AnimationGPT is a project focused on generating combat style character animations based on text. This project is trained on the MotionGPT and has produced the first character animation dataset dedicated to combat styles, named CombatMotion, which comes with textual descriptions.

Compare to current text-to-motion dataset

Compared to the current text-to-motion datasets, CombatMotion has the following characteristics:

1. Derived from game assets.
2. Features a fighting style, where the animation style in action games tends to be concentrated, and the types of actions are biased.
3. More detailed textual annotations.

1. Obtain game assets in FBX format, redirect them to SMPL, and read the coordinates of human body joints (refer to Fbx2SMPL);

2. Add textual annotations. For each animation, manually annotate it from the following aspects: action type, weapon type, attack type, locational words, power descriptor words, speed descriptor words, and confusion descriptor words. A partial list of terms is shown below:

   Action type	Weapon type	Attack type	Locative words	Power	Speed	Fuzzy
   Idle	Bare Hand	Left-Handed	In-Place	Light-Weighted	Swift	Piercing
   Get Hit	Sacred Seal	Right-Handed	Towards Left	Steady	Relative Fast	Slash
   Death	Fist	One-Handed	Towards Right	Heavy-Weighted	Uniform Speed	Blunt
   …	…	…	…	…	…	…

   Then, use GPT-4 to combine these annotations into sentences.

3. Process the animation and annotated data into a format compatible with HumanML3D.

Download: google drive

CombatMotionProcessed(CMP) is a refined dataset that, in terms of character animation, retains 8,700 high-quality animations with a strong fighting style. In terms of textual annotations, we provide three text annotations for each animation: a concise description, a concise description with sensory details, and a detailed description.

Taking CMP008388 as an example, its corresponding text annotations are:

weapon attack a man holding a Katana,executing a Charged Heavy Attack,Dual Wielding,root motion get Forward, Steady,Powerful and Relative Slow,First slow then fast,Cleanly.
weapon attack a man holding a Katana,executing a Charged Heavy Attack,Dual Wielding,root motion get Forward, Steady,Powerful and Relative Slow,First slow then fast,Cleanly,which make a sense of Piercing,Wide Open,Charged,Accumulating strength.
The character grips the wedge with both hands and charges for a powerful strike. They firmly lower their body, twist to the left, lunge forward with a bow step, and stab with the sword held in both hands.

Download: google drive

CombatMotionRaw (CMR) is an unrefined dataset containing 14,883 animation entries (CMP is a subset of CMR), but each animation is only provided with one textual annotation. Moreover, the textual annotations in CMR consist of simple concatenations of annotated words. It was found during project development that models trained with this type of annotation performed poorly, thus this format was ultimately not adopted.

Example of textual annotation:

weapon attack curved sword curved greatsword right-handed one-handed charged heavy attack forward steady powerful charged accumulating strength cleanly first slow then fast slash smooth and coherent wide open featherlike roundabout lean over and twist your waist to the left step forward with your right leg store your right hand from the left back swing it diagonally downward and swing two circles.

CMR has a richer set of animation data, unfortunately, the annotations are not detailed enough. You can read the textual annotations from the dataset yourself and refine them.

Here are models trained on the CMP dataset using different algorithms:

• MotionGPT Model：google drive
• MLD Model：google drive
• MDM Model：google drive

Evaluation on CMP

• If you need to train a model, please download the CMP dataset. Then, follow the tutorials for MotionGPT or other text-to-motion algorithms to set up the environment and train your model.

• If you only need to use the AGPT model trained on the CMP dataset, please follow these steps:

  1. Set up the environment

     Our experimental environment is Ubuntu 22.04, NVIDIA GeForce RTX 4090, and CUDA 11.8

     git clone https://github.com/OpenMotionLab/MotionGPT.git
     cd MotionGPT
     conda create python=3.10 --name mgpt
     conda activate mgpt
     pip install -r requirements.txt
     python -m spacy download en_core_web_sm
     mkdir deps
     cd deps
     bash prepare/prepare_t5.sh
     bash prepare/download_t2m_evaluators.sh
     

  2. Download the CMP dataset

     Unzip the dataset into the datasets/humanml3d directory.

     .
     └── humanml3d
         ├── new_joint_vecs
         ├── new_joints
         └── texts
     

  3. Generate animations using the model

     • git clone https://github.com/fyyakaxyy/AnimationGPT.git

     • Copy the tools folder and config_AGPT.yaml into the MotionGPT directory

     • Download the AGPT model, place it in the MotionGPT directory

     • Save the prompt in input.txt

     • Run python demo.py --cfg ./config_AGPT.yaml --example ./input.txt

     The generated result is id_out.npy, stored in results/mgpt/debug--AGPT/

  4. File format conversion

     • Convert the generated npy files to mp4 files: modify the file path in tools/animation.py, then run: python animation.py
     • Convert the generated npy files to bvh files: modify the file path in tools/npy2bvh/joints2bvh.py, then run: python joints2bvh.py Note: The code for npy2bvh is sourced from Momask

During the process of dataset creation and model training/tuning, you might encounter some issues in aspects like textual annotations, model training, and data augmentation. Based on our experience, we offer the following suggestions:

If you process data using the HumanML3D pipeline, you might encounter the following issues, which can lead to model training crashes:

• The textual description contains Chinese characters or Chinese punctuation.
• Some words fail to be successfully annotated with part-of-speech tags.
• Certain mathematical symbols, such as the degree symbol "°", are recognized as abnormal characters.

