SnapMoGen: Text-Driven Motion Benchmark

• SnapMoGen is a high-quality motion dataset and benchmark featuring 20,450 clips, 43.7 hours of motion data, and 122,565 detailed text descriptions.
• The dataset provides continuous, long-duration motion segments paired with rich, expressive annotations to improve evaluation of text-conditioned motion models.
• MoMask++ introduces a multi-scale residual quantization and masked-token approach that boosts motion quality and text alignment in human motion generation.

Searching arXiv for SnapMoGen and related motion-generation papers to ground the article in recent literature. SnapMoGen is a text–motion dataset and benchmark introduced for human motion generation from expressive texts, together with the model MoMask++. It was proposed to address a specific limitation in text-to-motion research: many prior systems were trained on short, generic prompts, which constrained fine-grained controllability, generalization to unseen prompts, and the handling of long, realistic motion descriptions. In the primary dataset paper, SnapMoGen is defined as a high-quality motion capture corpus with expressive textual annotations, comprising 20,450 motion clips, 43.7 hours of motion data at 30 fps, and 122,565 text descriptions averaging 48 words per description; the clips preserve original temporal continuity from long recordings, making the benchmark relevant not only for clip-level generation but also for long-term motion generation and blending (Guo et al., 12 Jul 2025).

1. Definition and scope

SnapMoGen occupies two closely related roles in the literature. First, it is a dataset / benchmark for text-to-motion generation. Second, the name is also associated with the source paper titled "SnapMoGen: Human Motion Generation from Expressive Texts", which introduced the benchmark and the MoMask++ baseline (Guo et al., 12 Jul 2025).

As a dataset, SnapMoGen is consistently treated in later work as a modern benchmark for expressive text-driven human motion generation. ScaleMoGen evaluates on HumanML3D and SnapMoGen and characterizes SnapMoGen as a more challenging text-to-motion benchmark with long, expressive descriptions (Hwang et al., 12 May 2026). FlowCoMotion likewise uses SnapMoGen as one of its two main evaluation datasets and describes it as a 43.7-hour, 30 fps, high-quality motion dataset in which each motion segment is accompanied by 6 detailed text descriptions spanning daily tasks, fitness, social interactions, and dance (Guan et al., 13 Apr 2026). CMDM positions SnapMoGen as the benchmark most relevant to temporally continuous, long-horizon, expressive text-conditioned motion under causal / streaming constraints (Yu et al., 26 Feb 2026).

A recurring misconception is to treat SnapMoGen primarily as a model name. In the later literature, its dominant use is as a benchmark dataset. This is especially explicit in ScaleMoGen, where SnapMoGen functions primarily as an evaluation benchmark, even though the prior baseline paper cited in that work bears the title "SnapMoGen: Human Motion Generation from Expressive Texts" (Hwang et al., 12 May 2026).

2. Dataset composition and collection protocol

The dataset paper defines SnapMoGen as a high-quality motion capture dataset with expressive, detailed textual descriptions. It contains 20,450 motion clips, totaling 43.7 hours of motion data at 30 fps, or about 4.7 million frames, paired with 122,565 text descriptions. Each description averages 48 words, compared with 12 words for HumanML3D; SnapMoGen clips average 7.8 seconds, versus 7.1 seconds for HumanML3D (Guo et al., 12 Jul 2025).

The motion content spans daily activities, fitness routines, social interactions, dances, and explicitly stylized performances, including examples such as princess, elderly person, and zombie. Data was captured from 10 participants using Xsens and Rokoko motion capture suits. Collection was guided by LLM-generated action scenarios and curated internet videos and images. Performers were shown these references and asked to perform them in their own style; motions with visible artifacts such as jittering or foot sliding were filtered out (Guo et al., 12 Jul 2025).

A distinguishing feature is the preservation of original temporal continuity from long motion recordings. Rather than assembling only isolated short clips, the dataset was recorded as long sequences containing multiple actions and then segmented into clips. The segmentation principle was to cut at motionless moments. The reported procedure was to compute average positional velocities of the hip and end-effector joints, smooth them with a Gaussian filter, detect velocity troughs, normalize them to obtain $\rho_{1:n}\in[0,1]$, and select trough $i$ as a segmentation point with probability $0.5\rho_i$. The resulting clips satisfy hard constraints of minimum 4 seconds and maximum 12 seconds (Guo et al., 12 Jul 2025).

