Paper deep dive
A Unified Conditional Flow for Motion Generation, Editing, and Intra-Structural Retargeting
Junlin Li, Xinhao Song, Siqi Wang, Haibin Huang, Yili Zhao
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 92%
Last extracted: 4/18/2026, 1:27:33 AM
Summary
The paper introduces a unified conditional flow-based generative framework for motion generation, editing, and intra-structural retargeting. By treating these tasks as conditional transport problems, the authors use a single rectified-flow motion model conditioned on text prompts and target skeletal structures. The architecture features a DiT-style transformer with per-joint tokenization and explicit joint self-attention to enforce kinematic dependencies, enabling zero-shot editing and retargeting without task-specific fine-tuning.
Entities (5)
Relation Signals (3)
Rectified-flow motion model → supports → Intra-structural retargeting
confidence 95% · a single trained model supports text-to-motion generation, zero-shot editing, and zero-shot intra-structural retargeting.
FlowEdit → enables → Zero-shot motion editing
confidence 90% · FlowEdit demonstrates inversion-free text-based editing using pretrained flow models
DiT-style transformer → implements → Rectified-flow motion model
confidence 90% · Our architecture extends a DiT-style transformer... We implement this vision via a rectified-flow motion model
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Text-driven motion editing and intra-structural retargeting, where source and target share topology but may differ in bone lengths, are traditionally handled by fragmented pipelines with incompatible inputs and representations: editing relies on specialized generative steering, while retargeting is deferred to geometric post-processing. We present a unifying perspective where both tasks are cast as instances of conditional transport within a single generative framework. By leveraging recent advances in flow matching, we demonstrate that editing and retargeting are fundamentally the same generative task, distinguished only by which conditioning signal, semantic or structural, is modulated during inference. We implement this vision via a rectified-flow motion model jointly conditioned on text prompts and target skeletal structures. Our architecture extends a DiT-style transformer with per-joint tokenization and explicit joint self-attention to strictly enforce kinematic dependencies, while a multi-condition classifier-free guidance strategy balances text adherence with skeletal conformity. Experiments on SnapMoGen and a multi-character Mixamo subset show that a single trained model supports text-to-motion generation, zero-shot editing, and zero-shot intra-structural retargeting. This unified approach simplifies deployment and improves structural consistency compared to task-specific baselines.
Tags
Links
- Source: https://arxiv.org/abs/2604.13427v1
- Canonical: https://arxiv.org/abs/2604.13427v1
Trouble viewing inline? Open PDF directly →
Full Text
53,563 characters extracted from source content.
Expand or collapse full text
A Unified Conditional Flow for Motion Generation, Editing, and Intra-Structural Retargeting Junlin Li ByteDanceSan JoseUSA 1306043330a@gmail.com , Xinhao Song ByteDanceSan JoseUSA xinhao.song13@gmail.com , Siqi Wang ByteDanceSan JoseUSA siqi9494@gmail.com , Haibin Huang ByteDanceSan JoseUSA jackiehuanghaibin@gmail.com and Yili Zhao ByteDanceSan JoseUSA yilizhao.cs@gmail.com Abstract. Text-driven motion editing and intra-structural retargeting, where source and target share topology but may differ in bone lengths, are traditionally handled by fragmented pipelines with incompatible inputs and representations: editing relies on specialized generative steering, while retargeting is deferred to geometric post-processing. We present a unifying perspective where both tasks are cast as instances of conditional transport within a single generative framework. By leveraging recent advances in flow matching, we demonstrate that editing and retargeting are fundamentally the same generative task, distinguished only by which conditioning signal, semantic or structural, is modulated during inference. We implement this vision via a rectified-flow motion model jointly conditioned on text prompts and target skeletal structures. Our architecture extends a DiT-style transformer with per-joint tokenization and explicit joint self-attention to strictly enforce kinematic dependencies, while a multi-condition classifier-free guidance strategy balances text adherence with skeletal conformity. Experiments on SnapMoGen and a multi-character Mixamo subset show that a single trained model supports text-to-motion generation, zero-shot editing, and zero-shot intra-structural retargeting. This unified approach simplifies deployment and improves structural consistency compared to task-specific baselines. motion editing, motion retargeting, text conditioning, T-pose conditioning, flow matching †ccs: Computing methodologies Animation†ccs: Computing methodologies Machine learning Figure 1. One rectified-flow model unifies motion generation, editing, and intra-structural retargeting. Conditioned on text and skeleton, it enables (left) generation, (middle) zero-shot editing by changing only the text condition, and (right) zero-shot retargeting by changing only the skeleton condition. 1. INTRODUCTION Character animation relies heavily on the ability to reuse and modify existing motion data. Traditionally, cross-character motion reuse is handled by motion retargeting; here we focus on intra-structural retargeting, which adapts motion between characters that share the same skeletal topology but differ in proportions. Recently, text-to-motion generative models enable prompt-driven motion editing, allowing users to specify high-level semantic changes in natural language (Tevet et al., 2023; Chen et al., 2023; Zhang et al., 2022; Guo et al., 2024, 2025). However, these two paradigms currently exist in isolation: editing is typically handled by generative models using scarce edit-specific supervision (Athanasiou et al., 2024; Li et al., 2025) or heuristic steering (Hong et al., 2025), while retargeting remains a distinct post-process relying on inverse kinematics or learned mapping networks (Gleicher, 1998; Aberman et al., 2020; Zhang et al., 2023b). This fragmentation stems from the fact that editing and retargeting are usually formulated with