Answer: We present MotionGPT to address various human motion-related tasks within one single unified model, by unifying motion modeling with language through a shared vocabulary. To train this unified model, we propose an instructional training scheme under the protocols for multiple motion-language, which further reveals the potential of Large Language Models (LLMs) in motion tasks beyond the success of language generation. However, it is non-trivial for this combination since it needs to model and generate two distinct modes from scratch. Contrary to the previous work leveraging CLIP to extract text embedding as motion generation conditions, like T2M-GPT, MotionGPT introduces the motion-language pre-training on LLM so it can leverage the strong language generation and zero-shot transfer abilities of pre-trained language models, as well as generates human language and motion in a unified model.

Answer: We propose instruction tuning to train a single MotionGPT across all motion-related tasks, while task-specific tuning is to train and evaluate MotionGPTs on a single task. We employ these two training schemes to study the ability of MotionGPT across multi-tasks. As shown in this figure, we provide zero-shot cases. Benefitting from strong language models, MotionGPTs can understand unseen works in the text-to-motion training set, like "scuttling" and "barriers", and generate correct motions based on the meaning of sentences. However, it still struggles to generate unseen motions, like gymnastics, even if MotionGPTs understand the text inputs.

Answer: We have faced this limited dataset issue while implementing MotionGPT and in our further research. It is a hard but valuable work to unify and collect a larger motion dataset. Fortunately, some researchers are working on this problem, as seen in recent work like Motion-X and other datasets, which hold promise for advancing large-scale motion models. We intend to further evaluate MotionGPT on these larger datasets once they become available.

Answer: Unlike the previous motion generators using the text encoder of CLIP for conditions, please note that MotionGPTs leverage language models to learn the motion-language relationship, instead of relying on text features from CLIP. According to our zero-shot results (cf. Fig. 12) and performances on multi-tasks (cf. Fig. 10), MotionGPTs establish robust connections between simple/complex texts and simple motions in evaluations, but they fall short when it comes to complex-text to complex motion translation.

Answer: The first language model that we used to build MotionGPTs is LLaMA-13B. However, it shows insufficient performance and low training efficiency. We assume the reason is the limited dataset size compared to the large parameters and language data of LLaMA. We tried a smaller size decoder-only backbone GPT2-Medium and provide the results in Tab. 15. Then, we thus chose T5-770M, a small but common language model, as our final backbone, because many previous vision-language multimodal works, like Unified-IO and BLIP, have chosen T5, this encoder-decoder architecture. It shows a strong power to address multi-modal tasks. In addition, the decoder-only model has the advantage for self-supervised without pair data while we have paired data which this advance is greatly weakened. We are still working on collecting a large motion dataset for larger motion-language models.

Answer: To ensure a shared distribution between language and motion, we initialize the motion tokens separately and concatenate them alongside the language tokens. This step ensures a balanced representation that encompasses both modalities. Besides the token embeddings are actively trained during the entirety of stages 2 and 3, ensuring a comprehensive fusion of language and motion knowledge.

Answer: To address individual tasks, we adopt a focused approach where the entire model is fine-tuned. Our rationale lies in the fact that, for each specific task, our emphasis is on optimizing task-specific performance, without retaining an excessive amount of intelligence learned from other tasks. Besides, we only exclusively fine-tune the text-to-motion task, while other tasks are reported without specific tuning.

Comparison of motion prediction on HumanML3D dataset using motion data only.

Answer: Referring to MDM, motion editing has two categories: body part editing and motion completion in the temporal domain. MotionGPT is capable of the latter, which includes motion prediction and motion in-between. It outperforms both MDM and T2M-GPT in the table above. However, when it comes to body part editing, the vector quantization(VQ)-based methods, like MotionGPT and T2M-GPT, are not as suitable as diffusion-based models that utilize diffusion inpainting on raw motion data. Editing body parts with LLM and prompts is a promising direction but still needs exploration.

Answer: Please follow the approach outlined in Appendix B.4 and Line-296 of our paper, where we highlight that MDM achieves the motion in-between task using a masked motion "in-painting" technique. Specifically, this involves fixing the initial and final portions of the motion and allowing the model to generate the central portion. To adapt this concept for motion prediction, we similarly fix a portion of the motion – in our case, the first 20% – and generate the subsequent sequence.

Answer: VQ-based methods, such as MotionGPT and T2M-GPT, employ downsampling tricky to enhance the density of the codebook or tokens and reduce computing costs. This indeed becomes a constraint when the operation granularity is smaller than the down-sample rate. However, to address this issue, only the start and end frames are provided as in-between inputs. Some technical tricks can be used, such as repeating a single start or end frame up to the window size as inputs and removing the redundant parts in outputs. This does not significantly impact the effectiveness of the model, as there are often static beginnings or endings in the ground truth (GT) motion data.

