X-MoGen: Unified Cross-Species Motion Generation

• X-MoGen is a unified cross-species text-driven motion generation framework that creates realistic 3D motion sequences with species-specific anatomical fidelity.
• It standardizes diverse human and animal skeletal structures into a shared 25-joint topology, leveraging UniMo4D’s benchmark data for enhanced cross-species learning.
• Its two-stage architecture employs a CGAE for canonical T-pose priors and a masked Transformer for diffusion-based motion completion, ensuring morphological consistency.

Searching arXiv for the specified X-MoGen paper and closely related work to ground the article in current literature. X-MoGen is a unified text-driven motion generation framework for both humans and animals, introduced as the first unified framework for cross-species text-driven motion generation covering both humans and animals (Wang et al., 7 Aug 2025). It addresses a setting in which a free-form text prompt must be converted into a plausible 3D motion sequence for a character that may be a human or one of many animal species, while simultaneously satisfying text alignment, motion realism, and morphological consistency. Its central premise is that if human and animal motions can be standardized into a shared skeletal topology and encoded in a shared latent space, then a single model can learn motion semantics more structurally, transfer better across species, and generalize to unseen morphologies (Wang et al., 7 Aug 2025).

1. Problem formulation and cross-species motivation

X-MoGen studies text-driven motion generation under a harder setting than standard text-to-motion because the model must accommodate large morphological variation across species, different kinematic statistics even for similar actions, topology mismatch between humans and animals, and unseen species generalization (Wang et al., 7 Aug 2025). The task is not merely to synthesize motion matching an action phrase; it must also preserve species-specific body shape, limb proportions, and topology.

The framework is motivated by the limitations of prior systems that model human and animal motion separately. The data identify human-only methods such as MDM, T2M-GPT, MoMask, and MARDM, and animal-only methods such as AniMo (Wang et al., 7 Aug 2025). This separation creates disjoint representation spaces because human and animal datasets typically use incompatible skeletons and joint conventions. It also impedes cross-species transfer: a model trained only on humans cannot capture quadruped motion dynamics, and an animal-only model cannot robustly generalize to humans or new animal species.

A central technical difficulty is that morphology matters at multiple levels. Even within animals, body lengths, limb counts, tails, and posture styles vary substantially. The paper states that models ignoring these differences often generate anatomically implausible motion (Wang et al., 7 Aug 2025). This suggests that unified modeling is not primarily a question of scaling up a conventional text-to-motion architecture; it requires explicit mechanisms for encoding skeletal structure and regularizing generation toward species-appropriate geometry.

2. UniMo4D and the shared skeletal topology

Unified learning in X-MoGen is enabled by UniMo4D, a large-scale benchmark constructed by combining HumanML3D, KIT-ML, and AniMo4D (Wang et al., 7 Aug 2025). UniMo4D contains 115 species total, consisting of humans and 114 animal species, with 118,663 motion sequences; the abstract rounds this to 119k. Motion lengths are retained in the range 18 to 300 frames.

A key contribution is standardization into a common 25-joint skeleton. KIT-ML joints are remapped, AniMo4D joints are reordered to align with HumanML3D conventions, bone lengths are scaled accordingly, and three virtual tail joints are added to accommodate diverse morphologies; for tailless species, these tail joints are initialized at the pelvic root (Wang et al., 7 Aug 2025). The result is a consistent topology across all motions, allowing joint training without changing the model input or output structure.

The dataset is also structured to support generalization studies. Five species with 5,179 sequences are held out as an unseen-species test set (Wang et al., 7 Aug 2025). This makes unseen-species evaluation possible rather than incidental. The significance of UniMo4D is therefore methodological as well as empirical: it enables joint human-animal training, provides sufficient diversity to learn species-conditional priors, and standardizes representation for a shared latent space. A plausible implication is that the dataset serves as both an infrastructure contribution and an architectural precondition, since the rest of X-MoGen depends on a unified skeletal topology.

3. Two-stage architecture

X-MoGen adopts a two-stage architecture separating shape modeling from motion generation (Wang et al., 7 Aug 2025). In Stage 1, a Conditional Graph Variational Autoencoder (CGAE) learns canonical T-pose priors, while an autoencoder (AE) encodes motion into a compact shared latent space with morphological reconstruction loss. In Stage 2, masked motion modeling with a Masked Transformer (M-Trans) generates latent motion embeddings conditioned on text, using a diffusion-based completion head and a Morphological Consistency Module (MCM) to promote anatomical plausibility.