The annotation pipeline combined human labeling with controlled LLM augmentation. Each clip received 2 manual descriptions from distinct human annotators and 4 LLM-augmented descriptions, yielding about six descriptions per clip. The paper reports 40,859 manual and 81,706 LLM-augmented descriptions, totaling 122,565. The human annotation workforce consisted of 55 professional native English-speaking annotators, instructed to describe action, context, style, moving direction, speed, trajectory shape, body parts involved, spatial relation/location, posture if applicable, and timing if applicable. Manual annotations underwent a second-round review for descriptive accuracy; typographical errors were corrected with an LLM, and the LLM then rewrote each manual description twice while preserving semantics and temporal order (Guo et al., 12 Jul 2025).

3. Representation, benchmark protocol, and evaluation usage

For SnapMoGen, the motion representation in the original paper follows the HumanML3D-style feature design and includes root angular velocity about Y, root linear velocity on XZ, root height, 6D local joint rotations, local joint positions, and local joint velocities. SnapMoGen uses a 24-joint skeleton, producing 296-dimensional pose features, and the authors note that these are directly convertible to standard mocap formats like BVH (Guo et al., 12 Jul 2025).

Later work preserves SnapMoGen as a standard benchmark rather than redefining its representation ad hoc. A unified conditional flow model describes SnapMoGen as containing 20,450 motion clips, 43.7 hours, 122,565 text descriptions, average text length 48 words, clips lasting 4 to 12 seconds at 30 FPS, and a canonical BVH skeleton representation. In that setting, SnapMoGen is explicitly treated as a single-skeleton dataset for fair comparison in text-to-motion evaluation (Li et al., 15 Apr 2026).

Evaluation practice on SnapMoGen differs in detail from HumanML3D. The original SnapMoGen paper reports FID, R-Precision (Top-1/2/3), MultiModal Distance, Multimodality, and additionally CLIP Score on SnapMoGen, adopting TMR as the evaluation model because it evaluates both alignment and reconstruction/fidelity (Guo et al., 12 Jul 2025). FlowCoMotion’s SnapMoGen table reports R-precision (Top-1, Top-2, Top-3), FID, CLIP Score, and MModality (Guan et al., 13 Apr 2026). CMDM likewise evaluates SnapMoGen text-to-motion with R-Precision Top1/2/3, FID, MModality, and CLIP-score, and further adds a dedicated long-horizon evaluation on selected continuous sequences (Yu et al., 26 Feb 2026).

This suggests that SnapMoGen is used not merely as a larger dataset, but as a benchmark emphasizing semantic alignment under richer language supervision and, in some later work, temporal continuity and long-horizon consistency. That interpretation is directly supported by the repeated use of CLIP-style or retrieval-style alignment metrics alongside fidelity metrics on SnapMoGen (Guo et al., 12 Jul 2025).

4. MoMask++ and the original benchmark baseline

The model introduced together with SnapMoGen is MoMask++, an improved masked-token text-to-motion generator derived from prior generative masked modeling, especially MoMask. The central architectural change is to use multi-scale residual vector quantization with a shared single codebook across all scales, enabling all motion tokens to be generated by a single generative masked transformer rather than separate generators for first-layer and residual-layer tokens (Guo et al., 12 Jul 2025).

In the paper’s formulation, a motion encoder $\mathcal{E}$ maps motion $\mathbf{m}\in\mathbb{R}^{N\times D}$ to latent features $f\in\mathbb{R}^{n\times d}$, which are quantized by $\mathcal{Q}$ into token indices $q$. MoMask++ reformulates residual quantization across temporal scales $h^0 < \cdots < h^V = n$ as

$$q^v = \mathcal{Q}(\mathcal{I}(r^v,h^v)), \qquad r^{v+1} = r^v - \mathcal{I}(\hat f^v,h^V), \qquad r^0 = f,$$

with final reconstruction

$i$0

Because every scale uses the same vocabulary, all tokens can be concatenated and modeled jointly (Guo et al., 12 Jul 2025).