incompatible inputs, objectives, and representations. First, the inputs differ: editing conditions on a source motion and a text prompt, whereas retargeting conditions on a source motion and a target skeleton. Second, the objectives diverge: text-driven editing seeks to preserve the source motion’s structure while altering its semantics, whereas retargeting seeks to preserve the semantic intent while adapting the motion to a new skeletal morphology. Third, and most critically, the underlying representations are mismatched. Most text-to-motion generators operate in a canonical output space (e.g., a fixed kinematic tree with standardized bone lengths) to simplify training (Tevet et al., 2023; Chen et al., 2023). Consequently, structural adaptation is deferred to downstream tools rather than handled by the generative model itself. This separation complicates deployment—requiring multiple models and inconsistent representations—and hinders the composition of operations, such as simultaneous editing and retargeting, within a single interactive system. We propose a unifying perspective where text-based editing and skeletal retargeting are fundamentally the same generative task, distinguished only by the conditioning signal being modulated. In our framework, both operations are cast as instances of conditional transport: editing corresponds to modifying the semantic condition (text) while fixing the structure, whereas retargeting corresponds to modifying the structural condition (skeleton) while fixing the semantics. Recent flow-based editing methods facilitate this unification by demonstrating that ordinary differential equations (ODEs) can map between source- and target-conditioned distributions without explicit inversion (Kulikov et al., 2025). This suggests a simple yet powerful principle: if a motion generator is conditioned on both language and skeletal structure, changing either condition at inference time yields the desired manipulation “for free,” purely through the dynamics of the generative flow. Guided by this principle, we introduce a single rectified-flow motion model jointly conditioned on a text prompt and a target character T-pose (serving as a proxy for bone lengths and rest pose). Our model is trained on motion clips paired with text and skeleton information. During inference, it is manipulated via a unified flow-based update rule: (i) Text-based editing changes the prompt while holding the target skeleton fixed; (i) Intra-structural Retargeting swaps the skeleton condition while leaving the text fixed. To ensure effective skeletal conditioning, we adopt a DiT-style transformer backbone (Peebles and Xie, 2022) operating on per-joint tokens and introduce a dedicated joint attention mechanism to explicitly model intra-frame kinematic dependencies. We also employ a multi-condition variant of classifier-free guidance to strictly control text adherence and skeletal conformity during both sampling and editing. We evaluate our system on SnapMoGen (Guo et al., 2025) and a multi-character Mixamo subset. Across experiments, the same trained model and inference pipeline support text-to-motion generation, zero-shot text edits, and zero-shot intra-structural retargeting. This eliminates the need for task-specific toolchains. Beyond quantitative improvements in fidelity and consistency, qualitative results demonstrate that our dual-conditioning strategy better preserves motion structure during edits and produces target-conforming outputs even under significant morphological changes. To summarize, the main contributions of this study include: • We formulate text-driven motion editing and intra-structural motion retargeting under a unified conditional transport framework, treating them as equivalent generative tasks distinguished only by the modulated condition. • We propose a dual-conditioning architecture—integrating text and target skeleton vectors—featuring a joint-token transformer with explicit joint attention to handle both semantic edits and cross-character transfer. • We provide a comprehensive evaluation spanning generation, editing, and retargeting, demonstrating that a single model can serve as a versatile motion manipulation engine without task-specific fine-tuning. 2. Related Work 2.1. Text-to-Human Motion Generation Human motion generation from text has seen rapid progress recently, fueled by gradual scale-growing datasets (Guo et al., 2022, 2025; Fan et al., 2025) and advances in generative modeling. The progress is largely driven by new auto-regressive methods and diffusion formulations. Token-based approaches first compress motion into discrete codes, typically via VQ-VAE style quantization (van den Oord et al., 2018), and then model the token sequence autoregressively, e.g. T2M-GPT (Zhang et al., 2023a). MoMask further adopts masked modeling with hierarchical quantization for high-fidelity generation (Guo et al., 2024), and Guo et al. (2025) improves this method by using a multi-scale RVQ. In parallel, diffusion-based generators have become a dominant framework, mitigating discretization and compression information loss in VQ-based methods. Early efforts include MDM (Tevet et al., 2023) and MotionDiffuse (Zhang et al., 2022). Subsequent work advances this direction through iterative generation in StableMoFusion (Huang et al., 2024a) and latent-diffusion formulations such as MARDM (Meng et al., 2024). Building on these diffusion formulations, SALAD introduces skeleton-aware latent diffusion to strengthen the generation and editing (Hong et al., 2025). Skeleton priors during generation. Despite architectural differences, most text-to-motion generators are trained and sampled on a fixed canonical kinematic tree and do not explicitly incorporate a skeleton prior during the generative process. Instead, structural cues are typically encoded implicitly through the motion parameterization or through the learned encoder and token codebooks in VQ-based pipelines (van den Oord et al., 2018; Guo et al., 2024). Consequently, adapting generated motions to different skeletons is usually deferred to a separate retargeting stage, see Sec. 2.3. Only a recent attempt (Cao et al., 2025) has been made by using skeleton-motion data and human specified key joint correspondence to guide the generation and retarget based on Gat et al. (2025). 