• Adding descriptions of root motion direction in the annotated text can help the model learn directional words.
• Adding frame number information to the annotated text does not enable the model to learn how to control the duration (or number of frames) of generation.
• The more detailed the textual annotations and the greater the number of different annotations for the same animation, the better the performance of the model.

Mixing the HumanML3D, KIT-ML, and CMP datasets for model training can result in significant improvements in evaluation metrics.

However, evaluation metrics and visual effects are not equivalent. For some generated results, models trained on a mixed dataset perform worse than those trained solely on the CMP dataset. Because differences in action styles between datasets change the data distribution, thereby affecting model performance.

We have attempted to convert Motion-X into the HumanML3D format for pre-training models or to expand the codebook length of VQ-VAE to increase the richness and stylization of actions. However, the work on data conversion failed. The specific content and code can be viewed here.

• Algorithm: Thanks to MotionGPT, MLD, MDM and Momask.
• Dataset: Thanks to HumanML3D and Motion-X.

Our code is partially borrowing from them.

If you find this repository useful, please consider citing it as follows:

@misc{CombatMotion,
  title={AnimationGPT:An AIGC tool for generating game combat motion assets},
  author={Yihao Liao, Yiyu Fu, Ziming Cheng, Jiangfeiyang Wang},
  year={2024},
  howpublished={\url{https://github.com/fyyakaxyy/AnimationGPT}}
}

@InProceedings{Guo_2022_CVPR,
    author    = {Guo, Chuan and Zou, Shihao and Zuo, Xinxin and Wang, Sen and Ji, Wei and Li, Xingyu and Cheng, Li},
    title     = {Generating Diverse and Natural 3D Human Motions From Text},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2022},
    pages     = {5152-5161}
}

@article{jiang2024motiongpt,
  title={Motiongpt: Human motion as a foreign language},
  author={Jiang, Biao and Chen, Xin and Liu, Wen and Yu, Jingyi and Yu, Gang and Chen, Tao},
  journal={Advances in Neural Information Processing Systems},
  volume={36},
  year={2024}
}

AnimationGPT是一个基于文本生成格斗风格角色动画的项目。本项目基于MotionGPT训练模型，并且制作了首个专注于格斗风格、并配备文本描述的角色动画数据集CombatMotion。

对比现有文本-动作数据集

与现有文本-动作数据集相比，CombatMotion具有如下特点：

1. 来源于游戏资产。
2. 具有格斗风格，动作类游戏当中的动画风格相对来说是集中的，动作类型有偏。
3. 具有更详细的文本标注。

1. 获取fbx格式的游戏资产，重定向到SMPL，并读取人体关节点坐标（参考Fbx2SMPL）；

2. 添加文本标注。对于每一条动画，首先从动作类型、武器类型、攻击类型、方位词、力量感描述词、速度描述词和困惑描述词这几个方面添加人工标注，部分词表如下所示：

   Action type	Weapon type	Attack type	Locative words	Power	Speed	Fuzzy
   Idle	Bare Hand	Left-Handed	In-Place	Light-Weighted	Swift	Piercing
   Get Hit	Sacred Seal	Right-Handed	Towards Left	Steady	Relative Fast	Slash
   Death	Fist	One-Handed	Towards Right	Heavy-Weighted	Uniform Speed	Blunt
   …	…	…	…	…	…	…

   然后通过GPT-4将这些标注连接成句子。

3. 将动画和标注数据处理成HumanML3D格式的数据。

下载链接：google drive

CombatMotionProcessed(CMP)是精加工的数据集，在角色动画方面，我们保留了高质量、格斗风格强的8,700个动画，在文本标注方面，我们为每一条动画提供了3条文本标注，分别是精简版描述、带有感觉描述的精简版描述和详细版描述。

以CMP008388为例，其对应的文本标注是：

weapon attack a man holding a Katana,executing a Charged Heavy Attack,Dual Wielding,root motion get Forward, Steady,Powerful and Relative Slow,First slow then fast,Cleanly.
weapon attack a man holding a Katana,executing a Charged Heavy Attack,Dual Wielding,root motion get Forward, Steady,Powerful and Relative Slow,First slow then fast,Cleanly,which make a sense of Piercing,Wide Open,Charged,Accumulating strength.
The character grips the wedge with both hands and charges for a powerful strike. They firmly lower their body, twist to the left, lunge forward with a bow step, and stab with the sword held in both hands.

