T2M generation aims to synthesize 3D human motion sequences from natural language descriptions. Early studies relied on conditioning with simple action classes or text labels [17,18]. As the field expanded to include free-form natural language instructions and required precise semantic alignment between text and motion, controllability became a key challenge. Initial approaches employed variational autoencoders (VAEs) [19] to align text and motion in a shared embedding space [5,6] or to learn text–motion correspondence via conditional decoding [4]. Tevet et al. [7] used visual priors by encoding motion descriptions with a frozen Contrastive Language-Image Pre-training (CLIP) [20] text encoder and aligning the motion latent space accordingly. More recently, diffusion models [21] have been widely adopted in T2M due to their strong generation quality and temporal stability. The Motion Diffusion Model (MDM) [7] introduced a Transformer-based [22] denoising network that directly models temporal dependencies. Chen et al. [13] improved efficiency and quality by leveraging latent-space diffusion with optimized sampling. However, these methods often apply the same text condition across all body parts, which makes it difficult to capture accurate part-level semantic mappings and leads to mismatches in part motion [14,15]. In parallel, token-based approaches tokenize continuous motion into codebook indices using vector quantization (VQ) [23] and treat it as a discrete sequence modeling problem. Guo et al. [9] formulated text-to-motion as a discrete sequence generation problem, and Zhang et al. [10] improved token prediction via autoregressive generation. More recently, Guo et al. [11] combined hierarchical quantization with bidirectional masked prediction, mitigating the error accumulation inherent in sequential autoregressive generation. However, tokenization can result in information loss when compressing continuous motion into a limited codebook, potentially restricting motion diversity and fidelity [24]. Although generation architectures have evolved in diverse directions, from a conditioning perspective, most methods still share the limitation of relying on fixed-length global embeddings, such as the CLIP [CLS] token. This reliance dilutes fine-grained semantics for specific body parts in composite prompts, motivating the need for conditioning mechanisms that faithfully reflect part-level semantics while maintaining expressiveness.

Ghosh et al. [3] began with coarse divisions such as upper and lower body, treating each part through a separate generation pathway. Athanasiou et al. [16] automated spatial composition by using GPT-3 to extract action-body part mappings and combining independently generated part-wise motions post-hoc. Subsequent research has focused on increasing the number of segmented parts to achieve finer motion representation. Part-Coordinating (ParCo) [14] decomposes the full body into multiple semantic parts and performs part-wise generation while considering inter-part coordination. More recent work has combined part-based modeling with text-based motion generation, directly connecting detailed elements to part-level representations. The Local-to-Global pipeline for Text-to-Motion generation (LGTM) [15] decomposes global text into part-specific descriptions to strengthen semantic alignment at the part level. Wang et al. [25] extract body-part-related cues through large language model (LLM)-based semantic parsing and reflect sentence structure to condition detailed semantics more precisely. Fan et al. [26] identify salient body parts and enforce semantic alignment for interaction motion. Improvements have also been attempted from the perspective of the conditioning mechanism itself. Chang et al. [27] introduce a composite-aware text encoder and a text-motion aligner, enabling dynamic word-level correspondence rather than fixed-length global embeddings. Li and Feng [28] improve the generation accuracy of composite actions by simultaneously leveraging coarse- and fine-grained descriptions. These studies supply fine-grained cues that would otherwise be lost in a global summary embedding, enabling body-part actions specified in text to be more faithfully reflected in the generated motion. However, existing methods combine global and local conditions at fixed ratios or perform partial generation and global optimization as separate stages, which can cause one condition to overshadow the other. Additionally, post-hoc composition approaches do not incorporate inter-part coordination into the generation process itself, and learning-based part-wise generation methods that inject part conditions independently through dedicated pathways struggle to capture full-body context. A mechanism that dynamically adjusts the relative contributions of global and local conditions as generation progresses has yet to be explored.