2.2. Motion Editing Motion editing aims to modify an existing motion while preserving non-edited content. Text-based motion editing has recently been studied under supervised and zero-shot settings. Supervised approaches rely on triplets source motion, edit text, target motion; MotionFix introduces a semi-automatically collected benchmark of such triplets and trains a conditional diffusion model for text-driven edits (Athanasiou et al., 2024). Building on this formulation, SimMotionEdit augments diffusion-based editing with an auxiliary motion similarity prediction objective (Li et al., 2025). Zero-shot motion editing often uses mask-based inpainting to localize edits within a sequence (Tevet et al., 2023). Related work also explores editing by steering a pretrained generator via attention or conditioning control (Chen et al., 2025; Hong et al., 2025), or by applying structured latent or code edits as in CoMo (Huang et al., 2024b). Connections to image and video editing. Beyond motion-specific methods, our approach draws inspirations on text-driven image and video editing. Diffusion-based editing commonly realizes edits via attention control or inversion, as in Prompt-to-Prompt (Hertz et al., 2022) and FateZero (Qi et al., 2023). FlowEdit demonstrates inversion-free text-based editing using pretrained flow models, offering an alternative to inversion-based diffusion editing and directly motivating our design choices (Kulikov et al., 2025). Edit prompts and data scarcity. Many text-based motion editing setups require an edit instruction, which is qualitatively different from standard text-to-motion captions and substantially less available at scale; hence, edit-text datasets are typically much smaller than text–motion pair counterparts and often require semi-automatic construction (Athanasiou et al., 2024). 2.3. Motion Retargeting Motion retargeting transfers motion from a source character to a target character with different morphology, such as bone lengths, skeleton structures, and mesh shape, while preserving motion semantics and physical plausibility. In recent years, learning-based approaches have become a dominant direction. Many methods formulate retargeting as unpaired translation between skeleton domains, combining cycle consistency with adversarial losses (Villegas et al., 2018; Aberman et al., 2020; Lim et al., 2019). SAME adopts supervised training to learn a skeleton-agnostic motion latent and uses a single model to handle arbitrary skeleton retargeting (Lee et al., 2023). Beyond skeleton-only formulations, mesh-aware retargeting highlights that avoiding interpenetrations and maintaining plausible surface interactions requires geometry-aware reasoning. Recent works therefore incorporate mesh-derived distance cues to enforce body-part relationship constraints: R2ET uses voxel-based distance-field losses (Zhang et al., 2023b), MeshRet relies on dense surface sensors (Ye et al., 2024), and SMRNet enforces relationships via surface edges sampling and convex-hull constraints (Zhang et al., 2025). Although the broader retargeting literature spans cross-topology and mesh-aware transfer, our work focuses on the intra-structural setting. Even within this narrower regime, motion retargeting has rarely been studied with diffusion- or flow-based generative models; most existing methods instead rely on adversarial training or directly paired mappings. 3. METHOD We propose treating motion retargeting as a conditional generative problem rather than a geometric optimization task. Under this view, the target skeleton (defined by bone lengths and rest pose) serves as a conditioning signal analogous to a “style” or “domain” code, while the motion semantics serve as the “content.” This parallels the logic of text-driven editing: just as changing a text prompt modifies the semantic trajectory of a motion, changing a skeleton condition modifies its morphological trajectory. By leveraging the FlowEdit framework (Kulikov et al., 2025) on top of a unified generative backbone, we enable both text-driven semantic edits and structure-driven retargeting within a single model, without requiring task-specific modules or training. To achieve this vision, our approach relies on three integrated components: a dual-conditioned transformer backbone, a conditional rectified flow training strategy, and a unified FlowEdit inference scheme, as summarized in Figure 2 and Figure 3. In the subsequent sections, we first formalize the problem definition and data representation, followed by detailed descriptions of the model architecture and the unified inference procedure. Model architecture diagram for the dual-conditioned transformer with text and skeleton conditioning. Figure 2. Model Architecture. Input frame tokens are reshaped into per-joint tokens for processing. Time and skeleton conditions are injected via AdaLN, while text embeddings are integrated through cross-attention. 3.1. Problem Formulation Problem Definition We represent a human motion clip as a sequence of feature vectors ∈ℝT×Dx ^T× D, where T is the sequence length and D is the feature dimension. Each motion x is associated with two conditioning signals: (i) a semantic condition txtc_txt, derived from a natural language prompt, and (i) a structural condition skelc_skel, representing the character’s skeletal morphology (e.g., bone lengths and rest pose). Our objective is to learn a single conditional distribution p(∣txt,skel)p(x _txt,c_skel) that unifies generation and manipulation under a shared conditional transport framework. Specifically, the model must support three operations: (1) Text-to-Motion Generation: Sampling ∼p(∣txt,skel)x p(x _txt,c_skel) for any target skeleton skelc_skel within the supported topology. (2) Zero-Shot Text-Based Editing: Given a source motion srcx_src and its conditions, generating a modified motion tgtx_tgt that adheres to a new prompt txttgtc_txt^tgt while preserving the non-edited part of the txtsrcc_txt^src and the original motion. (3) Zero-Shot Intra-Structural Retargeting: Given a source motion srcx_src generated with skeleton skelsrcc_skel^src, producing a new motion tgtx_tgt that preserves the source semantics but conforms strictly to a new target skeleton skeltgtc_skel^tgt (where the skeletons share a graph topology but differ in bone lengths). Data Representation. To simultaneously satisfy the requirements of standard evaluation metrics and precise geometric control, we operate in a concatenated feature space: (1) t=[tgen;tret],tgen∈ℝDgen,tret∈ℝDret.x_t= [x^gen_t\,;\,x^ret_t ], ^gen_t ^D_gen,\;x^ret_t ^D_ret. Generative Features (genx^gen). The first block follows the standard SnapMoGen layout (Dgen=8+12JD_gen=8+12J), consisting of root velocities (yaw/planar) and heights, local joint rotations, positions and velocities relative to the root, and foot contacts. While genx^gen is essential for compatibility with existing text-to-motion metrics (e.g., Fréchet Inception Distance (FID)), it is ill-suited for retargeting. It uses root angular velocities instead of absolute angles, and temporally smooths rotations, discarding the absolute spatial frame and accumulating position error, observed as “root drift” when retargeting with genx^gen alone. Retargeting Features (retx^ret). To address the geometric limitations of genx^gen, we append AnyTop-style (Gat et al., 2025) features (Dret=12JD_ret=12J) for each joint j∈1,…,Jj∈\1,…,J\: (2) t,jret=[t,j;t,j6D;t,j]∈ℝ12.x^ret_t,j= [p_t,j\,;\,r^6D_t,j\,;\,v_t,j ] ^12. Here, t,jp_t,j and t,jv_t,j denote canonicalized root-relative joint positions and velocities. Unlike using the root angular velocities to reconstruct the root transformations in genx^gen, our root rotation is directly obtained from t,j6Dr^6D_t,j. To accurately predict t,jp_t,j conditioned on a skeleton vector skelc_skel, the model must learn the underlying kinematic constraints (i.e., bone lengths). This implicit structure modeling is the foundation of our retargeting capability: it allows the FlowEdit inference process to continuously ”warp” the skeleton geometry from source to target. Condition Inputs. We parameterize the semantic condition txtc_txt using text tokens =wii=1L P=\w_i\_i=1^L from a frozen T5 encoder. The structural condition skelc_skel is parameterized as a vector ∈ℝDret S ^D_ret following the layout of Eq. (2). In practice, we set the position channels of S to the target character’s T-pose offsets (defining bone lengths and rest pose) and set the remaining rotation/velocity channels to identity/zero. 3.2. Model Architecture We adopt a DiT-style transformer backbone that is jointly conditioned on text and skeleton. Let ∈ℝT×DX ^T× D denote an input state such as τx_τ, and let J be the joint count. Unlike prior motion transformers that treat each frame feature as a single token, we encourage body-part-level generation by operating on per-joint tokens. Concretely, instead of projecting tX_t to an arbitrary hidden vector, we project it to J joint chunks and reshape: (3) t=reshape(int+in)∈ℝJ×d,d=DhJ,H_t=reshape (W_inX_t+b_in ) ^J× d, d= D_hJ, where DhD_h is the model total hidden size and t,jH_t,j is the token for joint j at frame t. The input projection learns how to split the whole feature dimensions into joint parts. Joint attention. Standard transformer backbones capture temporal structure via attention between frame tokens, while spatial inter-joint structure is typically handled indirectly by the per-token feed-forward network that mixes all joint parts inside the hidden frame vector. Without a strong skeletal inductive bias, this mixing must learn inter-joint relations through a generic MLP, which is parameter-heavy and often slow to learn. We keep the existing temporal attention and feed-forward updates, and additionally introduce a dedicated joint self-attention module that attends over joints within each frame by forcing the hidden dimension to be split into J joint chunks. This module is conceptually similar to attention module in graph transformers, but implemented as dense attention over the J joint tokens at each frame. Within each layer, we apply joint self-attention independently at each time step: (4) t,⋅←t,⋅+gjnt()⋅Attnjnt(AdaLN(t,⋅,)),H_t,· _t,·+g_jnt(c)·Attn_jnt\! (AdaLN(H_t,·,c) ), where c is the conditioning vector defined below, AttnjntAttn_jnt attends over joints j=1…Jj=1… J using joint rotary embeddings, and gjnt()g_jnt(c) is the AdaLN-Zero residual gate. This directly models intra-frame joint dependencies, which is critical when generating on characters with different bone lengths. Frame attention and feed-forward. After updating joint tokens, we form frame tokens back by concatenating all joints per frame, t=[t,1;…;t,J]∈ℝDhF_t=[H_t,1;…;H_t,J] ^D_h, and run temporal attention over the motion sequence with frame rotary embeddings. Finally, a SwiGLU feed-forward network updates frame tokens and the outputs are reshaped back to joints. Skeleton and flow-time condition injection. We encode flow time τ with a sinusoidal embedding and map it to τ∈ℝDhc_τ ^D_h with an MLP. We encode the target T-pose vector with a skeleton MLP, yielding s∈ℝDhc_s ^D_h. The final conditioning vector is (5) =ϕ(τ+s)∈ℝDh,c=φ\! (c_τ+c_s ) ^D_h, where ϕφ is a small condition-merging MLP. We inject c into every transformer block using AdaLN-Zero modulation with learned shift, scale, and a residual gate. This conditioning controls both joint-level and frame-level updates; for example, it drives the gate gjnt()g_jnt(c) in Eq. (4). Text injection at joint and frame resolutions. We encode text prompt tokens wii=1L\w_i\_i=1^L with a frozen T5 encoder, producing token embeddings i∈ℝdtexte_i ^d_text. We then form: (i) frame-level text tokens frm∈ℝL×DhE^frm ^L× D_h and (i) joint-level text tokens jnt∈ℝL×dE^jnt ^L× d via two separate linear projections. After joint attention, joint tokens query the text via cross-attention using jntE^jnt as keys/values, while frmE^frm participates in the subsequent frame-text cross attention after the frame attention. This design injects language at both the spatial resolution over joints and the temporal resolution over frames. Output head. The final joint tokens are normalized and reshaped back to DhD_h, then projected to the motion feature dimension: (6) ^1=outreshape(Norm())+out. x_1=W_out\,reshape\! (Norm(H) )+b_out. We initialize out,outW_out,b_out to zero to stabilize early training. 