Multimodal dance style transfer

In recent years, style transfer has achieved impressive progress across various fields, including computer vision, speech and music processing, natural language processing, motion animation, etc.

In the field of computer vision, the pioneering work of Gatys et al. [1] introduces the concept of style transfer and leverages the hierarchical layers in convolutional neural networks (CNNs) to extract both the underlying content structures and stylistic elements. They utilize an optimization-based technique to transfer styles between arbitrary images. Later, Li et al. [15] propose whitening and coloring transform (WCTs) to stylize images by analyzing second-order correlations of content and style features. More generally, Huang et al. [16] introduce an adaptive instance normalization (AdaIN) layer to solve the challenge of arbitrary target style application in image style transfer, which has been broadly adopted to fuse the style and content information for image generation and image-to-image translation [17‐19]. Zhu et al. [6] introduce CycleGAN which uses a pair of generators and discriminators to learn the mapping between two unpaired image domains. This general idea has been further developed in StarGAN [13], which incorporates domain labels as additional input and enables image style transfer among multiple corresponding domains, such as facial appearances and expressions.

Voice conversion (VC) refers to a technique of converting non-linguistic or para-linguistic information from the original speech into the desired target speech while retaining the linguistic content unchanged. While some early VC frameworks have achieved success [20, 21], they rely on precisely aligned parallel data of source and target speech. To address this challenge, researchers have turned to non-parallel VC techniques. For example, Hsu et al. [22] construct a VC system from non-parallel speech with variational autoencoders and Wasserstein GANs. Kameoka et al. [23] build an auxiliary classifier VAE with information-theoretic regularization for the model training. Kaneko and Kameoka [24] propose CycleGAN-VC, which is a variation of the CycleGAN architecture using gated CNNs and an identity-mapping loss. This was later improved by CycleGAN-VC2 [10] with the addition of a 2-1-2D convolution structure and two-step adversarial losses to improve performance. This approach has also been extended to a StarGAN-based architecture to enable many-to-many mappings across different domains [14, 25]. Fu et al. [11] incorporated transformers and curriculum learning in voice conversion to facilitate training efficiency.

With the development of MIDI format parsing, research has also been carried out to transfer symbolic music styles, as demonstrated by studies such as Groove2Groove [26], which employs an encoder-decoder architecture and parallel data, and Malik et al. [27] that introduce StyleNet with a shared GenreNet, which aimed to learn various styles for music translation. Brunner et al. [2] use a CycleGAN-based approach for MIDI music. Ding et al. [28] design SteelyGAN, a symbolic-domain transfer approach that combines both pixel-level and latent-level features. Regarding style transfer in natural language processing (NLP), Mueller et al. [29] propose recurrent variational auto-encoders (VAE) to modify text sequences. Fu et al. [30] develop a multi-decoder and style-embedding model using adversarial networks to learn content and style representations. Dai et al. [31] propose a Style Transformer network with a tailored training scheme that integrates an attention mechanism and makes a latent representation-agnostic assumption. Finally, Xu et al. [32] introduce a cycled reinforcement learning approach focusing on unpaired sentiment-to-sentiment translation.

Our research focuses on transferring motion data, specifically dance movements. We use the CycleGAN-VC2 backbone, originally designed for voice conversion, as our foundation. To improve scalability, we extend the model to a StarGAN-based framework. We also augment an additional music modality in our approach to improve the training performance.

Motion style transfer has been a longstanding challenge in the field of computer animation, which involves transferring the motion style of a source animation to a target animation while preserving the key content, such as its structure, timing, spatial relationships, etc. Prior research in motion style transfer relied on handcrafted features [33‐38]. Since style is a challenging attribute to define precisely, most modern studies advocate data-driven approaches for feature extractions [4, 39‐45]. Commonly used models for style transfer include K nearest neighbors (KNNs) [46], convolutional auto-encoders [39], temporal invariant AdaIN layers [5], CycleGAN [45], spatial-temporal graph neural networks [44], and autoregressive flows [47]. Furthermore, certain studies focus on the issue of efficient real-time style transfer [41, 43, 46]. However, it should be noted that all these studies focus on relatively simple human movements, such as exercise and locomotion, where the stylistic variation is often limited. For example, the transfer between children and adult locomotion [45]. In contrast, our work deals with the transfer of dance movements that possess a significant level of complexity in terms of postures, transitions, rhythms, and artistic styles. Consequently, our research may have more empirical and practical value for video games or film industries. Given the intricacies involved, our method differs significantly from the reviewed research. We utilize transformer and curriculum learning on top of CycleGAN-VC2 to enable more effective training on more complex motion data.