Text descriptions are encoded using T5-base. The paper studies both in-context conditioning and cross-attention conditioning. Training uses a masked-token objective

$i$1

with masking rate $i$2 for $i$3, plus text-conditioning dropout with probability 10\% for classifier-free guidance (Guo et al., 12 Jul 2025).

On SnapMoGen, the reported results are stronger than on HumanML3D. The original paper reports the following SnapMoGen numbers: MoMask achieves Top1 0.777, Top2 0.888, Top3 0.927, FID 17.404, CLIP 0.664, MModality 8.183; MoMask++ in-context reaches Top1 0.805, Top2 0.904, Top3 0.938, FID 15.56, CLIP 0.684, MModality 6.556; MoMask++ cross-attention reaches Top1 0.802, Top2 0.905, Top3 0.938, FID 15.06, CLIP 0.685, MModality 7.259 (Guo et al., 12 Jul 2025).

The paper emphasizes that MoMask++ improves both motion quality and text alignment on expressive prompts while using only two VQ layers, amounting to about a quarter of MoMask’s token count. It also notes an important benchmark-specific effect: in-context works better on HumanML3D, whereas cross-attention works slightly better on SnapMoGen because the in-context design tends to overfit on long prompts in SnapMoGen (Guo et al., 12 Jul 2025).

5. SnapMoGen as a benchmark in subsequent model development

Subsequent motion-generation papers use SnapMoGen as a principal testbed for new generative paradigms. In these works, the benchmark serves as a probe of semantic richness, motion realism, continuity, or causal generation, depending on the method under study.

Paper	Role of SnapMoGen	Representative claim
ScaleMoGen (Hwang et al., 12 May 2026)	Evaluation benchmark	A more challenging text-to-motion benchmark with long, expressive descriptions
Unified Conditional Flow (Li et al., 15 Apr 2026)	Main text-to-motion benchmark and training corpus backbone	Standard benchmark for text-to-motion; single canonical skeleton for fair comparison
FlowCoMotion (Guan et al., 13 Apr 2026)	One of two main benchmarks	High-quality motion dataset with 6 detailed text descriptions per motion segment
CMDM (Yu et al., 26 Feb 2026)	Key benchmark for causal, long-horizon evaluation	Allows evaluation of smooth, consistent motion generation

ScaleMoGen proposes scale-wise autoregressive next-scale prediction rather than standard next-token prediction, using a multi-scale skeletal-temporal representation and bitwise quantization. On SnapMoGen, it reports a CLIP Score of 0.693 (vs. 0.685 for MoMask++), and the paper presents SnapMoGen as one of its two main evaluation benchmarks (Hwang et al., 12 May 2026).

The unified conditional flow paper uses the SnapMoGen test split for standard text-to-motion evaluation while keeping the single canonical SnapMoGen skeleton fixed for fair comparison. Because SnapMoGen alone has only one canonical skeleton, the authors curate a multi-character Mixamo subset, convert it to SnapMoGen’s topology, and merge it with SnapMoGen to support zero-shot intra-structural retargeting. In this setting, SnapMoGen functions both as the canonical benchmark domain and as part of the joint training corpus (Li et al., 15 Apr 2026).

FlowCoMotion uses SnapMoGen to evaluate a token-latent coupling design under a text-conditioned flow-matching objective. On the SnapMoGen test set, its stronger Orbit variant achieves Top-1 0.776, Top-2 0.890, Top-3 0.934, FID 14.678, CLIP Score 0.672, and MModality 8.553. The same paper reports a SnapMoGen ablation in which replacing a latent only, dim 256 representation with a Hybrid token-latent, dim 64/192 representation improves T2M Generation FID from 19.465 ± .037 to 14.678 ± .061 and R-Precision@3 from 0.895 ± .002 to 0.934 ± .001 (Guan et al., 13 Apr 2026).