3.3. Rectified Flow Matching We model motions with a rectified flow (RF) (Liu et al., 2022) defined on the continuous-time path τ∈ℝT×Dx_τ ^T× D, τ∈[0,1]τ∈[0,1]. Let 1x_1 be a data sample and 0∼(0,)x_0 (0, I) be base noise. To match FlowEdit’s assumption that the noisy state is a linear interpolant, we use rectified flow matching with the linear path (7) τ=(1−τ)0+τ1,x_τ=(1-τ)x_0+ _1, whose ground-truth velocity is (8) ⋆=dτdτ=1−0.u = dx_τdτ=x_1-x_0. Predicting the clean target. Instead of predicting velocity, we predict the clean target sample ^1=fθ(τ,τ;,) x_1=f_θ(x_τ,τ; P, S) following the previous work (Tevet et al., 2023; Cuba et al., 2025), which empirically yields smoother outputs. To support inference and editing, we then convert the target prediction to a velocity field via (9) θ(τ,τ;,)=^1−τ1−τ,v_θ(x_τ,τ; P, S)= x_1-x_τ1-τ, where we clamp τ away from 11 for numerical stability. Flow matching loss. Training minimizes a mean-squared error on the predicted target directly. Because τx_τ concatenates τgenx_τ^gen and τretx_τ^ret blocks, we enforce the objective per feature block: (10) ℒsimple= _simple= λgen1,0,τ[‖θgen(τ,τ;,)−1gen‖2] _genE_x_1,x_0,τ [ ^gen_θ(x_τ,τ; P, S)-x_1^gen ^2 ] +λret1,0,τ[‖θret(τ,τ;,)−1ret‖2], + _retE_x_1,x_0,τ [ ^ret_θ(x_τ,τ; P, S)-x_1^ret ^2 ], where θgenf^gen_θ and θretf^ret_θ are the corresponding slices of the concatenated prediction. This prevents the higher-dimensional block from implicitly dominating the loss signal. We set λgen=λret=1 _gen= _ret=1 unless otherwise stated. Multi-Condition Classifier-Free Guidance. To support flexible conditioning and strong text adherence, we use a multi-condition extension of classifier-free guidance (CFG). At sampling time, we evaluate the model under an unconditioned input and several conditioned variants: uncond (=∅,=)( P= , S=0), text (,=)( P, S=0), skeleton (=∅,)( P= , S), text+skeleton (,)( P, S). We fill the text tokens with padding tokens when =∅ P= , and set all skeleton vector channels to zero when = S=0. Let uv_u denote the unconditioned velocity, and let kv_k denote a conditioned velocity. We combine them as (11) CFG=u+∑kwk(k−u),v_CFG=v_u+ _kw_k (v_k-v_u ), where wkw_k are user-controlled guidance weights (e.g., wtextw_text, wskeletonw_skeleton, wtext+skeletonw_text+skeleton). During training we randomly drop conditions to learn the branches required by Eq. (11). Specifically, we drop both text and skeleton condition with probability pbothp_both, and additionally drop only the text condition with probability ptextp_text. This reflects our downstream use cases: we never need to generate motions with text-only conditioning, and without a skeleton condition the position channels can induce a random skeleton that does not match standard downstream formats (T-pose + joint rotations). Table 1. Text-to-motion generation quantitative evaluation on the SnapMoGen test split. Baseline numbers are taken from SnapMoGen (mean ± 95% CI). Bold and underlined numbers indicate the best and second-best results, respectively. R Precision↑ Methods Top 1 Top 2 Top 3 FID↓ CLIP Score↑ MModality↑ MDM 0.503± 0.002 0.653± 0.002 0.727± 0.002 57.783± 0.092 0.481± 0.001 13.412± 0.231 T2M-GPT 0.618± 0.002 0.773± 0.002 0.812± 0.002 32.629± 0.087 0.573± 0.001 9.172± 0.181 StableMoFusion 0.679± 0.002 0.823± 0.002 0.888± 0.002 27.801± 0.063 0.605± 0.001 9.064± 0.138 MARDM 0.659± 0.002 0.812± 0.002 0.860± 0.002 26.878± 0.131 0.602± 0.001 9.812± 0.287 MoMask 0.777± 0.002 0.888± 0.002 0.927± 0.002 17.404± 0.051 0.664± 0.001 8.183± 0.184 MoMask++ 0.802± 0.001 0.905± 0.002 0.938± 0.001 15.060± 0.065 0.685± 0.001 7.259± 0.180 Ours 0.917± 0.001 0.973± 0.001 0.987± 0.001 16.567± 0.045 0.663± 0.001 11.259± 0.293 Diagram illustrating text-based editing and retargeting under a shared, inversion-free inference scheme. Figure 3. Applications of our unified inference scheme. Left: text-based editing by changing the text condition. Right: intra-structural retargeting by changing only the skeleton condition. Both use the same pre-trained model and an inversion-free update rule, where the edit velocity is obtained by combining velocity predictions under different conditions. 3.4. FlowEdit for Zero-Shot Editing and Retargeting Unlike SDEdit and ODEdit, which edit by adding noise to the input and re-running the generative process, FlowEdit (Kulikov et al., 2025) constructs a direct transport ODE between the source- and target-conditioned distributions, without an explicit inversion or round-trip through pure noise. This shorter shared-noise path empirically yields lower transport cost and better structure preservation, while remaining optimization-free and model-agnostic. FlowEdit treats editing as solving for an updated motion y by integrating a difference-of-velocities field under two conditions along a shared noise path. The original FlowEdit formulation parameterizes time as t∈[0,1]t∈[0,1] with t=0t=0 representing data and t=1t=1 representing noise. In this paper we instead use τ∈[0,1]τ∈[0,1] with τ=0τ=0 representing noise and τ=1τ=1 representing data, consistent with our rectified-flow interpolation in Eq. (7). Let the source motion be x with source condition csc_s and a target condition ctc_t. We define a noisy source trajectory (12) ~τ=(1−τ)ϵ+τ,ϵ∼(0,). x_τ=(1-τ) _τ+ , _τ (0,I). Update rule. Starting from τmin=y_ _ =x, we integrate: (13) dτdτ=CFG(τ+~τ−,τ;ct)−CFG(~τ,τ;cs), dy_τdτ=v_CFG\! (y_τ+ x_τ-x,τ;c_t )-v_CFG\! ( x_τ,τ;c_s ), where τ∈[τmin,1]τ∈[ _ ,1]. The two velocity evaluations share the same ~τ x_τ and thus the same noise. This is shown in Fig. 3. Edit strength. We initialize at τmin≥0 _ ≥ 0 so the first target evaluation is anchored to the same partially noised state as the source: τmin+~τmin−=~τminy_ _ + x_ _ -x= x_ _ . Smaller τmin _ injects more noise and yields stronger edits, while larger τmin _ produces weaker edits that better preserve the source motion. We additionally use separate classifier-free guidance strengths for the source and target branches. We mirror this by allowing different guidance weights in Eq. (11) for the two velocity evaluations: wk(s)\w_k^(s)\ for CFG(⋅;cs)v_CFG(·;\,c_s) and wk(t)\w_k^(t)\ for CFG(⋅;ct)v_CFG(·;\,c_t). Intuitively, a smaller source guidance helps preserve the input motion, while a larger target guidance enforces the desired prompt or skeleton. Task instantiations. (i) Zero-shot text-based editing: keep the skeleton fixed and change only text, cs=(,)c_s=( P, S), ct=(′,)c_t=( P , S). (i) Zero-shot intra-structural retargeting: keep text empty and change only skeleton, cs=(∅,)c_s=( , S), ct=(∅,′)c_t=( , S ). Both tasks use the same trained model and the same update rule, without extra training or fine-tuning. 