Another important task that accompanies the transfer of motion styles is to evaluate the quality of the synthesized animation. While subjective surveys help estimate movement quality, such as recruiting a group of dance experts with defined requirements, relying on them for evaluation can be expensive, time-consuming, and have low reproducibility [38]. Utilizing objective metrics for quantitative evaluation eliminates the need for human involvement, avoiding the issues associated with subjective surveys. The Fréchet Inception Distance (FID) metric [48], which has proven effective in assessing synthesized images in computer vision, has become a standard in evaluating image generative models. Building on the success of FID, Wang et al. [49] extended the FID concept to motion data. Yoon et al. [50] defined the Fréchet Gesture Distance (FGD) as a metric to evaluate speech-driven gesture generation based on the distance between gesture feature distributions. Maiorca et al. [38] transform motions into image representations and introduced the Fréchet Motion Distance (FMD) to assess the quality and diversity of synthesized motion. Valle-Pérez et al. [8] evaluated the realism of music-based dance generation by measuring the Fréchet distance between the distributions of poses and movements. For the motion style transfer task, we propose a Fréchet Pose Distance (FPD) based on the distribution of key poses to assess the content preservation, as well as a Fréchet Motion Distance (FMD), the Fréchet distance between the distribution of the true dance motion and the generated dance motion to evaluate transfer strength.

Numerous studies have focused on human motion synthesis, utilizing various techniques such as deep feedforward networks [51], convolutional networks [52], recurrent models [53], graph neural networks [54], and autoencoders [55]. Dance and music are often intertwined, leading to an emerging research topic known as cross-modal motion generation. This field aims to understand the association between different modalities better and improve music-conditioned motion synthesis. Early works in cross-modal motion generation mostly focused on statistical models [56‐58]. Specifically, these models typically generate motions by selecting pre-existing dance moves that match particular music features, such as rhythm, intensity, and structure. With the recent advances in deep learning and the availability of large-scale datasets, learning-based methods have been developed to learn the patterns between music and motion. For example, ChoreoMaster [9] propose an embedding module to capture music-dance connections, while in DeepDance [59], a cross-modal association system is designed to correlate dance motion with music. Lee et al. [60] propose a decomposition-to-composition framework that leverages MM-GAN for music-based dance unit organization. The DanceNet model, as proposed in [60], uses a musical context-aware encoder to fuse music and motion features. In DanceFormer [61], kinematics-enhanced transformer-guided networks are utilized to perform motion curve regression. In a recent work by Valle-Pérez et al. [8], cross-modal transformers were successfully employed to model the relationship between music and motion distributions.

Music-conditioned dance synthesis refers to the task of generating dance motion sequences that are synchronized with a given musical context. In contrast, our work focuses on the dance style transfer task, which involves manipulating the style of existing dance movements while preserving contextual information. Although our style transfer model does not require music as an input for conditioning, incorporating it can enhance the quality of the generated movements.

Our study aims to develop mapping functions between two distinct domains, denoted as X and Y, without relying on paired data between these domains. In our scenario, we focus on the transfer of dance movements between two different styles, corresponding to domains X and Y given dance sample \(x\sim P_X\) and \(y\sim P_Y\), where x is a sample from domain X, drawn from the probability distribution \(P_X\), and y is a sample from domain Y, drawn from the probability distribution \(P_Y\). The dance samples may also be accompanied by music, with \(m_x \in M_x\) and \(m_y \in M_y\), respectively, with associated styles. The inclusion of music modality is optional in the style transfer task.

To tackle the problem described above, we employ a strategy inspired by CycleGAN-like framework [6], which is depicted in Fig. 3. The CycleDance architecture we use incorporates two discriminators, \(D_X\) and \(D_Y\), which differentiate between real and generated data. In addition, the architecture includes two mappings, \(G_{X\rightarrow Y}\) and \(G_{Y\rightarrow X}\), which are responsible for generating patterns in the target style. The mappings are cycled to enable the generated patterns to be converted back to their original domains. To achieve this goal, we adopt a strategy similar to that used in CycleGAN-VC2 [10] and utilize four types of losses, also see Fig. 3.

