑t∈ℳsmooth-ℓ1(P(fθ(⊙))t,sg[fθ¯()t]),L_JEPA= _t smooth- _1 (P(f_θ( x m))_t,\,sg[f_ θ( x)_t] ), (11) where P is the predictor, ℳM denotes masked positions, and θ¯←μθ¯+(1−μ)θ θ←μ θ+(1-μ)θ. We use temporal-span masking over contiguous frames and joint-group masking over selected joints across time, encouraging the encoder to infer temporal dynamics and inter-joint dependencies. After pretraining, we freeze f and obtain contextual features =f()∈ℝT×768 H=f( x) ^T× 768. Although structurally informative, these features remain wide and spectrally ill-conditioned (Table 1); the following tokenizer compresses them into a compact generative latent. 4.2 Compact Coupled Tokenizer With f frozen, a single-attention variational encoder maps H to qϕ(∣) q_φ( z H) =(,diag(2)), =N ( μ,diag( σ^2) ), (12) z =+⊙ϵ∈ℝT×32, = μ+ σ ε ^T× 32, where ϵ∼(,) ε ( 0, I). A feature decoder DfD_f reconstructs the frozen representation, ^=Df() H=D_f( z), and a motion decoder DmD_m maps it back to motion. During tokenizer training, we add a small Gaussian perturbation to the reconstructed feature before motion decoding: ^=Dm(^+ηϵ′),ϵ′∼(,). x=D_m ( H+η ε ), ε ( 0, I). (13) The tokenizer objective is ℒtok= _tok= ∥^−∥⏟ℒfeat+λ‖Dm(^+ηϵ′)−‖⏟ℒmot H- H _L_feat+λ D_m( H+η ε )- x _L_mot (14) +βKL(qϕ(∣)∥(,)). +β\,KL (q_φ( z H)\,\|\,N( 0, I) ). The compact bottleneck and feature loss preserve the frozen representation in a low-dimensional latent, while the KL regularizes its marginal distribution. Crucially, ℒmotL_mot backpropagates through DmD_m, DfD_f, and E, coupling the latent representation to motion decoding. The decoupled control in Probe 2 uses the same architecture and objective but stops the motion-loss gradient before it reaches DfD_f and E. 4.3 Latent Flow Matching After tokenizer training, f, E, DfD_f, and DmD_m are frozen. We train a DiT-based flow-matching model (Lipman et al. 2023; Ma et al. 2024) over the full sequence of latent tokens. Given 0∼(,),1∼qϕ(∣f()), z_0 ( 0, I), z_1 q_φ( z f( x)), (15) we use the linear interpolant t=t1+(1−t)0 z_t=t z_1+(1-t) z_0 and regress the constant target velocity: ℒFM=t,0,1,c[‖θ(t,t,c)−(1−0)‖22].L_FM=E_t, z_0, z_1,c [ v_θ( z_t,t,c)-( z_1- z_0) _2^2 ]. (16) At inference, we integrate the probability-flow ODE from Gaussian noise to a generated latent ^1 z_1, then decode ^=Dm(Df(^1)) x=D_m(D_f( z_1)). Text condition c is encoded by a frozen CLIP text encoder (Radford et al. 2021) and injected through adaptive layer normalization following DiT/SiT (Peebles and Xie 2023; Ma et al. 2024). Classifier-free guidance is applied during sampling. No vector quantization or autoregressive factorization is used. 5 Experiments 5.1 Datasets and Evaluation Protocol We evaluate on HumanML3D (Guo et al. 2022) and on KIT-ML (Plappert et al. 2016) under the 2020 FPS, following the official split. Evaluation protocol. Following MARDM (Meng et al. 2024), our main experiments use the 6767-D essential representation—the root state and relative joint positions used to recover motion—and the corresponding evaluator. We report FID, R-Precision, Matching Distance, and Multimodality, together with a CLIP-style motion–text alignment score (Radford et al. 2021). We additionally measure foot skating, temporal jerk, and bone-length variation for contact stability, smoothness, and skeletal consistency. Lower foot skating and bone variation are better; jerk is judged by proximity to the ground-truth distribution. Implementation. Motion-JEPA and DiT are trained on the official splits. The reported generator is a DiT-XL trained for 800800 epochs with batch size 6464, learning rate 2×10−42×10^-4, and cosine decay. We integrate the probability-flow ODE for 2424 steps with classifier-free guidance (cfg=8.5cfg\!