4. Experiments 4.1. Dataset SnapMoGen. We use the SnapMoGen dataset (Guo et al., 2025), which is designed for text-to-human-motion generation. It contains 20,450 motion clips that total 43.7 hours, with 122,565 text descriptions whose average length is 48 words. Each clip lasts 4 to 12 seconds at 30 FPS. SnapMoGen is provided in a canonical BVH skeleton representation. We treat it as a single-skeleton dataset and use the provided skeleton for all text-to-motion evaluations. Mixamo retargeting subset. Since SnapMoGen contains only a single canonical skeleton, we constructed a supplementary multi-character dataset from Mixamo (Adobe Systems Inc., 2018) to enable structural retargeting training. We curated a diverse subset of motions by filtering for common locomotion and action keywords (e.g., run, walk, sit) and subsampling 30% of the matches to minimize redundancy. To model cross-morphology transfer, we paired each selected motion with three distinct characters, creating varied source-skeleton combinations. All Mixamo characters are selected in the dataset, making it multiple times larger than the previous methods. These assets were standardized to the SnapMoGen topology by normalizing character heights, pruning incompatible joints (e.g., fingers), and reordering the kinematic chain to match the canonical depth-first traversal. To integrate this data with the text-driven SnapMoGen pipeline, we generated synthetic captions using Qwen3-VL (Bai et al., 2025). For each motion, we produced six caption variants using the original Mixamo filename as a prompt prior, ensuring the style and length distribution aligned with SnapMoGen’s ground-truth labels. Finally, we merged this Mixamo subset into the SnapMoGen dataset, maintaining an 80/20 train-test split across both data sources. 4.2. Implementation details Model Configuration. We employ a 12-layer transformer backbone with 12 attention heads and a feed-forward expansion ratio of 3. To align with our per-joint tokenization strategy, we set the total hidden dimension to Dh=432D_h=432. Given a standard character topology of J=24J=24 joints, this allocates d=18d=18 hidden channels per joint token, ensuring sufficient capacity for the joint self-attention mechanism. We apply a dropout rate of 0.1 throughout the network. For text conditioning, we utilize a frozen T5-Large encoder (Ni et al., 2022) to extract semantic embeddings. Training Protocol. We train the model with fixed-length motion windows of 320 frames. Optimization is performed using AdamW (Loshchilov and Hutter, 2019) with a learning rate of 5×10−55× 10^-5 and a weight decay of 5×10−35× 10^-3. We use a global batch size of 512 and train 500 epochs. The entire training process takes approximately 8 hours on a single node equipped with 8 NVIDIA H100 GPUs. 4.3. Text-to-motion generation We first evaluate the model’s core generative capability on the SnapMoGen test split. Note that for this standard benchmark, we use the single canonical SnapMoGen skeleton to ensure fair comparison with baselines. Metrics. Following previous works (Tevet et al., 2023; Guo et al., 2025), we report FID, R-Precision (Top 3), CLIP score, and Multimodality. All metrics are computed in the feature space of a pre-trained TMR model (Petrovich et al., 2023). We sample motions using the fourth-order Runge–Kutta (RK4) integrator with 100 steps (Hairer et al., 1993). For classifier-free guidance (Eq. 11), we empirically set the guidance scales to wtxt=0.5w_txt=0.5, wskel=0w_skel=0, and wboth=1.0w_both=1.0. This configuration prioritizes the joint text-skeleton condition while providing a moderate boost to semantic adherence via the text-only branch. Figure 5 visualizes samples generated from long, complex prompts. Table 1 compares our method against state-of-the-art baselines. Our model achieves the highest R-Precision scores across all k-values, indicating superior semantic alignment with complex prompts. On FID and Multimodality, we consistently rank second-best, performing on par with top-tier methods like MoMask++(Guo et al., 2025). Crucially, prior methods often trade alignment for diversity. In contrast, Our method maintains a robust balance: it delivers state-of-the-art alignment and competitive realism without collapsing diversity. Most importantly, these results are achieved using a unified model trained to support varying skeletons within a shared topology. This demonstrates that our dual-conditioning architecture incurs no performance penalty on the standard single-skeleton task. Table 2. Ablation study of the generation task on the SnapMoGen test split. Methods FID↓ CLIP↑ M↑ Ours w/o jnt attn 17.205± 0.061 0.664± 0.001 10.941± 0.321 Ours w/o jnt attn, larger DhD_h 17.085± 0.035 0.668± 0.001 10.376± 0.345 Ours 16.567± 0.045 0.663± 0.001 11.259± 0.293 Ablation Study. We investigate the impact of our architectural choices in Table 2. Removing the joint self-attention causes a sharp degradation in FID, indicating its vital role in modeling realistic kinematic chains. Moreover, despite adding only ∼ 0.4M parameters, the joint-attention model outperforms a wider baseline (hidden dim 576; 90M parameters) that lacks this mechanism. This confirms that the performance gain stems from the inductive bias of explicit spatial modeling rather than raw capacity. 4.4. Motion Editing We evaluate model’s zero-shot editing capabilities, focusing on semantic modification while maintaining the structural integrity of the source motion. Experimental Setup. We evaluate global and local editing, including motion addition, deletion, and replacement on 20 motions randomly sampled from the test split. We compare against MDM (Tevet et al., 2023), MoMask++ (Guo et al., 2025), and an ablation integrating MDM’s editing protocol into our framework. Inference Configuration. We use 100 integration steps and control edit strength via starting time τmin≈0.1 _ ≈ 0.1. The source text guidance is set to 1.51.5, while target text guidance is increased to ≈3.5≈ 3.5 to drive the edit. Skeleton guidance is fixed at 1.01.0 for both branches. Figure 4. Perceptual user study. Pairwise preferences (Ours/Equal/Baseline) for Source Preservation (Pres.), Edit Accuracy (Edit), and Overall Quality (Over.). Ours is preferred across all three criteria vs. MDM (Tevet et al., 2023) (65/78/77%), MoMask++ (Guo et al., 2025) (75/75/72%), and the mask-based ablation (46/70/63%) (Pres./Edit/Over.). Perceptual Evaluation. Figure 7 provides a visual comparison of editing results. Given the subjective nature of motion editing, we conducted an A/B user study (N=76N=76 valid participants). Participants evaluated 20 motion pairs across 6 survey sets using 3 criteria: Source Preservation, measuring consistency of unedited parts; Edit Accuracy, measuring alignment with the target text; and Overall Quality, measuring realism and smoothness. As summarized in Figure 4, users preferred our model in 70% of comparisons on average. These results demonstrate a robust preference for our flow-based transport over mask-based alternatives. Unlike MDM, our method is fully automatic and does not require manual masking, offering a significant advantage in usability. 