Adversarial loss \({\mathcal {L}}_{{\text {adv}}}^{X\rightarrow Y}\): this loss measures the difference between the transferred data \(G_{X\rightarrow Y}(x, m_x)\) and the target y, where the discriminator \(D_Y\) distinguishes between the transferred data and the real data:Similarly, we define the adversarial loss \({\mathcal {L}}_{{\text {adv}}}^{Y\rightarrow X}\) for \(G_{Y\rightarrow X}\) and discriminator \(D_X\).

Cycle-consistency loss \({\mathcal {L}}_{{\text {cyc}}}\): this loss helps to account for the loss of contextual information by recovering the original x and y from generated patterns, by \(G_{X\rightarrow Y}(x, m_x)\) and \(G_{Y\rightarrow X}(y, m_y)\), where 1-norm is adopted to minimize the absolute difference:Identity-mapping loss \({\mathcal {L}}_{{\text {id}}}\): this loss further promotes content preservation by enforcing an identity transformation when applying \(G_{X\rightarrow Y}\) and \(G_{Y\rightarrow X}\) to the other domain:Two-step adversarial loss \({\mathcal {L}}_{{\text {adv2}}}\): a second adversarial loss is used to mitigate the over-smoothing reconstruction statistics in the cycle-consistency loss [10]:Note that this introduces an additional discriminator, denoted as \(D_{X}^{'}\). The loss \({\mathcal {L}}_{{\text {adv2}}}^{Y\rightarrow X\rightarrow Y}\) can be defined in a similar manner.

The final objective is expressed as a weighted sum of the above-mentioned loss termswhere \(\lambda _{{\text {cyc}}}\) and \(\lambda _{{\text {id}}}\) trade off the consistency and identity loss terms.

In addition, we utilize a curriculum learning algorithm as a training scheme. The rationale for this is that training can be made more efficient by initially handling simpler data and then gradually increasing the complexity of the training task. Similar strategies have been applied in various applications and scenarios, demonstrating its ability to improve the convergence rate, generalization capacity, and training stability [62]. We implement a curriculum learning strategy based on the length of the data by learning from short samples to longer ones as the training process progresses. The training data are truncated by following the method in [11].

While CycleDance allows for generation of dance movement when a sufficient number of training examples are available, one limitation is that it only learns one-to-one mappings. New models need to be trained from scratch to transfer a new pair of dance styles. We thus extend CycleDance to a StarGAN-based backbone for a more scalable solution called StarDance. The StarDance architecture has one generator G and one discriminator D to accommodate for all styles with the help of an additional domain classifier C. The generator can take a dance sample x of style attribute \(c_x\), with a target style attribute \(c_y\), accompanied by music m, and generate a target dance sample \({\hat{y}}=G(x, m, c_y)\). The discriminator D produces a probability \(D(x, c_x)\) that can be used to distinguish between real and transferred data, while the additional domain classifier C in StarDance is designed to predict which style an input dance sequence belongs to. Here we represent c as a one-hot vector, where each element is associated with a dance style. Similar to CycleDance, we define four types of losses in StarDance, as well as a domain classification loss for the classifier.

Adversarial loss \({\mathcal {L}}_{{\text {adv}}}\):Cycle-consistency loss \({\mathcal {L}}_{{\text {cyc}}}\):Identity-mapping loss \({\mathcal {L}}_{{\text {id}}}\):Two-step adversarial loss \({\mathcal {L}}_{{\text {adv2}}}\):Domain classification loss \({\mathcal {L}}_{{\text {cls}}}\):To summarize, the full objective of StarDance to be minimized with respect to G, D, and C is given bywhere \(\lambda _{{\text {cyc}}}, \lambda _{{\text {id}}}\), and \(\lambda _{{\text {cls}}}\) are regularization parameters to weight the losses. The training strategy of StarDance is illustrated in Fig. 4. With the StarDance architecture, there are fewer coefficients to tune, given a more compact generator.

Our Cycle/StarDance framework builds upon a cross-modal transformer, as illustrated in Fig. 2. In the CycleDance framework, the cross-modal transformer is employed to concatenate motion and music encodings, both of which are obtained through a sequence of layers including 2D convolution (purple blocks in Fig. 2), 2D–1D reshaping (red), residual convolution (green), and modality-specific transformers (yellow). The 2D convolutional layers are utilized for downsampling while keeping the sequential structure of the input data. The resulting downsampled features are then reshaped and passed through the residual blocks of 1D CNNs. The reshaped 1D sequences are further processed by transformers, which use a position embedding to output encodings that capture temporal relationships among timesteps. Finally, the generator takes the concatenated encodings and feeds them into a 1D–2D reshape block (red) and an upsampling block (purple) to synthesize transferred dance motions. Within these blocks, we employ gated linear units (GLUs) [63], a tunable activation function, to learn hierarchical and sequential structures. In our StarDance framework, the transformer is adapted to incorporate style labels. Before feeding the concatenated encodings into the 1D–2D reshape block, we concatenate the encoded features and style attribute along the channel dimension.