The challenge of fine-grained semantics being diluted under global conditioning has been widely recognized in conditional generative modeling. In text-to-image synthesis, efforts to address this issue have included decomposing composite prompts into per-concept diffusion processes [29] and leveraging linguistic structure to restructure cross-attention [30]. Zarei et al. [31] demonstrated that the output space of CLIP is suboptimal for compositional prompts, showing that attention contributions from unrelated tokens are mixed into the final token embeddings, leading to failures in attribute-object binding. In the facial synthesis domain, Song et al. [32] improved attribute-level alignment under multi-attribute conditions by introducing a module that dynamically balances global text features and local attribute features through learnable gating. It has also been observed that the role of conditioning changes qualitatively across stages of the diffusion process. Balaji et al. [33] empirically showed that text-to-image diffusion models rely heavily on text conditioning during early denoising but largely ignore it in later stages. More recently, Cho et al. [34] argued that static conditioning cannot flexibly adapt to the dynamic nature of multi-stage denoising, which evolves from coarse structure to fine detail, and proposed TC-LoRA to dynamically adjust conditioning based on both the denoising timestep and the control signal. These studies consistently demonstrate that hierarchical semantic decomposition and dynamic conditioning are critical to generation quality, yet these insights remain largely unexplored in T2M. In this paper, we combine such adaptive conditioning strategies with part-level semantic decomposition for T2M and show that timestep-conditioned gating improves motion-text alignment.

Recent text-to-motion research, particularly following HumanML3D [4], typically represents full-body motion using a canonical pose that incorporates root information and joint features relative to the root. A motion sequence consists of F frames with J joints, where each frame xi=[r˙a,r˙xz,rh,jp,jr,jv,cf]. In this formulation, r˙a∈R1 denotes the root yaw angular velocity, r˙xz∈R2 denotes the root linear velocity on the XZ plane, and rh∈R1 denotes the root height. The features jp∈R3×(J−1),jr∈R6×(J−1), and jv∈R3×J represent joint position, rotation, and velocity in the root coordinate system, respectively, while cf∈R4 is a binary feature indicating foot-ground contact. This representation encodes both global trajectory and relative joint motion, and is widely adopted for stable modeling of diverse motions. Building on this canonical representation, our approach reorganizes motion into semantic parts, reflecting the fact that text descriptions frequently target specific body parts. We partition joints into six groups (root, backbone, left arm, right arm, left leg, right leg), and construct part-wise motion by selecting only the joint features (position, rotation, velocity, and foot contact) for each group.

We propose the EGN, which consists of a transformation module and a GlocalGate, as illustrated in Fig. 1a. In diffusion-based T2M models, text is typically condensed into a single global embedding, which can obscure fine-grained cues in complex sentences. Motivated by this, we separate text conditions into global and local components. The global text Tg represents the original full-body description, providing overall action identity and coarse context. Part-specific local texts {Tl,p}p=1P are derived by decomposing Tg into short action phrases for each of the six body parts using an LLM. For instance, given the global description “a person bends forward and picks up an object in their right hand,” the LLM assigns “bends forward” to the pelvis and “picks up an object” to the right arm, while the remaining parts receive a null descriptor indicating no part-specific action. Each text input is independently encoded by the pretrained CLIP text encoder, extracting the end-of-sequence token representation as a fixed-dimensional embedding. This process yields a global embedding egCLIP∈RD and part-specific local embeddings {el,pCLIP∈RD}p=1P. Since the text encoder’s embedding space may not align with the conditioning space required for motion generation, a transformation module is introduced to expand the representational capacity of both global and local embeddings. This module is implemented as a residual feed-forward adapter that learns motion-relevant correction terms on top of the original CLIP embeddings. We formulate this transformation as a residual function, allowing the embeddings to be progressively adapted for motion conditioning while retaining the original CLIP embeddings as a semantic anchor. The final conditioning embeddings are obtained as (1)eg=egCLIP+fg(egCLIP),el=elCLIP+fl(elCLIP)where fg and fl denote the learnable nonlinear mappings for global and local conditions, respectively.

Diffusion models are known to recover global structures at high noise levels and fine-grained details at low noise levels [35,36]. We therefore hypothesize that the relative importance of global and local conditions shifts across timesteps and generation states. Based on this hypothesis, we dynamically modulate their contributions rather than combining them at a fixed ratio. Accordingly, GlocalGate predicts global and local weights from the noisy motion xt and timestep t. This module is defined as:(2)[wg,wl]=Linear(SiLU(Linear([xt,emb(t),t]))),where [⋅] denotes concatenation and emb(t) represents the timestep embedding. The output logits wg and wl are subsequently normalized via softmax to yield the final gating coefficients:(3)αg=exp⁡(wg)exp⁡(wg)+exp⁡(wl),αl=exp⁡(wl)exp⁡(wg)+exp⁡(wl).