=\!8.5 on HumanML3D and 7.57.5 on KIT-ML). Method MDM MLD T2M-GPT M MoMask MARDM-SiT MoRAE AIT (s) ↓ 14.3114.31 0.210.21 0.320.32 0.060.06 0.040.04 2.402.40 0.060.06 Table 4: Inference efficiency. 5.2 Main Results Methods R-Precision↑ FID↓ Matching↓ MModality↑ CLIP-score↑ FootSkate Jerk Bone Var.↓ Top-1 Top-2 Top-3 HumanML3D GT 0.503±.0030.503 ±.003 0.697±.0020.697 ±.002 0.796±.0020.796 ±.002 0.0000.000 3.242±.0103.242 ±.010 — 0.640±.0010.640 ±.001 11.6±.011.6 ±.0 44.4±.044.4 ±.0 0.00±.00.00 ±.0 T2M-GPT 0.470±.0030.470 ±.003 0.659±.0020.659 ±.002 0.758±.0020.758 ±.002 0.335±.0030.335 ±.003 3.505±.0173.505 ±.017 2.018±.0532.018 ±.053 0.607±.0050.607 ±.005 17.66±.0717.66 ±.07 74.5±.7374.5 ±.73 2.83±.12.83 ±.1 M 0.487±.0030.487 ±.003 0.683±.0020.683 ±.002 0.782±.0010.782 ±.001 0.132±.0040.132 ±.004 3.359±.0093.359 ±.009 1.241±.0731.241 ±.073 0.635±.0030.635 ±.003 16.20±.0616.20 ±.06 63.5±.4563.5 ±.45 2.12±.12.12 ±.1 MoMask 0.490±.0040.490 ±.004 0.687±.0030.687 ±.003 0.786±.0030.786 ±.003 0.116±.0060.116 ±.006 3.353±.0103.353 ±.010 1.263±.0791.263 ±.079 0.637±.0030.637 ±.003 15.81±.0815.81 ±.08 60.3±.1960.3 ±.19 2.51±.02.51 ±.0 MDM 0.440±.0070.440 ±.007 0.636±.0060.636 ±.006 0.742±.0040.742 ±.004 0.518±.0320.518 ±.032 3.640±.0283.640 ±.028 3.604±.0313.604 ±.031 0.578±.0030.578 ±.003 15.71±.0615.71 ±.06 36.5¯±.62 36.5 ±.62 1.82¯±.2 1.82 ±.2 MotionDiffuse 0.450±.0060.450 ±.006 0.641±.0050.641 ±.005 0.753±.0050.753 ±.005 0.778±.0050.778 ±.005 3.490±.0233.490 ±.023 3.179±.0463.179 ±.046 0.606±.0040.606 ±.004 15.79±.0615.79 ±.06 22.5±.2022.5 ±.20 2.97±.22.97 ±.2 MLD 0.461±.0040.461 ±.004 0.651±.0040.651 ±.004 0.750±.0030.750 ±.003 0.431±.0140.431 ±.014 3.445±.0193.445 ±.019 3.506¯±.031 3.506 ±.031 0.610±.0030.610 ±.003 15.99±.0815.99 ±.08 20.1±.4520.1 ±.45 2.43±.12.43 ±.1 MARDM (DiT-XL) 0.500¯±.004 0.500 ±.004 0.695¯±.003 0.695 ±.003 0.795¯±.003 0.795 ±.003 0.114¯±.007 0.114 ±.007 3.270¯±.009 3.270 ±.009 2.231±.0712.231 ±.071 0.642¯±.002 0.642 ±.002 20.33±.0920.33 ±.09 76.3±.5476.3 ±.54 2.10±.02.10 ±.0 MoRAE (DiT-XL) 0.512±.0030.512 ±.003 0.704±.0020.704 ±.002 0.798±.0020.798 ±.002 0.089±.0040.089 ±.004 3.233±.0063.233 ±.006 1.310±.0591.310 ±.059 0.651±.0010.651 ±.001 15.64±.0615.64 ±.06 40.6±.1540.6 ±.15 1.61±.01.61 ±.0 KIT-ML GT 0.3770.377 0.6160.616 0.7590.759 0.0000.000 3.3013.301 — 0.6990.699 29.5±.029.5 ±.0 62.7±.062.7 ±.0 0.00±.00.00 ±.0 T2M-GPT 0.359±.0070.359 ±.007 0.553±.0070.553 ±.007 0.690±.0130.690 ±.013 0.593±.0530.593 ±.053 3.765±.0463.765 ±.046 1.798¯±.157 1.798 ±.157 0.651±.0050.651 ±.005 33.54±.1033.54 ±.10 56.50±.5856.50 ±.58 3.23±.023.23 ±.02 M 0.363±.0050.363 ±.005 0.569±.0060.569 ±.006 0.724±.0060.724 ±.006 0.478±.0340.478 ±.034 3.629±.0283.629 ±.028 1.455±.1061.455 ±.106 0.660±.0030.660 ±.003 32.74±.0932.74 ±.09 59.10¯±.44 59.10 ±.44 3.34±.033.34 ±.03 MoMask 0.369±.0050.369 ±.005 0.588±.0050.588 ±.005 0.731±.0050.731 ±.005 0.411±.0260.411 ±.026 3.577±.0213.577 ±.021 1.309±.0581.309 ±.058 0.669±.0020.669 ±.002 32.39±.0732.39 ±.07 47.60±.6947.60 ±.69 7.05±.057.05 ±.05 MDM 0.333±.0120.333 ±.012 0.561±.0090.561 ±.009 0.689±.0090.689 ±.009 0.585±.0430.585 ±.043 