For the discriminator, CycleDance first downsamples the motion data with a 2D CNN. We adopt the same approach as proposed in [10] and use convolution only at the last layer of the discriminator to alleviate training instability. The output layer of the discriminator employs a sigmoid activation function to make the final prediction on the motion clip. Similarly, the StarDance discriminator first concatenates the motion data and the style attribute along the channel dimension before the downsampling block.

We also devise a domain classifier in StarDance, which downsamples a motion data input with 2D CNNs and produces a sequence of class probability distributions at the last layer that predicts how likely the motion sequence is to belong to the respective style attributes.

We implement and train the Cycle/StarDance model with the PyTorch-Lightning framework, with Adam optimizers used for training. The batch size is fixed to 1, given that \(batch\_size=1\) with instance normalization achieves invariance to mean and variance of features, which is beneficial for the style transfer task. We set the initial learning rate of 0.0002 for the generator and 0.0001 for the discriminator and domain classifier, with momentum terms of 0.5 for both optimizers. As for loss weights, we set \(\lambda _{{\text {cyc}}}=10\), \(\lambda _{{\text {id}}}=5\), and \(\lambda _{{\text {cls}}}=5\). The identifying-mapping loss \({\mathcal {L}}_{{\text {id}}}\) is only active for the first \(10^4\) iterations for regulating consistency in the initial training stage. Our framework adopts a curriculum learning strategy to alleviate the challenges of training models for extremely long sequential data. Specifically, we gradually increase the sample length from 32 frames to 800 frames, adding 8 frames to data fragments every 500 epochs. In total, we train the entire framework for \(5\times 10^5\) iterations.

We utilize the AIST+ Dance Database [42] as our source of data to generate 3D dance motion samples with paired music. The AIST++ database reconstructs 3D motions with SMPL parameters from multi-view videos in the AIST Dance Database [64], a large-scale dance video database containing various dance genres recorded by professional dancers. To extract motion features, we first downsample all motion data to 30 frames per second (fps) and retarget it to a 21-body-joint skeleton using Autodesk MotionBuilder. We adopt exponential map parametrization to represent the 3D rotation of all joints. The root joint (hip) is characterized by four additional features representing changes in the vertical root position, ground-projected position, and 2D facing angle. Consequently, a 67-dimensional vector is employed to represent the motion features of each frame. The music features are extracted with the Librosa toolbox in a similar way to [64]. We combine 20-dim MFCC, 12-dim chroma, 1-dim one-hot peaks, and 1-dim one-hot beats, resulting in a total 35-dim audio feature. Six dance genres were selected for analysis based on suggestions from professional dancers we recruited, specifically ballet-jazz, locking, hip-hop, pop, waacking, and house dance. Each dance set consists of 141 motion sequences and six songs, spanning approximately 2000 s in total. We make the preprocessed data publicly available at https://​urlis.​net/​aist to further facilitate research in the community.

To evaluate the impact of design choices such as the cross-modal transformer and curriculum learning strategy, we compare our proposed CycleDance and StarDance models to the CycleGAN-VC2 baseline. Additionally, we implement three alternative architectures for an ablation study. In the first ablation configuration, CycleTransGAN, we eliminate the music pathway, cross-modal transformer, and curriculum learning strategy. Our aim is to highlight the effectiveness of the transformer architecture we introduced. In the CycleTransGAN+CL ablation, we apply curriculum learning to the CycleTransGAN model. We aim to gauge the performance gains achieved by meticulously modulating the complexity of the samples used to train the model. The final ablation, CycleCrossTransGAN, employs cross-modal transformers for motion and music information in the encoder without using the curriculum learning strategy. Through this ablation, we aim to evaluate the influence of the cross-modal transformers by comparing the discrepancies between CycleTransGAN and CycleCrossTransGAN.