These gating coefficients, αg(t) and αl(t), determine the relative contribution of global and local semantics at each timestep. Using these coefficients, we compute the part-specific condition cp(t) for part p as the weighted sum of global and local embeddings:(4)cp(t)=αg(t)⋅eg+αl,p(t)⋅el,p.

GlocalGate thus adaptively balances global and local influences based on its inputs. Specifically, the noisy motion xt conveys the current geometric state and pose configuration, while the timestep t encodes the prevailing noise level. The learned coefficients αg and αl therefore shift the balance between global coherence and local specificity at each generation step. The EGN integrates both the transformation module and GlocalGate to produce the final part-specific conditioning embedding cp(t).

Our approach maintains part-wise pathways within a single model: each part representation is generated under its global-local embedding, after which a lightweight fusion step exchanges inter-part information, as illustrated in Fig. 1b. Accordingly, we separate the network into part-wise attention for conditional generation and PartFuse(⋅) for part-wise fusion. At each timestep t, the generation process is divided into two specialized stages: semantic synthesis and structural coordination. Initially, part-wise attention utilizes the part-specific condition cp(t) to update the latent representation of each part p as follows:(5)h^tp=Attention(htp,cp(t)),where h^tp represents the semantically updated state focusing on synthesizing the local action structure. Subsequently, PartFuse(⋅) enables controlled inter-part coordination while maintaining the separation of generation pathways. It is defined for each part i as follows:(6)PartFuse(h^ti;[h^t1∣⋯∣h^tj]j≠i)=LN(h^ti+MLPi∘LN([h^t1∣⋯∣h^tj]j≠i)),where MLPi functions as a part-specific alignment module that incorporates features from other parts into the current pathway.

A key advantage is the explicit separation, at the layer level, between part-wise attention and PartFuse(⋅). Decoupling semantic interpretation from inter-part coordination introduces an inductive bias that keeps part representations distinct. This architectural choice stabilizes training and provides consistent part-level control, even for complex textual conditions involving multiple parts.

Dataset. We evaluate our method on HumanML3D [4], a widely used public benchmark for text-to-motion generation. HumanML3D consists of 3D human motion sequences paired with natural language annotations and has served as a standard dataset in prior T2M studies. It contains 14,616 motion sequences and 44,970 text descriptions. Since global text annotations in HumanML3D are often insufficient for part-level semantic alignment, we augment the dataset with local texts derived from the global annotations using the decomposition procedure described in Section 3.2.

Metrics. We evaluate motion quality and text–motion alignment using five metrics: (1) R-Precision measures motion–text retrieval accuracy in the feature space of a pretrained T2M evaluation network. For each motion sequence, we compute precision based on whether the ground-truth text is ranked within Top-1/2/3 among 32 candidate texts. (2) Fréchet Inception Distance (FID) measures the distributional gap between generated and real motions by computing FID on motion features extracted by the T2M evaluator. (3) Multi-Modal Distance (MM-Dist) computes the average Euclidean distance between each text feature and the motion features generated from that text. (4) Diversity splits generated motions into two random subsets of equal size and computes the average Euclidean distance between their motion features [4,5,37]. (5) Part-level Multi-Modal Similarity (PMM Sim), adopted from Sun et al. [15], trains part-level text and motion encoders with contrastive learning following Petrovich et al. [38] and measures the correspondence between part-specific text descriptions and generated part motions.

Baselines. We select three baselines to evaluate our part-wise diffusion generation framework, which jointly uses global and local text conditions: (1) MDM, a diffusion-based T2M model conditioned on global text, as a reference for full-body diffusion performance; (2) ParCo, a part-wise generation approach with separate body-part pathways, to assess the impact of part-wise generation on performance and consistency; and (3) LGTM, which decomposes global text into part-wise descriptions, to evaluate the effect of part-wise text conditioning.

Implementation Details. We employ Qwen2.5-7B-Instruct [39] as the LLM for decomposition, prompting it to produce a local text for each of the six body parts given the global annotation. Our model consists of 155M total parameters, of which 21.2M are trainable. Training was conducted on a single NVIDIA RTX 4090 GPU for 83.3 h on the HumanML3D dataset. At inference, the model requires 0.45 s per sample (DDIM [40], 50 steps) and peaks at 1184 MB of GPU memory.