下载链接：google drive

CombatMotionRaw(CMR)是未经过精加工的数据集，具备14,883个的动画数据（CMP是CMR的子集），但每条动画只提供一个文本标注。另外，CMR中的文本标注是标注词的简单连接，在项目研发中发现这种标注训练的模型性能较差，因此最终未采用这种格式。

文本标注示例：

weapon attack curved sword curved greatsword right-handed one-handed charged heavy attack forward steady powerful charged accumulating strength cleanly first slow then fast slash smooth and coherent wide open featherlike roundabout lean over and twist your waist to the left step forward with your right leg store your right hand from the left back swing it diagonally downward and swing two circles.

CMR具备更丰富的动画数据，可惜标注不够精细，您可以自行读取数据集中的文本标注并优化。

以下分别是在CMP数据集上用不同算法训练的模型：

• MotionGPT Model：google drive
• MLD Model：google drive
• MDM Model：google drive

CMP数据集上的评估结果

• 如果需要训练模型，请下载CMP数据集，然后按照MotionGPT或其它text-to-motion算法的教程配置环境，训练模型。

• 如果只需要使用CMP数据集上训练好的AGPT模型，请参考以下步骤：

  1. 配置环境

     我们的实验环境是Ubuntu22.04，NVIDIA GeForce RTX 4090，CUDA 11.8

     git clone https://github.com/OpenMotionLab/MotionGPT.git
     cd MotionGPT
     conda create python=3.10 --name mgpt
     conda activate mgpt
     pip install -r requirements.txt
     python -m spacy download en_core_web_sm
     mkdir deps
     cd deps
     bash prepare/prepare_t5.sh
     bash prepare/download_t2m_evaluators.sh
     

  2. 下载CMP数据集

     将数据集解压到datasets/humanml3d路径下。

     .
     └── humanml3d
         ├── new_joint_vecs
         ├── new_joints
         └── texts
     

  3. 使用模型生成动画

     • git clone https://github.com/fyyakaxyy/AnimationGPT.git

     • 将tools文件夹和config_AGPT.yaml复制到MotionGPT文件夹下

     • 下载AGPT model，放置在MotionGPT路径下

     • 将prompt保存到input.txt中

     • python demo.py --cfg ./config_AGPT.yaml --example ./input.txt

     生成的结果是id_out.npy，保存在results/mgpt/debug--AGPT/路径下

  4. 文件格式转换

     • 将生成的npy文件转为mp4文件：修改tools/animation.py中的文件路径，然后运行：python animation.py
     • 将生成的npy文件转为bvh文件：修改tools/npy2bvh/joints2bvh.py中的文件路径，然后运行：python joints2bvh.py 备注：npy2bvh的代码来源于Momask

在数据集制作和模型训练调优的过程中，您可能会在文本标注、模型训练、数据增强等方面遇到一些问题。基于我们的经验，给出以下建议：

如果采用HumanML3D的pipline处理数据，可能会遇到以下问题，它们将会导致模型训练崩溃：

• 文本描述中包含中文字符或中文标点。
• 部分词语无法成功添加词性标注。
• 部分数学符号，例如角度"°"被识别为异常字符。

• 在标注文本中添加对root motion的方位词描述，可以让模型学习到方位词。
• 在标注文本中添加帧数信息，并不能让模型学会控制生成时长（或帧数）。
• 文本标注越详细、同一条动画的不同标注数量越多，模型的性能越好。

将HumanML3D、KIT-ML和CMP数据集混合起来训练模型，在评估指标上会带来巨大提升，但评估指标和视觉效果并不等价，对于部分生成结果，混合训练的模型表现不如单独使用CMP数据集训练的模型，这是因为不同数据集动作风格的差异改变了数据分布，进而影响了模型的性能。

我们尝试过将Motion-X转换成HumanML3D的格式，用于预训练模型，或者扩充VQ-VAE的码本长度来增加动作的丰富性和风格化程度，但数据转换的工作失败了。具体内容和代码在此处查看。

• 算法：感谢 MotionGPT 、MLD、 MDM以及Momask。
• 数据集：感谢 HumanML3D 和 Motion-X。

我们的代码部分借鉴了以上工作。

如果您觉得这个仓库对您有用，请考虑引用：

@misc{CombatMotion,
  title={AnimationGPT:An AIGC tool for generating game combat motion assets},
  author={Yihao Liao, Yiyu Fu, Ziming Cheng, Jiangfeiyang Wang},
  year={2024},
  howpublished={\url{https://github.com/fyyakaxyy/AnimationGPT}}
}

@InProceedings{Guo_2022_CVPR,
    author    = {Guo, Chuan and Zou, Shihao and Zuo, Xinxin and Wang, Sen and Ji, Wei and Li, Xingyu and Cheng, Li},
    title     = {Generating Diverse and Natural 3D Human Motions From Text},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2022},
    pages     = {5152-5161}
}

@article{jiang2024motiongpt,
  title={Motiongpt: Human motion as a foreign language},
  author={Jiang, Biao and Chen, Xin and Liu, Wen and Yu, Jingyi and Yu, Gang and Chen, Tao},
  journal={Advances in Neural Information Processing Systems},
  volume={36},
  year={2024}
}