Answer: We selected the down-sample rate based on the frames-per-second (FPS) of the HumanML3D and KIT-ML datasets, which is 20 fps. Therefore, down-sampling by a factor of 4 to achieve 5 fps can ensure distinctiveness in motion frames, and prevents redundancy, and acceleration training. This choice was also made to ensure a fair comparison, as we utilized the same down-sample rate as T2M-GPT. As shown in the above table, we provide an ablation study on these parameters, where a factor of 4 achieves the best Frechet Inception Distance (FID) in motion reconstructions.

Answer: As shown in Fig. 12, we provide both zero-shot cases and failure cases. Benefitting from strong language models, MotionGPTs can understand unseen works in the text-to-motion training set, like "scuttling" and "barriers", and generate correct motions based on the meaning of sentences. However, it still struggles to generate unseen motions, like gymnastics, even if MotionGPTs understand the text inputs.

Motion reconstruciton comparision.

Comparison of FID in text-to-motion task on KIT-ML dataset.

Answer: Given sufficient training or testing data from the same dataset, motion reconstruction is not a challenging task for both VAE and VQ-VAE. We have provided the evaluation on motion reconstruction in Tab.8. However, when dealing with a limited amount of motion data, like the KIT dataset, the VAE model shows better ability in motion interpolation, surpassing VQ-VAE. A relevant evaluation is shown above (also in Tab.7), where MLD (VAE) outperforms MotionGPT and T2M-GPT (VQ-VAEs) on FID. The real challenge lies in reconstructing complex motions, such as diving or gymnastics sports. Existing motion generators struggle to accurately reconstruct complex motions using a codebook extracted from daily motion datasets. Collecting these complex yet valuable motions is still a significant challenge to the motion research community.

Comparison of FID in text-to-motion task on HumanML3D dataset.

Comparison of FID in text-to-motion task on KIT-ML dataset.

Answer: The FID metrics primarily focus on the motion quality rather than the correlation between motion and text. While MDM serves as a successful benchmark for motion generation, both MotionGPT and T2M-GPT outperform MDM by a margin of 0.38~0.43 on the FID scale. However, the difference in motion quality among these three works is not significant in video supply. Additionally, MDM outperforms two vector quantized methods, MotionGPT and T2M-GPT, in terms of FID on the KIT dataset. This can be attributed to the limited number of 3,911 motion sequences, which makes it challenging to construct a comprehensive motion codebook. More importantly, MotionGPT contributes to multiple motion tasks with LLM, particularly in generating both text and motion within a single model, rather than aiming to improve the FID metric.

Answer: We thought MotionGPT, using a significantly larger language model, would surpass all existing methods in all tasks. However, the evaluation shows MotionGPT achieves SOTA results in 18 out of 23 metrics, where many improvements are only small gains. This can be attributed to the limited size of the dataset. Both HumanML3D (14,616 motions) and KIT (3,911 motions) are limited in vocabulary size and overall dataset size, particularly when compared to billion-level language datasets, which affects the efficacy of large-scale models. Benefitting from recent dataset works, like Motion-X, we will evaluate the performance gain of MotionGPT in larger datasets once they become available.

Answer: The evaluation of R-Precision in the KIT dataset relies on the text encoder, which is built using a limited set of 6,353 textual descriptions. In contrast, MotionGPTs benefit from LLM and large language data, enabling them to generate longer and more natural language descriptions for motion. However, this leads to a discrepancy between the generated descriptions and the GT descriptions, resulting in a lower R-Precision.

Answer: As shown in Fig. 10, MotionGPT achieves SOTA on 18 out of 23 metrics across four motion-related tasks. Additionally, both HumanML3D and KIT are limited in overall dataset size, particularly when compared to billion-level language datasets. This affects the efficacy of large-scale models. We will further employ a larger motion-text dataset to evaluate MotionGPT. Besides, MotionGPTs introduce motion-language pre-training, as well as its zero-shot ability, which is a promising direction worth exploring and could stimulate self-training procedures for further research.

Answer: As shown in Fig.13, we visualize these motion tokens in motion vocabulary $V_m$ and their corresponding localized spatial-temporal contexts, depicted within 4-frame motion segments. However, MotionGPT falls short in generating descriptions for each individual token, as the training is conducted on token sequences.

You can run the script below to visualize more tokens:

python -m scripts.get_code_visual --cfg configs/config_h3d_stage2.yaml