Fig. 2 compares results on composite prompts in which a single sentence specifies a full-body action alongside multiple part-level instructions. MDM largely preserves natural full-body motion, but tends to converge to a similar global motion pattern, yielding limited changes when part instructions vary and occasionally omitting part-specific constraints. ParCo can reflect some parts’ instructions due to part-wise generation, but under composite prompts, inter-part interactions often become unstable, leading to motion collapse or suppression where constraints on one part inhibit the motions of other parts. LGTM improves responsiveness to part instructions by providing part-wise text conditions, but when part conditions conflict or when the fusion with global context is not sufficiently stable, some constraints are weakened, or the overall action consistency degrades. In contrast, our method injects part-wise signals progressively while preserving the global condition throughout denoising, producing motions that preserve full-body coherence while simultaneously satisfying multiple part instructions under composite prompts.

Fig. 3 evaluates how well part-level text descriptions, decomposed from input sentences using an LLM, are preserved during full-body motion synthesis. For example, given the prompt “A person sits down and stretches his legs straight.”, LGTM generates part-wise motions from decomposed part texts and then applies strong global conditioning during the full-body motion optimization stage [15]. In this process, features induced by part conditions (e.g., stretches legs straight) are partially suppressed by the global context (e.g., sits down), leading to cases where the legs converge to a bent posture rather than being fully extended. In contrast, our method maintains the global condition as an anchor while explicitly injecting and aligning part-wise conditions during generation, ensuring full-body consistency while preserving part-level semantics. As a result, local semantics such as “stretches legs straight” are more consistently reflected in the final motion, even in composite sentences.

Overall, our method reflects part-wise conditions more consistently when satisfying multiple part constraints simultaneously in composite prompts. Even in sentences with a strong global context, local prompts decomposed at the part level are stably preserved in the final motion, ensuring that key actions of specific parts appear without being weakened. This observation qualitatively supports the effectiveness of the global-local conditioning scheme, which stabilizes the integration of part-wise conditions throughout the diffusion process while preserving the global condition.

Table 1 presents the quantitative evaluation results of the proposed method and baselines (MDM, ParCo, LGTM) on the HumanML3D test set. From the perspective of text-motion alignment, our method demonstrates consistent improvements across all metrics. Our method achieves substantial gains in R-Precision and MM-Dist compared with MDM and outperforms ParCo and LGTM in text-motion alignment. This quantitatively confirms that part-wise conditioning improves alignment over global-only conditioning, which tends to obscure fine-grained elements in complex sentences. In terms of generation quality (FID), our method shows slightly higher FID than ParCo, yet R-Precision and MM-Dist consistently show improvements. Meng et al. [24] have shown that the standard evaluation protocol can disproportionately favor VQ-based methods over diffusion-based methods. Accordingly, the FID gap between the proposed diffusion-based method and the VQ-based ParCo likely reflects intrinsic differences between the two generation paradigms. Among diffusion-based methods, our method achieves the best FID, suggesting that dynamically modulating global and local conditions throughout generation is effective for improving generation quality. Finally, for Diversity, our model maintains a level of diversity comparable to that of real motions, suggesting that it improves text-motion alignment without sacrificing variation in the generated results.

Table 2 reports the evaluation of part-level semantic alignment using PMM Sim. Our method achieves the highest scores on Left Arm and Right Arm, and remains competitive on Head, ranking second only to ParCo by a narrow margin. This is because arm movements are explicitly described in text annotations, yielding semantically rich local embeddings via LLM-based decomposition, which enables EGN’s gating and dedicated generation pathways to operate effectively. In contrast, performance on Torso, Left Leg, and Right Leg is slightly below LGTM. These parts are less frequently described in text, often receiving null descriptors, which makes the local embeddings semantically sparse and causes EGN’s gating to shift toward the global condition, reducing the benefit of part-specific conditioning. Nevertheless, our method consistently surpasses MDM and ParCo on these parts, maintaining balanced alignment performance across all body parts.

Overall, by preserving the global condition while adaptively injecting part-level conditions during generation, our method improves text-motion semantic alignment over existing baselines and remains stable in generation quality and diversity.

This section analyzes the impact of the key components of the proposed framework on performance. We first examine the contributions of individual components within EGN, then investigate how framework-level design choices—including global descriptions, local descriptions, the PartFuse module, and part-wise generation pathways—affect overall performance.