This decomposition is explicit in the pipeline description. Stage 1 learns how species are shaped and how motions can be compressed; Stage 2 learns how to generate motions from text while respecting those learned constraints (Wang et al., 7 Aug 2025). The architecture therefore treats morphology as a prior and as a regularizer rather than leaving it to emerge implicitly from text-motion co-occurrence.

The two-stage design also clarifies the role of latent representations. The AE learns a continuous latent space rather than a quantized codebook-based representation. According to the paper’s ablation table, this AE strongly outperforms VQ, FSQ, and RVQ compressors on UniMo4D reconstruction (Wang et al., 7 Aug 2025). This suggests that continuous latent modeling is better suited to highly diverse cross-species motion data, where quantization artifacts may be especially harmful for rare morphologies and fine-grained skeletal variation.

4. Morphology modeling in Stage 1

The CGAE models species-specific skeletal morphology using bone-length vectors extracted from canonical T-poses. Its inputs are a bone-length vector $\mathbf{b} \in \mathbb{R}^{N_b}$ and a species semantic condition $\mathbf{c}$ obtained from the BioCLIP-2 text encoder (Wang et al., 7 Aug 2025). The skeletal structure is explicitly represented as a graph, and both encoder and decoder use graph convolutional layers. The goal is to learn the conditional latent distribution

$$q_\phi(\mathbf{z}_{\text{pose} \mid \mathbf{b}, \mathbf{c}})$$

with prior

$p(\mathbf{z}_{\text{pose}}) = \mathcal{N}(0, I).$

Its loss is given as

$$\mathcal{L}_{\text{CGAE}} = \|\hat{\mathbf{b}} - \mathbf{b}\|_2^2 + \beta \, D_{\text{KL}}\!\left( q_\phi(\mathbf{z}_{\text{pose} \mid \mathbf{b}, \mathbf{c}}) \,\|\, p(\mathbf{z}_{\text{pose}}) \right).$$

Once trained, the CGAE can sample a latent code $\mathbf{z}_{\text{pose}}$ and decode a canonical T-pose bone-length vector $\hat{\mathbf{b}}$, from which joint positions $\mathbf{P}$ are recovered by forward kinematics using the predefined kinematic tree (Wang et al., 7 Aug 2025). The paper identifies this as a species-aware canonical prior for morphology, especially useful for unseen quadrupeds.

The motion AE compresses raw motion sequences

$\mathbf{X} = [\mathbf{x}_1, \ldots, \mathbf{x}_L]$

into latent motion sequences

$$\mathbf{Z} = [\mathbf{z}_1, \ldots, \mathbf{z}_T] \in \mathbb{R}^{T \times d},$$

with temporal downsampling

$\mathbf{c}$0

The encoder uses convolutional layers with strides and residual connections, and the decoder uses transposed convolutions (Wang et al., 7 Aug 2025). Its loss is

$\mathbf{c}$1

where

$\mathbf{c}$2

and

$\mathbf{c}$3

Here $\mathbf{c}$4 extracts bone lengths. The morphological reconstruction term is intended to preserve morphology in the latent motion encoding. The paper states that this continuous latent space preserves motion details and morphology better across diverse species than quantized alternatives (Wang et al., 7 Aug 2025).

5. Text-conditioned generation and morphological guidance

Stage 2 performs masked motion modeling in latent space. Given latent motion tokens $\mathbf{c}$5, a random subset of positions

$\mathbf{c}$6

is masked and replaced by a learnable token $\mathbf{c}$7 (Wang et al., 7 Aug 2025). The canonical T-pose embedding $\mathbf{c}$8 is produced from the CGAE and concatenated as the first token, giving the query

$\mathbf{c}$9

Text conditioning uses CLIP text features: a sentence feature $$q_\phi(\mathbf{z}_{\text{pose} \mid \mathbf{b}, \mathbf{c}})$$0 and word features $$q_\phi(\mathbf{z}_{\text{pose} \mid \mathbf{b}, \mathbf{c}})$$1, with

$$q_\phi(\mathbf{z}_{\text{pose} \mid \mathbf{b}, \mathbf{c}})$$2

Fusion is then performed as

$$q_\phi(\mathbf{z}_{\text{pose} \mid \mathbf{b}, \mathbf{c}})$$3

The Transformer output conditions a diffusion-like completion module implemented by MLPs as a velocity network $$q_\phi(\mathbf{z}_{\text{pose} \mid \mathbf{b}, \mathbf{c}})$$4 (Wang et al., 7 Aug 2025). For each masked position $$q_\phi(\mathbf{z}_{\text{pose} \mid \mathbf{b}, \mathbf{c}})$$5, the model predicts a deterministic velocity field from noisy latent $$q_\phi(\mathbf{z}_{\text{pose} \mid \mathbf{b}, \mathbf{c}})$$6 to target latent $$q_\phi(\mathbf{z}_{\text{pose} \mid \mathbf{b}, \mathbf{c}})$$7 using flow matching. The interpolated latent state is