The primary objective of all these models is to transfer the dance style from a specified source to a particular target dance style. We conduct assessments from both objective and subjective perspectives to ensure a comprehensive evaluation of the complex motion patterns common in dance. For the objective evaluation, we analyze 17 dance sequences for each style. Style transfer was performed on each ablated model, and two metrics were used to evaluate the performance on transfer strength and content preservation. The two metrics are designed based on Fréchet distance, which is computed by Eq. 12:where \((\mu _{r},\Sigma _{r})\) and \((\mu _{g},\Sigma _{g})\) are, respectively, the mean and the covariance matrix of the real and generated dance movement distribution.

Transfer strength The transfer strength, which measures the extent that the source style is converted to the target style, is the most crucial aspect of style transfer. Specifically, we adopt the Fréchet distance between the true and generated dance motion distributions. To capture features that are more style-correlated such as intensity, we utilize joint velocity and acceleration. To be specific, we use pairs of consecutive raw poses without normalization \((x_{i-1}, x_{i})\) to convert the pose representations to joint velocity \(v_i\). Similarly, we use triples of consecutive poses \((x_{i-1}, x_{i}, x_{i+1})\) to estimate the joint acceleration \(a_i\). We refer to this metric as the Fréchet motion distance (FMD), which is used to quantify the transfer strength between the generated motion and target style.

Content preservation We use the Fréchet pose distance (FPD) as a measure of content preservation, which evaluates how well the salient poses of the original dance movement are preserved after the transfer. For this dimension, we use the Fréchet distance between distributions of key poses \(x_k\) for a given dance movement. We detect frames presenting local maxima in joint acceleration as the key frames that segment the full dance movement. To ensure comparability across frames, key poses in these frames are normalized with respect to the hip-centric origin.

In Table 1, we present the quantitative results of the proposed model and ablations for style transfer across three pairs of dance genres in both directions. These pairs include ‘ballet-jazz to locking dance’ (BJ2LC) and ‘locking dance to ballet-jazz’ (LC2BJ), ‘waacking to hip-hop dance’ (WK2HP) and ‘hip-hop to waacking dance’ (HP2WK), as well as ‘pop to house dance’ (PO2HO) and ‘house to pop dance’ (HO2PO).

The baseline model, CycleGAN-VC2, appears to struggle with the style transfer task for locking dance to ballet-jazz, as evidenced by the significantly higher MFD compared to all other ablation methods. The complete framework, CycleDance, outperforms all other ablation methods and achieves the best performance on both metrics and almost all transfer pairs. This emphasizes the necessity of all introduced techniques in this task.

Figure 5 presents an example of a synthesized motion clip that demonstrates the transfer of dance style from locking to ballet-jazz. An example of style transfer from ballet-jazz to locking dance is shown in Fig. 6. The top keyframe sequence shows the original locking dance. The middle sequence is generated by the baseline model, CycleGAN-VC2, while the bottom sequence is generated by the proposed CycleDance. By comparing the poses of each column in these two figures, it can be observed that the extracted key gestures are representative of the pose sequences. Notably, CycleDance achieves a higher similarity to the source gestures, preserving more content while aligning better with the target dance style.

In addition, the ablation study revealed that CycleTransGAN (the combination of CycleGAN-VC2 and transformer) achieved lower FMD scores, suggesting that the model benefited from capturing richer intra-relations among frames with the help of the transformer. By comparing the results of ‘CycleTransGAN and CycleCrossTransGAN’, we can observe an improvement in both FMD and FPD metrics. This suggests that the music information aids in generating accurate target-style movements and that this contextual information is effectively encoded by the cross-modal transformer. The comparison between ‘CycleTransGAN and CycleTransGAN+CL’ as well as ‘CycleCrossTransGAN and CycleDance’ indicates that curriculum learning greatly enhances transfer strength. This demonstrates the effectiveness of gradually increasing the level of difficulty by training with longer dance sequences.

In Table 1, we also present results from StarDance, which demonstrates an ability to solve the task of transferring among multiple dance styles using a more scalable model. The performance of StarDance exhibits a similar trend to that of CycleDance. It performs reasonably well on certain pairs, such as transferring from ’house to pop dance,’ but shows poor performance when transferring from ’ballet-jazz to locking dance’. The poorer performance in general is natural, given that CycleDance is specifically trained for particular pairs of dance styles, whereas StarDance is expected to cover them all. It is worthwhile to investigate the factors that influence style transfer in different dance pairs in future studies.

To obtain a more comprehensive evaluation of our model and the baseline, we conducted a user study in addition to the objective assessment, asking participants to rate three aspects: motion naturalness, transfer strength, and content preservation. This study also includes open-ended questions to gather additional feedback and suggestions for future work. We make the user study questionnaire publicly available at https://​urlis.​net/​ques to further support relevant studies.