4.5. Motion Retargeting We further evaluate the model’s ability to transfer motion between characters with identical topological graphs but significantly different bone lengths and proportions. Inference Configuration. While structural retargeting is robust across a range of parameters, we found that maintaining skeletal guidance scales at wskel=1.0w_skel=1.0 for both source and target branches yields the most consistent structural fidelity. We use a 100-step integrator and sweep the start step between 5 and 40 (stride 5) to select the optimal trade-off between source motion preservation and target skeletal adaptation for each pair. Baselines & Metrics. We compare against three retargeting methods: SAN (Aberman et al., 2020) (without the adversarial loss), SAME (Lee et al., 2023), and R2ET (Zhang et al., 2023b) (using only its skeleton retargeting module). We also report a copy baseline, which naively applies the source joint rotations and root translations to the target skeleton. Quantitative performance is measured using Mean Squared Error (MSE) on global positions, normalized by character height to account for scale differences. Quantitative Results. Table 3 presents the comparative results. Our method achieves the lowest MSE across all methods, significantly outperforming both the copy baseline and dedicated retargeting networks. We report errors for two decoding strategies: direct prediction, which is reconstructing from the position channels (t,jp_t,j in Eq. 2), and FK reconstruction, recomputing positions from predicted rotations applied to the target skeleton. The low error in the direct prediction variant confirms that the model successfully learns the correlation between the source and target skeletons. It is not merely “hallucinating” rotations; it is generating a physically consistent pose sequence that respects the target’s bone constraints. Furthermore, the ablation study (Table 3, bottom rows) underscores the critical role of our joint token architecture. Figure 6 further visualizes retargeting outcomes across different methods. Table 3. Retargeting evaluation on the Mixamo test split. Method Error (×103× 10^3)↓ Copy 15.89 SAN (Aberman et al., 2020) 17.52 SAME (Lee et al., 2023) 13.47 R2ET (Zhang et al., 2023b) 8.13 Ours w/o jnt attn, direct prediction 5.73 Ours, direct prediction 5.24 Ours, FK reconstruction 4.91 5. Conclusion and Limitations We introduced a unified motion generation model jointly conditioned on text and skeleton, enabling text-based motion editing and intra-structural retargeting with the same inference procedure. This shared framework simplifies production workflows and enables new interactive editing tools. Our current implementation assumes a consistent joint ordering and requires a well-defined rest pose for each target character. Extreme morphology changes (e.g., non-humanoid characters) and highly out-of-distribution motions (e.g., climbing) remain challenging. Besides, very long text prompts can introduce small velocity differences during text-based editing, resulting in subtle edits. Figure 5. Qualitative comparison on text-to-motion generation. For visualization, motions with little or no root translation are manually time-shifted. Prompt words in red denote actions, while words in green indicate motion modifiers. Our model successfully synthesizes coherent full-body motions that faithfully reflect fine-grained textual details over long time spans. This temporal coherence is critical for our downstream editing tasks, where the model must preserve the narrative structure of the source motion. Figure 6. Qualitative retargeting comparison. We compare against SAN (Aberman et al., 2020), SAME (Lee et al., 2023), and R2ET (Zhang et al., 2023b). The proposed method better preserves fine-grained local motion and adapts to varying skeleton proportions. As highlighted in the zoomed-in regions, baseline methods often introduce artifacts such as unnatural twisting or over-stretching when adapting to different skeletons. In contrast, our method preserves delicate local details (e.g., leg bending scale, hand facing direction) while naturally adapting the global motion to fit the new body shape. Figure 7. Qualitative results on text-based motion editing. Edited prompt words are highlighted in red and green. For visualization, motions shifted as in Fig. 5. Our method demonstrates superior adaptability across all edit types. Unlike mask-based approaches (MDM) which can introduce stiffness at mask boundaries, or regenerative baselines (MoMask++) which often drift from the source semantics, Our method produces accurate, coherent edits. References K. Aberman, P. Li, D. Lischinski, O. Sorkine-Hornung, D. Cohen-Or, and B. Chen (2020) Skeleton-aware networks for deep motion retargeting. ACM Trans. Graph. 39 (4). External Links: ISSN 0730-0301, Link, Document Cited by: §1, §2.3, §4.5, Table 3, Figure 6. Adobe Systems Inc. (2018) Mixamo. Note: Accessed: 2017-09-28; 2018-12-27 External Links: Link Cited by: §4.1. N. Athanasiou, A. Ceske, M. Diomataris, M. J. Black, and G. Varol (2024) MotionFix: text-driven 3d human motion editing. In SIGGRAPH Asia 2024 Conference Papers, Cited by: §1, §2.2, §2.2. S. Bai, Y. Cai, R. Chen, K. Chen, X. Chen, Z. Cheng, L. Deng, W. Ding, C. Gao, C. Ge, W. Ge, Z. Guo, Q. Huang, J. Huang, F. Huang, B. Hui, S. Jiang, Z. Li, M. Li, M. Li, K. Li, Z. Lin, J. Lin, X. Liu, J. Liu, C. Liu, Y. Liu, D. Liu, S. Liu, D. Lu, R. Luo, C. Lv, R. Men, L. Meng, X. Ren, X. Ren, S. Song, Y. Sun, J. Tang, J. Tu, J. Wan, P. Wang, P. Wang, Q. Wang, Y. Wang, T. Xie, Y. Xu, H. Xu, J. Xu, Z. Yang, M. Yang, J. Yang, A. Yang, B. Yu, F. Zhang, H. Zhang, X. Zhang, B. Zheng, H. Zhong, J. Zhou, F. Zhou, J. Zhou, Y. Zhu, and K. Zhu (2025) Qwen3-vl technical report. arXiv preprint arXiv:2511.21631. Cited by: §4.1. Z. Cao, B. Liu, S. Li, W. Zhang, and H. Chen (2025) G-dream: graph-conditioned diffusion retargeting across multiple embodiments. External Links: 2505.20857, Link Cited by: §2.1. L. Chen, S. Lu, W. Dai, Z. Dou, X. Ju, J. Wang, T. Komura, and L. Zhang (2025) Pay attention and move better: harnessing attention for interactive motion generation and training-free editing. External Links: 2410.18977, Link Cited by: §2.2. 