Table 3 reports the results of separately removing the two core components of EGN: the Transformation module and the Noisy Motion conditioning. “w/o Transformation” removes the residual feed-forward adapter fg and fl (Eq. (1)), directly using the raw CLIP embeddings as conditioning inputs. “w/o Noisy Motion” removes xt from the GlocalGate input (Eq. (2)), so that the gating coefficients are determined solely by timestep information. Removing the Transformation module degrades R-Precision and MM-Dist, while FID rises sharply from 0.1567 to 0.4152. Without a learned transformation, the raw embeddings are insufficient for effective conditioning, destabilizing the generation distribution. Removing the Noisy Motion conditioning also results in a decline in R-Precision and MM-Dist, along with a modest increase in FID. This indicates that reflecting the current noisy motion state in the gating contributes to adaptive condition adjustment as generation progresses. Collectively, these results confirm that both the embedding transformation and the motion-state-based gating within EGN contribute to text-motion alignment and generation quality, and that neither component alone is sufficient.

Table 4 analyzes the contributions of four framework-level components—global descriptions, local descriptions, the PartFuse module, and part-wise generation pathways—by evaluating their combinations. Removing PartFuse results in a slight decline in R-Precision and MM-Dist, while FID increases notably from 0.1567 to 0.2973. This suggests that simply combining independently generated part motions without a dedicated fusion mechanism degrades the quality of the full-body distribution, confirming that PartFuse helps ensure inter-part coherence. When the global description is removed and the generation is based solely on local descriptions, a sharp performance drop is observed across all metrics. This can be attributed to two factors. First, text annotations in the current dataset do not explicitly describe every body part, so relying on local descriptions alone leaves certain parts without conditioning information. Second, without the global context itself, there is no basis for capturing inter-part relationships and the overall semantic structure of the motion, making it difficult for part-wise generators to form coherent full-body motion. When local descriptions are removed, each part-wise generation pathway receives only the same global summary as its condition. Although the pathways are physically separated, the absence of differentiated semantic information prevents part-wise separation from being fully exploited, leading to degradation in both text-motion alignment and generation quality. When part-wise generation pathways are removed and generation relies solely on the global description, text-motion alignment metrics (MM-Dist, R-Precision) decline, and FID also increases. This indicates that relying solely on a global description through a single generation pathway limits the model’s ability to capture fine-grained semantics, and that explicitly separating part-wise generation pathways is effective for improving both alignment and generation quality.

In summary, the global context serves as the structural foundation for overall motion, while local descriptions and part-wise pathways function complementarily to achieve fine-grained semantic alignment. PartFuse integrates independently generated part motions at the full-body level. When these components are combined, both text-motion alignment and generation quality improve jointly.

To quantify the relative contribution of local conditioning with respect to global conditioning at diffusion timestep t, we define α(t) as the ratio of the local gating coefficient to the global gating coefficient, i.e., α(t)=αl(t)/αg(t). Figs. 4 and 5 visualize the changes in α(t) across diffusion timesteps. α(t) is high in the early generation stages and gradually decreases as denoising progresses, indicating that local conditions contribute strongly to part-level structure formation at the beginning and then shift toward refining full-body context under global conditions. This suggests that local information is not merely used as a late-stage detail refinement signal; rather, it serves as a structural constraint in the early steps to shape the outline and skeleton of part-specific actions. As sampling proceeds, the relative contribution of global conditions increases, strengthening overall motion context and full-body consistency. α(t) is not uniform across body parts: parts that are essential for establishing the global action tend to exhibit larger α(t). This discrepancy is particularly pronounced in the early diffusion steps, implying that the gating network prioritizes injecting local signals into globally critical parts to stabilize the formation of the initial motion structure. As timesteps progress, local contributions decrease across all parts, while the early-stage prioritization for globally important parts remains consistent.

Table 5 provides quantitative evidence supporting this interpretation. Removing local injection (α(t)=0) degrades part-level semantic alignment, confirming that explicit local conditioning is necessary for part-level semantic alignment under composite prompts. Mirroring α(t) over timesteps—so that early local contributions shrink and late ones grow—causes an overall performance drop, demonstrating that sufficiently strong local signals in the early stage are crucial. Furthermore, using a fixed α also underperforms the learned schedule, suggesting that a static global–local mixture is insufficient and that timestep-dependent role transition is necessary.

In summary, the proposed EGN functions as a mechanism that uses local signals as early structural constraints to form part-wise outlines and progressively refines full-body consistency through global context.