我们希望训练一个视觉语言大模型成为成美妆大师,给你提供穿衣和美妆建议。
欢迎大家fork和star我们的项目!!!
- 主要
- 训练视觉语言大模型具备基本的审美能力;
- 对齐人类的审美能力;
- 让大模型为我们提供穿搭建议;
- 将大模型的穿搭建议展示出来。
- 流程图
- 数据处理(WIP)
- 搜集时尚穿搭数据集;
- 使用视觉语言大模型为数据打标 internVL;
- 做一个标记工具;
- 用标记工具校正大模型打标的数据。
- RLHF instruct
- DPO DPO
- ORPO ORPO(https://github.com/xfactlab/orpo/tree/main)
- RLHF-V RLHF-V(https://github.com/RLHF-V/RLHF-V)
- 视觉语言大模型会给出穿搭建议(WIP)
- 视觉语言大模型会给出穿搭建议 internVL
-
Segment Tool
- segment-anything: Features: it can be used to generate masks for all objects in an image.
- Self Correction for Human ParsingFeatures: An out-of-box human parsing representation extractor.
-
Virtual Try-on(WIP)
- IDM-VTON: Features: it could keep background of the VToN, Techs: IP-Adapter, TryonNet, GarmentNet.
- COTTON-size-does-matter: Features: it can adjust the size of VToN with a parameter.
- OOTDiffusion: Techs: LDM-based, outfitting UNet, outfitting dropout, Classifier-free guidance.
- OutfitAnyone: Ultra-high quality virtual try-on for Any Clothing and Any Person (Including Anime character pictures in demo).
- StableVITON: Features: it could keep background of the VToN, Techs: LDM-based, zero cross-attention blocks, attention total variation loss and augmentation.
-
Visualize the results of the suggestions to the user
- ControlNet
- 4d-dress: A 4D Dataset of Real-world Human Clothing with Semantic Annotations
- champ: Controllable and Consistent Human Image Animation with 3D Parametric Guidance
- MOSS
├── README.md
├── README_cn.md
├── docs
├── scripts
├── beautymaster
│ ├── datasets
│ ├── 3rdparty
│ │ ├── internvl
│ │ ├── IDM-VTON
│ │ ├── champ
│ │ ├── archpp
│ │ ├── ControlNet
│ │ └── OOTDiffusion
│ ├── demo
│ ├── src
│ ├── train
│ └── utils
├── requirements.txt
- 下载代码
git clone https://github.com/RayTang88/BeautyMaster.git
cd BeautyMaster
- 安装基础环境
conda create -n beautyMaster python=3.10 -y
conda activate beautyMaster
- 安装依赖
pip install -r requirements.txt
请参考 Data.md
cd beautymaster
python demo/infer.py
使用和许可声明: 数据、代码和检查点仅供研究使用,不得用于商业用途。 它们还仅限于遵循 InternVL、RLHF、DPO、ORPO、RLHF-V、segment-anything、SCHP、IDM-VTON、ControlNet、4d-dress、champ 许可协议的使用。 数据集为 CC BY NC 4.0(仅允许非商业用途),使用该数据集训练的模型不应在研究目的之外使用。