4.002±.0334.002 ±.033 1.681±.1071.681 ±.107 0.605±.0070.605 ±.007 31.77±.0831.77 ±.08 44.00±.5144.00 ±.51 2.34¯±.02 2.34 ±.02 MotionDiffuse 0.344±.0090.344 ±.009 0.536±.0070.536 ±.007 0.658±.0070.658 ±.007 3.845±.0873.845 ±.087 4.167±.0544.167 ±.054 1.774±.2171.774 ±.217 0.626±.0060.626 ±.006 29.05¯±.06 29.05 ±.06 39.90±.3539.90 ±.35 2.22±.022.22 ±.02 MLD 0.351±.0070.351 ±.007 0.536±.0070.536 ±.007 0.658±.0070.658 ±.007 0.492±.0470.492 ±.047 3.746±.0443.746 ±.044 1.803±.1641.803 ±.164 0.646±.0060.646 ±.006 30.20±.0830.20 ±.08 38.20±.4338.20 ±.43 2.55±.022.55 ±.02 MARDM (DiT-XL) 0.387¯±.006 0.387 ±.006 0.610¯±.006 0.610 ±.006 0.749¯±.006 0.749 ±.006 0.242¯±.014 0.242 ±.014 3.374¯±.019 3.374 ±.019 1.312±.0531.312 ±.053 0.692¯±.002 0.692 ±.002 36.54±.1136.54 ±.11 64.70±.6264.70 ±.62 2.85±.032.85 ±.03 MoRAE (DiT-XL) 0.410±.0060.410 ±.006 0.630±.0070.630 ±.007 0.761±.0050.761 ±.005 0.169±.0090.169 ±.009 3.254±.0203.254 ±.020 1.429±.0461.429 ±.046 0.697±.0020.697 ±.002 28.47±.1528.47 ±.15 63.2±.4563.2 ±.45 1.99±.011.99 ±.01 Table 5: Quantitative evaluation under the essential-dimension evaluators. We report the mean over 2020 repeats with 95%95\% confidence intervals, identical to prior work. Bold is best, underline second best. Jerk is best when closest to GT (44.444.4). MoRAE achieves the best overall performance on both HumanML3D and KIT-ML, improving distributional quality, text–motion alignment, and retrieval accuracy. This supports our central claim: a compact, transport-stable, and motion-faithful representation allows a standard flow to match or surpass discrete or autoregressive generators. Physical metrics further reveal failures hidden by standard scores. VQ and autoregressive methods can accumulate temporal discontinuities, increasing jerk, foot skating, and skeletal variation, while diffusion in raw or ill-conditioned spaces can produce contact and kinematic errors during transport toward the thin valid-motion set. MoRAE mitigates both through a compact, transport-stable self-supervised latent shaped by motion-coupled decoding. Figure 2 qualitatively confirms this advantage: for a compositional prompt, MoRAE preserves the ordered forward-walking, right-arm, and backward-walking phases, whereas competing methods omit or entangle parts of the sequence. Inference efficiency. We tested average inference time per sample (batched), with methods ordered by generation FID. Baselines measured on RTX 4090. Since our method only need a single diffusion process, it achieves the best balance between speed and quality. Figure 2: Qualitative comparison on a compositional text prompt. 5.3 Ablations We vary one factor at a time and report generation FID (gFID); Table 6 uses the diagnostic protocol, Table 7 the final one. Autoregression compensates poor geometry. Masked-AR rescues AE-512 (0.592→0.1200.592\!→\!0.120) but hurts MoRAE (0.089→0.2210.089\!→\!0.221; Table 6(a)). AE-512 lies beyond the collapse onset identified by Probe 1, so factorization avoids one global transport through its weak directions. Once MoRAE restores stable geometry, this benefit disappears and the factorization becomes a net cost. Autoregression is therefore a workaround for poor latent geometry, not an intrinsic requirement of motion generation. Reconstruction does not predict generation. rFID changes little across bottleneck widths, whereas gFID is minimized at z=32z\!=\!32 (Table 6(b)). Over-compression (z=16z\!