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, p. 18000–18010. Cited by: §1, §1. M. C. Cuba, V. do Carmo Melício, and J. P. Gois (2025) FlowMotion: target-predictive conditional flow matching for jitter-reduced text-driven human motion generation. External Links: 2504.01338, Link Cited by: §3.3. K. Fan, S. Lu, M. Dai, R. Yu, L. Xiao, Z. Dou, J. Dong, L. Ma, and J. Wang (2025) Go to zero: towards zero-shot motion generation with million-scale data. External Links: 2507.07095, Link Cited by: §2.1. I. Gat, S. Raab, G. Tevet, Y. Reshef, A. H. Bermano, and D. Cohen-Or (2025) AnyTop: character animation diffusion with any topology. In Proceedings of the Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers, SIGGRAPH Conference Papers ’25, New York, NY, USA. External Links: ISBN 9798400715402, Link, Document Cited by: §2.1, §3.1. M. Gleicher (1998) Retargetting motion to new characters. In Proceedings of the 25th Annual Conference on Computer Graphics and Interactive Techniques, SIGGRAPH ’98, New York, NY, USA, p. 33–42. External Links: ISBN 0897919998, Link, Document Cited by: §1. C. Guo, I. Hwang, J. Wang, and B. Zhou (2025) SnapMoGen: human motion generation from expressive texts. External Links: 2507.09122, Link Cited by: §1, §1, §2.1, Figure 4, §4.1, §4.3, §4.3, §4.4. C. Guo, Y. Mu, M. G. Javed, S. Wang, and L. Cheng (2024) MoMask: generative masked modeling of 3d human motions. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Vol. , p. 1900–1910. External Links: Document Cited by: §1, §2.1, §2.1. 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), p. 5152–5161. Cited by: §2.1. E. Hairer, S. P. Nørsett, and G. Wanner (1993) Solving ordinary differential equations i: nonstiff problems. 2 edition, Springer Series in Computational Mathematics, Vol. 8, Springer. Cited by: §4.3. A. Hertz, R. Mokady, J. Tenenbaum, K. Aberman, Y. Pritch, and D. Cohen-Or (2022) Prompt-to-prompt image editing with cross attention control. Cited by: §2.2. S. Hong, C. Kim, S. Yoon, J. Nam, S. Cha, and J. Noh (2025) SALAD: skeleton-aware latent diffusion for text-driven motion generation and editing. External Links: 2503.13836, Link Cited by: §1, §2.1, §2.2. Y. Huang, Y. Hui, C. Luo, Y. Wang, S. Xu, Z. Zhang, M. Zhang, and J. Peng (2024a) StableMoFusion: towards robust and efficient diffusion-based motion generation framework. arXiv preprint arXiv: 2405.05691. Cited by: §2.1. Y. Huang, W. Wan, Y. Yang, C. Callison-Burch, M. Yatskar, and L. Liu (2024b) Como: controllable motion generation through language guided pose code editing. In European Conference on Computer Vision, p. 180–196. Cited by: §2.2. V. Kulikov, M. Kleiner, I. Huberman-Spiegelglas, and T. Michaeli (2025) Flowedit: inversion-free text-based editing using pre-trained flow models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, p. 19721–19730. Cited by: §1, §2.2, §3.4, §3. S. Lee, T. Kang, J. Park, J. Lee, and J. Won (2023) SAME: skeleton-agnostic motion embedding for character animation. In SIGGRAPH Asia 2023 Conference Papers, SA ’23, New York, NY, USA. External Links: ISBN 9798400703157, Link, Document Cited by: §2.3, §4.5, Table 3, Figure 6. Z. Li, K. Cheng, A. Ghosh, U. Bhattacharya, L. Gui, and A. Bera (2025) SimMotionEdit: text-based human motion editing with motion similarity prediction. External Links: 2503.18211 Cited by: §1, §2.2. J. Lim, H. J. Chang, and J. Y. Choi (2019) PMnet: learning of disentangled pose and movement for unsupervised motion retargeting. In British Machine Vision Conference (BMVC), Cited by: §2.3. X. Liu, C. Gong, and Q. Liu (2022) Flow straight and fast: learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003. Cited by: §3.3. I. Loshchilov and F. Hutter (2019) Decoupled weight decay regularization. External Links: 1711.05101, Link Cited by: §4.2. Z. Meng, Y. Xie, X. Peng, Z. Han, and H. Jiang (2024) Rethinking diffusion for text-driven human motion generation. arXiv preprint arXiv:2411.16575. Cited by: §2.1. J. Ni, G. H. Abrego, N. Constant, J. Ma, K. B. Hall, D. Cer, and Y. Yang (Eds.) (2022) Sentence-t5: scaling up sentence encoder from pre-trained text-to-text transfer transformer. External Links: Link Cited by: §4.2. W. Peebles and S. Xie (2022) Scalable diffusion models with transformers. arXiv preprint arXiv:2212.09748. Cited by: §1. M. Petrovich, M. Black, and G. Varol (2023) TMR: text-to-motion retrieval using contrastive 3d human motion synthesis. External Links: Document Cited by: §4.3. C. Qi, X. Cun, Y. Zhang, C. Lei, X. Wang, Y. Shan, and Q. Chen (2023) FateZero: fusing attentions for zero-shot text-based video editing. arXiv:2303.09535. Cited by: §2.2. G. Tevet, S. Raab, B. Gordon, Y. Shafir, D. Cohen-or, and A. H. Bermano (2023) Human motion diffusion model. In The Eleventh International Conference on Learning Representations, External Links: Link Cited by: §1, §1, §2.1, §2.2, §3.3, Figure 4, §4.3, §4.4. A. van den Oord, O. Vinyals, and K. Kavukcuoglu (2018) Neural discrete representation learning. External Links: 1711.00937, Link Cited by: §2.1, §2.1. R. Villegas, J. Yang, D. Ceylan, and H. Lee (2018) Neural kinematic networks for unsupervised motion retargetting. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Cited by: §2.3. Z. Ye, J. Liu, J. Jia, S. Sun, and M. Z. Shou (2024) Skinned motion retargeting with dense geometric interaction perception. Advances in Neural Information Processing Systems. Cited by: §2.3. J. Zhang, M. Wang, F. Zhang, and F. Zhang (2025) Skinned motion retargeting with preservation of body part relationships. IEEE Transactions on Visualization and Computer Graphics 31 (9), p. 4923–4936. External Links: Document Cited by: §2.3. J. Zhang, Y. Zhang, X. Cun, S. Huang, Y. Zhang, H. Zhao, H. Lu, and X. Shen (2023a) T2M-gpt: generating human motion from textual descriptions with discrete representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Cited by: §2.1. J. Zhang, J. Weng, D. Kang, F. Zhao, S. Huang, X. Zhe, L. Bao, Y. Shan, J. Wang, and Z. Tu (2023b) Skinned motion retargeting with residual perception of motion semantics & geometry. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 13864–13872. Cited by: §1, §2.3, §4.5, Table 3, Figure 6. 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. External Links: 2208.15001, Link Cited by: §1, §2.1.