$$q_\phi(\mathbf{z}_{\text{pose} \mid \mathbf{b}, \mathbf{c}})$$8

and the flow loss is

$$q_\phi(\mathbf{z}_{\text{pose} \mid \mathbf{b}, \mathbf{c}})$$9

Morphological plausibility during generation is enforced by the Morphological Consistency Module. The MCM is an auxiliary regularizer that predicts skeletal bone lengths from latent motion: a GRU encodes the latent sequence $p(\mathbf{z}_{\text{pose}}) = \mathcal{N}(0, I).$0, the final hidden state passes through an MLP, and the output is a predicted bone-length vector (Wang et al., 7 Aug 2025). The MCM is pre-trained on the AE latent space and frozen during generation. Its guidance loss is

$p(\mathbf{z}_{\text{pose}}) = \mathcal{N}(0, I).$1

where $p(\mathbf{z}_{\text{pose}}) = \mathcal{N}(0, I).$2 is the canonical ground-truth bone-length vector. An important training detail is that gradients from this loss are backpropagated only through masked positions $p(\mathbf{z}_{\text{pose}}) = \mathcal{N}(0, I).$3, while unmasked positions are detached. The total generation loss is

$p(\mathbf{z}_{\text{pose}}) = \mathcal{N}(0, I).$4

At inference, all positions in $p(\mathbf{z}_{\text{pose}}) = \mathcal{N}(0, I).$5 are initialized with mask tokens, subsets of masked positions are iteratively filled over $p(\mathbf{z}_{\text{pose}}) = \mathcal{N}(0, I).$6 rounds, and for each selected position the method samples noise and runs $p(\mathbf{z}_{\text{pose}}) = \mathcal{N}(0, I).$7 denoising steps (Wang et al., 7 Aug 2025). Classifier-free guidance is used:

$p(\mathbf{z}_{\text{pose}}) = \mathcal{N}(0, I).$8

where $p(\mathbf{z}_{\text{pose}}) = \mathcal{N}(0, I).$9 is the guidance scale. The completed latent motion is then decoded through the AE decoder to obtain the final motion sequence. The paper notes that this iterative masked filling also enables cross-species motion transition.

6. Experimental evaluation, ablations, and qualitative behavior

X-MoGen is evaluated against MDM, T2M-GPT, AttT2M, MMM, MoMask, MARDM, and AniMo, with all baselines retrained on UniMo4D for a fair comparison (Wang et al., 7 Aug 2025). A text-motion retrieval model retrained on UniMo4D is used as the evaluation backbone. The reported metrics are R-Precision (Top-1, Top-2, Top-3), FID, MM-Dist, Diversity, and MME, where MME is Mean Morphological Error:

$$\mathcal{L}_{\text{CGAE}} = \|\hat{\mathbf{b}} - \mathbf{b}\|_2^2 + \beta \, D_{\text{KL}}\!\left( q_\phi(\mathbf{z}_{\text{pose} \mid \mathbf{b}, \mathbf{c}}) \,\|\, p(\mathbf{z}_{\text{pose}}) \right).$$0

Results are averaged over 10 runs and reported with 95% confidence intervals.

On the standard UniMo4D test split, X-MoGen achieves R-Precision Top-1 of 0.848, FID of 0.050, MM-Dist of 0.742, Diversity of 19.332, and MME of 0.201 (Wang et al., 7 Aug 2025). The paper states that it achieves the best results across all main metrics, outperforming the strongest baselines such as MARDM in R-Precision and FID, performing far better than AniMo on the unified dataset, and showing especially strong morphological consistency as reflected by low MME. On the unseen-species test set, performance decreases but remains strong: R-Precision Top-1 of 0.148, FID of 19.935, MM-Dist of 7.939, Diversity of 17.652, and MME of 0.229 (Wang et al., 7 Aug 2025). The unseen-species setting is explicitly described as substantially harder.