=\!16) removes conditional detail and has the lowest optimal guidance scale; under-compression (=768 H\!=\!768) preserves R-precision but generates worst, with nearly twice the GT jerk. Width therefore governs implicit motion validity, not merely semantic information. (a) Generator × latent latent full-sequence latent flow matching masked-AR AE-512 0.592 0.120 MoRAE z=32z\!=\!32 0.089 0.221 (b) Bottleneck width latent rFID↓ gFID↓ z=16z\!=\!16 0.090 0.228 z=32z\!=\!32 0.068 0.089 z=64z\!=\!64 0.066 0.320 =768 H\!=\!768 0.034 0.358 Table 6: Core ablations under the diagnostic protocol. Conditioning is necessary, not sufficient. The compact tokenizer moves AE-512 into the transport-stable regime (0.592→0.0890.592\!→\!0.089 gFID; Table 7), but the plateau is a floor, not a ranking. The non-SSL tokenizer is better conditioned than AE-32 (38.538.5 vs. 136136) yet generates worse, with severe jitter hidden by pooled FID. Conversely, AE-32 has the lowest residual alignment but an order-of-magnitude larger decoder gain, so its errors remain costly after decoding. Thus, κflow _flow determines transport stability, while ρalign _align and gisog_iso determine the decoded cost of the remaining error (Sec. 3.5). tokenizer κflow _flow gFID↓ Jerk (→ ) AE-512 7.2×1047.2×10^4 0.592 32.5 AE-32 136 0.127 40.3 two-stage w/o SSL 38.5 0.152 69.4 MoRAE 27 0.089 40.6 GT – 0.000 43.1 Table 7: Tokenizer ablations under the final protocol. 6 Conclusion MoRAE shows that semantic quality alone does not make a motion latent generation-ready. Compact, well-conditioned JEPA distillation with motion-coupled decoding enables a standard non-autoregressive flow to achieve state-of-the-art results. References M. Assran, Q. Duval, I. Misra, P. Bojanowski, P. Vincent, M. Rabbat, Y. LeCun, and N. Ballas (2023) Self-supervised learning from images with a joint-embedding predictive architecture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Cited by: §1, §2, §2, §4.1. M. Caron, H. Touvron, I. Misra, H. Jégou, J. Mairal, P. Bojanowski, and A. Joulin (2021) Emerging properties in self-supervised vision transformers. External Links: 2104.14294, Link Cited by: §2. X. Chen, B. Jiang, W. Liu, Z. Huang, B. Fu, T. Chen, and G. Yu (2023) Executing your commands via motion diffusion in latent space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Cited by: §2, Table 1. V. De Bortoli (2022) Convergence of denoising diffusion models under the manifold hypothesis. Transactions on Machine Learning Research (TMLR). Cited by: §1. C. Guo, Y. Mu, M. G. Javed, S. Wang, and L. Cheng (2024a) MoMask: generative masked modeling of 3D human motions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Cited by: §2. C. Guo, Y. Mu, X. Zuo, P. Dai, Y. Yan, J. Lu, and L. Cheng (2024b) Generative human motion stylization in latent space. arXiv preprint arXiv:2401.13505. Cited by: §2. C. Guo, S. Zou, X. Zuo, S. Wang, W. Ji, X. Li, and L. Cheng (2022) Generating diverse and natural 3D human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Cited by: §1, §5.1. K. He, X. Chen, S. Xie, Y. Li, P. Dollár, and R. Girshick (2022) Masked autoencoders are scalable vision learners. External Links: 2111.06377, Link Cited by: §2. J. Kim, J. Kim, and S. Choi (2023) Flame: free-form language-based motion synthesis & editing. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 37, p. 8255–8263. Cited by: §2. H. Liang, W. Zhang, W. Li, J. Yu, and L. Xu (2024) Intergen: diffusion-based multi-human motion generation under complex interactions. International Journal of Computer Vision, p. 1–21. Cited by: §2. Y. Lipman, R. T. Q. Chen, H. Ben-Hamu, M. Nickel, and M. Le (2023) Flow matching for generative modeling. In International Conference on Learning Representations (ICLR), Cited by: §1, §4.3. N. Ma, M. Goldstein, M. S. Albergo, N. M. Boffi, E. Vanden-Eijnden, and S. Xie (2024) SiT: exploring flow and diffusion-based generative models with scalable interpolant transformers. In European Conference on Computer Vision (ECCV), Cited by: §1, §4.3, §4.3. Z. Meng, Y. Xie, X. Peng, Z. Han, and H. Jiang (2024) Rethinking diffusion for text-driven human motion generation: redundant representations, evaluation, and masked autoregression. arXiv preprint arXiv:2411.16575. Cited by: §2, §3.5, §3.5, Table 1, §5.1. M. Oquab, T. Darcet, T. Moutakanni, H. Vo, M. Szafraniec, V. Khalidov, et al. (2024) DINOv2: learning robust visual features without supervision. External Links: 2304.07193, Link Cited by: §1, §2. W. Peebles and S. Xie (2023) Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), Cited by: §4.3. J. Pidstrigach (2022) Score-based generative models detect manifolds. In Advances in Neural Information Processing Systems (NeurIPS), Cited by: §1. E. Pinyoanuntapong, P. Wang, M. Lee, and C. Chen (2024) M: generative masked motion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Cited by: §2. M. Plappert, C. Mandery, and T. Asfour (2016) The kit motion-language dataset. Big data 4 (4), p. 236–252. Cited by: §1, §5.1. A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever (2021) Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), Cited by: §4.3, §5.1. J. Singh, B. Zheng, Z. Wu, R. Zhang, E. Shechtman, and S. Xie (2026) Improved baselines with representation autoencoders. External Links: 2605.18324, Link Cited by: §1, §2. G. Tevet, S. Raab, B. Gordon, Y. Shafir, D. Cohen-Or, and A. H. Bermano (2023) Human motion diffusion model. In International Conference on Learning Representations (ICLR), Cited by: §2, §2. A. van den Oord, O. Vinyals, and K. Kavukcuoglu (2017) Neural discrete representation learning. In Advances in Neural Information Processing Systems (NeurIPS), Cited by: §3.5. J. Wang, Y. Rong, J. Liu, S. Yan, D. Lin, and B. Dai (2022) Towards diverse and natural scene-aware 3d human motion synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 20460–20469. Cited by: §2. Y. Xie, V. Jampani, L. Zhong, D. Sun, and H. Jiang (2024) OmniControl: control any joint at any time for human motion generation. In The Twelfth International Conference on Learning Representations, Cited by: §2. S. Xu, Z. Li, Y. Wang, and L. Gui (2023) InterDiff: generating 3d human-object interactions with physics-informed diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), p. 14928–14940. Cited by: §2. J. Zhang, Y. Zhang, X. Cun, Y. Zhang, H. Zhao, H. Lu, X. Shen, and Y. Shan (2023) Generating human motion from textual descriptions with discrete representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 14730–14740. Cited by: §2. M. Zhang, Z. Cai, L. Pan, F. Hong, X. Guo, L. Yang, and Z. Liu (2022) MotionDiffuse: text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001. Cited by: §2. B. Zheng, N. Ma, S. Tong, and S. Xie (2025) Diffusion transformers with representation autoencoders. External Links: 2510.11690, Link Cited by: §1, §2.