Our subjective analysis primarily focuses on the style transfer between ballet-jazz and locking dance, as these dance styles are well-established and well-understood by dance professionals, providing a diverse representation of the challenges in style transfer. An online survey was performed to evaluate the transfer tasks for both ‘ballet-jazz to locking dance’ and ‘locking dance to ballet-jazz’, gathering participant feedback. To generate the videos for each source and target dance sequence, we utilized the Blender software to create 8-second clips with an x-bot character (representing ballet-jazz) and a y-bot character (representing locking dance). During the survey, participants were presented with a source dance video clip followed by a generated target dance clip. Prior to this, an introduction phase allowed participants to become familiar with the animated dance through video clips that could be played. To avoid potential order effects, the order of target dance clips was randomly selected and balanced. Each target dance clip was generated either from CycleDance or from the baseline. The participants were allowed to view the clips multiple times before answering three questions:During the study, 30 participants equipped with at least 5 years of dance experience, including training, choreographing, teaching, or performing, were recruited. Participants in the study were aged between 20 and 41 years (median 30), with 37.9% identifying as male, 58.6% as female, and 3.4% as other. According to the demographic questions, the participants reported their familiarity with ballet-jazz dance and locking dance as M = 3.93 (SD = 1.05) and M = 3.03 (SD = 1.18), respectively, on a scale from 1 (not at all familiar) to 5 (very familiar). As the generated motions were presented using virtual characters, we also assessed the frequency at which participants played video games. Results indicated that 34.5% of participants played video games weekly, 13.8% played monthly, 13.8% yearly, and 37.9% rarely.

We performed a statistical analysis of the subjective responses from the user study to support our findings, and evaluated whether our proposed method could be further enhanced. Based on Fig. 7, the results indicate that CycleDance received higher ratings from experts than the baseline model on both motion naturalness and transfer strength. The Wilcoxon signed-rank test was used to assess the statistical significance of the subjective responses. The results showed that both the median value of motion naturalness \((Z=-9.2262, p<0.0001)\) and transfer strength \((Z=-8.7677, p<0.0001)\) were significantly higher for CycleDance compared to the baseline model. Therefore, from the perspective of dance experts, CycleDance is preferred for its improved naturalness and similarity to the target dance style, which is consistent with what we observe from the objective quantitative results (Sect. 4.3). For the responses on content preservation, Fig. 8 presents the overall statistics for the four queried aspects. The experts chose CycleDance more often than the baseline model on all four aspects, indicating that they believed that CycleDance better preserved specific features of the source dance style. We conducted a McNemar test to assess the statistical significance of the differences between the baseline model and CycleDance. The test showed no significant differences between the two models on ‘Orientation through space’ \((p=0.1724)\) and ‘Shapes of the limbs’ \((p=0.1573)\). On the other hand, there was a strong statistical significance supporting the proposed CycleDance for ‘Shapes of the body trunk’ \((p=0.000002)\) and ’Rhythmic patterns’ \((p=0.00004)\), based on the median value. Both CycleDance and the baseline model received higher scores on rhythmic patterns and orientation through space among the four aspects evaluated. This suggests preserving dance orientation, and rhythm is relatively easier when performing dance style transfer. Preserving the shape of the limbs is found to be more challenging, where no significant differences were found between CycleDance and the baseline. However, CycleDance outperforms the baseline in preserving the shape of the body trunk.

In response to open-ended questions, dance experts provided feedback that, for the task from ‘ballet jazz to lock dance,’ both methods produced a jerky style that mimicked pop and lock dance. The example illustrated in Fig. 9 is frequently mentioned as a significant instance of successful "transfer" with a noticeable locking dance style from the perspective of experienced dancers. Regarding the CycleDance samples of transferring ‘locking dance to ballet-jazz’, the dance experts noted that the character arms clearly exhibit jazz or ballet characteristics and effectively maintain traditional shapes. The dance experts also provided feedback on some limitations of the proposed method. One common observation is that some generated motions appear wobbly, indicating a need for applying smoothing filters to improve the overall quality of the results. Experts have also pointed out that in jazz ballet, dancers typically point their toes, whereas the generated movements always show ankle flexion. This demonstrates the limitation that the available data do not adequately represent the nuances of foot movements required in ballet-jazz, highlighting the need for more comprehensive data collection in future studies. This limitation also leads to some physically unrealistic effects, such as the character appearing to float when their body does not have any contact with the floor.