The ablation study removes word-level CLIP features $$\mathcal{L}_{\text{CGAE}} = \|\hat{\mathbf{b}} - \mathbf{b}\|_2^2 + \beta \, D_{\text{KL}}\!\left( q_\phi(\mathbf{z}_{\text{pose} \mid \mathbf{b}, \mathbf{c}}) \,\|\, p(\mathbf{z}_{\text{pose}}) \right).$$1, the CGAE T-pose prior, $$\mathcal{L}_{\text{CGAE}} = \|\hat{\mathbf{b}} - \mathbf{b}\|_2^2 + \beta \, D_{\text{KL}}\!\left( q_\phi(\mathbf{z}_{\text{pose} \mid \mathbf{b}, \mathbf{c}}) \,\|\, p(\mathbf{z}_{\text{pose}}) \right).$$2, and the MCM guidance loss (Wang et al., 7 Aug 2025). Removing word-level features slightly hurts FID and alignment. Removing CGAE worsens both FID and MME, indicating the importance of species-aware T-pose priors. Removing the morphological reconstruction loss significantly degrades FID, MM-Dist, and MME. Removing MCM barely changes FID but noticeably worsens MME, showing that it is especially important for anatomical correctness. The paper characterizes these components as complementary: CGAE handles morphology prior, AE preserves motion detail, MCM enforces skeletal plausibility during generation, and text features improve semantic precision.

Qualitative comparisons are reported against MoMask, MARDM, and AniMo. For a human example with actions such as “running” and “cartwheel,” X-MoGen correctly preserves action order and transitions smoothly, whereas MoMask reverses the order and MARDM and AniMo omit part of the prompt, generating only the cartwheel (Wang et al., 7 Aug 2025). For an unseen animal example, X-MoGen produces a semantically faithful and morphologically consistent animal motion, whereas MoMask degenerates into a human-like motion, MARDM has unstable rotations, and AniMo misinterprets the action. These examples support the claim that the model is robust to species variation and better preserves both semantics and anatomy.

7. Significance, limitations, and related terminology

The paper identifies four main contributions: the first unified text-driven motion generation framework across humans and animals; the UniMo4D dataset with 115 species and 118,663 motion sequences under a standardized skeletal topology; a two-stage morphology-aware generation pipeline based on CGAE, AE, and masked motion modeling plus diffusion; and morphological regularization through both morphological reconstruction loss and the Morphological Consistency Module (Wang et al., 7 Aug 2025). The overall significance claimed is that unified modeling is not only possible but beneficial: a shared representation of human and animal motion yields strong text alignment, realistic motion generation, and better generalization to unseen species.

Several limitations are implied rather than presented in a long explicit limitations section. The method depends on a standardized 25-joint topology, which may simplify or distort some species-specific structures. Hard unseen-species generalization remains difficult, as indicated by the noticeable performance drop on the unseen-species split. Morphological modeling is bone-length based, capturing skeleton scale and proportions but not necessarily complex dynamics such as muscle-driven motion, gait subtleties, or soft tissue effects. Cross-species transitions are described as promising but not perfect, and dataset coverage, though large, is still limited to 115 species and the available captured motion diversity (Wang et al., 7 Aug 2025).

The terminology surrounding “X-MoGen” can invite confusion with later motion-generation systems. The related paper “Re$$\mathcal{L}_{\text{CGAE}} = \|\hat{\mathbf{b}} - \mathbf{b}\|_2^2 + \beta \, D_{\text{KL}}\!\left( q_\phi(\mathbf{z}_{\text{pose} \mid \mathbf{b}, \mathbf{c}}) \,\|\, p(\mathbf{z}_{\text{pose}}) \right).$$3MoGen: Open-Vocabulary Motion Generation via LLM Reasoning and Physics-Aware Refinement” is described as related to, but not identical with, “X-MoGen” (Zheng et al., 20 Apr 2026). Re$$\mathcal{L}_{\text{CGAE}} = \|\hat{\mathbf{b}} - \mathbf{b}\|_2^2 + \beta \, D_{\text{KL}}\!\left( q_\phi(\mathbf{z}_{\text{pose} \mid \mathbf{b}, \mathbf{c}}) \,\|\, p(\mathbf{z}_{\text{pose}}) \right).$$4MoGen focuses on open-vocabulary human text-to-motion generation using MCTS-enhanced LLM reasoning, pose prior optimization, spatiotemporal completion, and RL post-training for physics-aware refinement, whereas X-MoGen is specifically defined as a unified cross-species framework spanning humans and animals. This distinction is important because the two systems address different generalization problems: X-MoGen targets unification across morphology and species, while Re$$\mathcal{L}_{\text{CGAE}} = \|\hat{\mathbf{b}} - \mathbf{b}\|_2^2 + \beta \, D_{\text{KL}}\!\left( q_\phi(\mathbf{z}_{\text{pose} \mid \mathbf{b}, \mathbf{c}}) \,\|\, p(\mathbf{z}_{\text{pose}}) \right).$$5MoGen targets open-vocabulary generalization for human motion.