CMDM uses SnapMoGen to validate a causal-generation thesis. It describes SnapMoGen as containing 20,450 motion capture clips, 122K expressive captions, average caption length of 48 words, about 43.7 hours of data, and temporally continuous, long-horizon activities. On the main SnapMoGen text-to-motion benchmark, CMDM w/ FSS reports Top1 0.831, Top2 0.926, Top3 0.958, FID 14.451, MModality 9.521, and CLIP-score 0.702, exceeding the MoMask++ numbers reported in that paper’s comparison table (Yu et al., 26 Feb 2026).

6. Long-text prompting, continuity, and research significance

SnapMoGen was introduced to change the problem setting of text-to-motion generation from short action labels to expressive prompt following, finer controllability, and more ambitious temporal reasoning. The dataset paper argues that prior datasets and methods were optimized for short, coarse prompts, which made them weak on prompts requiring multiple sequential actions, detailed timing, body-part coordination, style or character cues, unusual action combinations, or long descriptive text (Guo et al., 12 Jul 2025).

One practical contribution associated with the benchmark is LLM rewriting of casual user prompts into the expressive narration style of SnapMoGen. The original paper describes an inference-time rewriting mechanism that preserves the user’s intent, keeps the prompt focused on human body movement, produces around 60 words with a maximum of 100, adds descriptive details such as posture, timing, and style, and suggests a motion duration typically 4 to 12 seconds. This is distinct from the training-time use of LLMs for annotation augmentation (Guo et al., 12 Jul 2025).

The preserved temporal continuity of SnapMoGen also makes it relevant to research areas beyond clip-level generation. The dataset paper explicitly states that continuity from long recordings facilitates research in long-term motion generation and blending (Guo et al., 12 Jul 2025). CMDM operationalizes this by selecting 128 samples with over five continuous motions for long-horizon evaluation, measuring both subsequence quality and transition smoothness (Yu et al., 26 Feb 2026). A plausible implication is that SnapMoGen bridges standard text-to-motion benchmarking and sequence-level problems such as transition generation and temporal compositionality more naturally than datasets assembled from isolated clips.

An additional significance of SnapMoGen is that it exposes failure modes that may be obscured on older benchmarks. CMDM notes that on SnapMoGen, lower jerk metrics can be misleading because a model may generate motions that are static or frozen, appearing smooth numerically while failing semantically (Yu et al., 26 Feb 2026). This suggests that the benchmark is especially useful for evaluating models that claim continuity, causality, or streaming capability, since its long expressive sequences make trivial failure modes easier to detect.

7. Limitations, caveats, and current interpretation

The dataset paper explicitly discusses limitations on both the data and model sides. Although SnapMoGen is mocap-based and quality-controlled, it uses inertial mocap suits, so some issues remain: global positions may be imprecise, and jitter can still occur in fast or complex motions. The dataset also cannot capture some motion categories well, including cartwheels, backflips, and outdoor activities like climbing (Guo et al., 12 Jul 2025).

Model-side caveats are similarly clear in the benchmark literature. MoMask++ still suffers from quantization error, struggles with rare motion patterns or uncommon prompts, and does not fully maintain physical plausibility, especially foot contacts (Guo et al., 12 Jul 2025). ScaleMoGen, FlowCoMotion, and CMDM all improve parts of the benchmark frontier, but none eliminates the reported gap to real motion data. The original paper states that despite improved results, there remains a large gap to real motions (Guo et al., 12 Jul 2025).

Within the current literature, SnapMoGen is best understood not simply as a larger replacement for HumanML3D, but as a benchmark designed to stress expressive text grounding, temporal continuity, and increasingly long-horizon or causal generation. Later papers use it for exactly those purposes: ScaleMoGen for coarse-to-fine autoregression (Hwang et al., 12 May 2026), unified conditional flow for joint generation, editing, and retargeting (Li et al., 15 Apr 2026), FlowCoMotion for hybrid token-latent flow modeling (Guan et al., 13 Apr 2026), and CMDM for causal diffusion-based streaming synthesis (Yu et al., 26 Feb 2026). Taken together, these uses establish SnapMoGen as a central benchmark for research on text-conditioned human motion under richer language supervision and more realistic temporal structure.