This repository provides the official PyTorch implementation for the following paper:
Text2Human: Text-Driven Controllable Human Image Generation
Yuming Jiang, Shuai Yang, Haonan Qiu, Wayne Wu, Chen Change Loy and Ziwei Liu
In ACM Transactions on Graphics (Proceedings of SIGGRAPH), 2022.
From MMLab@NTU affliated with S-Lab, Nanyang Technological University and SenseTime Research.
[Project Page] | [Paper] | [Dataset] | [Demo Video] | [Gradio Web Demo]
- [09/2022] π₯π₯π₯We have released a high-quality 3D human generative model EVA3D!π₯π₯π₯
- [07/2022] Release the model trained on SHHQ dataset!
- [07/2022] Try out the web demo of drawings-to-human! .
- [06/2022] Integrated into Huggingface Spaces π€ using Gradio. Try out the Web Demo:
- [05/2022] Paper and demo video are released.
- [05/2022] Code is released.
- [05/2022] This website is created.
Clone this repo:
git clone https://github.com/yumingj/Text2Human.git
cd Text2Human
Dependencies:
All dependencies for defining the environment are provided in environment/text2human_env.yaml
.
We recommend using Anaconda to manage the python environment:
conda env create -f ./environment/text2human_env.yaml
conda activate text2human
pip install mmcv-full==1.2.1 -f https://download.openmmlab.com/mmcv/dist/cu101/torch1.7.0/index.html
pip install mmsegmentation==0.9.0
conda install -c huggingface tokenizers=0.9.4
conda install -c huggingface transformers=4.0.0
conda install -c conda-forge sentence-transformers=2.0.0
If it doesn't work, you may need to install the following packages on your own:
- Python 3.6
- PyTorch 1.7.1
- CUDA 10.1
- sentence-transformers 2.0.0
- tokenizers 0.9.4
- transformers 4.0.0
In this work, we contribute a large-scale high-quality dataset with rich multi-modal annotations named DeepFashion-MultiModal Dataset. Here we pre-processed the raw annotations of the original dataset for the task of text-driven controllable human image generation. The pre-processing pipeline consists of:
- align the human body in the center of the images according to the human pose
- fuse the clothing color and clothing fabric annotations into one texture annotation
- do some annotation cleaning and image filtering
- split the whole dataset into the training set and testing set
You can download our processed dataset from this Google Drive. If you want to access the raw annotations, please refer to the DeepFashion-MultiModal Dataset.
After downloading the dataset, unzip the file and put them under the dataset folder with the following structure:
./datasets
βββ train_images
βββ xxx.png
...
βββ xxx.png
βββ xxx.png
βββ test_images
% the same structure as in train_images
βββ densepose
% the same structure as in train_images
βββ segm
% the same structure as in train_images
βββ shape_ann
βββ test_ann_file.txt
βββ train_ann_file.txt
βββ val_ann_file.txt
βββ texture_ann
βββ test
βββ lower_fused.txt
βββ outer_fused.txt
βββ upper_fused.txt
βββ train
% the same files as in test
βββ val
% the same files as in test
Unofficial Demo implemented by @neverix.
Pretrained models can be downloaded from the model zoo. Unzip the file and put them under the pretrained_models folder with the following structure:
pretrained_models
βββ index_pred_net.pth
βββ parsing_gen.pth
βββ parsing_token.pth
βββ sampler.pth
βββ vqvae_bottom.pth
βββ vqvae_top.pth
Model | Dataset | Annotations |
---|---|---|
Standard Model | DeepFashion-Multimodal | Follow the dataset preparation in Step(1) |
Extended Model | SHHQ | Replace the annotations with the following ones: densepose, segm, shape, texture |
Remark: For fair research comparisons, it is suggested to use the standard model.
You can generate images from given parsing maps and pre-defined texture annotations:
python sample_from_parsing.py -opt ./configs/sample_from_parsing.yml
The results are saved in the folder ./results/sampling_from_parsing
.
You can generate images from given human poses and pre-defined clothing shape and texture annotations:
python sample_from_pose.py -opt ./configs/sample_from_pose.yml
Remarks: The above two scripts generate images without language interactions. If you want to generate images using texts, you can use the notebook or our user interface.
python ui_demo.py
The descriptions for shapes should follow the following format:
<gender>, <sleeve length>, <length of lower clothing>, <outer clothing type>, <other accessories1>, ...
Note: The outer clothing type and accessories can be omitted.
Examples:
man, sleeveless T-shirt, long pants
woman, short-sleeve T-shirt, short jeans
The descriptions for textures should follow the following format:
<upper clothing texture>, <lower clothing texture>, <outer clothing texture>
Note: Currently, we only support 5 types of textures, i.e., pure color, stripe/spline, plaid/lattice,
floral, denim. Your inputs should be restricted to these textures.
Train the parsing generation network. If you want to skip the training of this network, you can download our pretrained model from here.
python train_parsing_gen.py -opt ./configs/parsing_gen.yml
Step 1: Train the top level of the hierarchical VQVAE. We provide our pretrained model here. This model is trained by:
python train_vqvae.py -opt ./configs/vqvae_top.yml
Step 2: Train the bottom level of the hierarchical VQVAE. We provide our pretrained model here. This model is trained by:
python train_vqvae.py -opt ./configs/vqvae_bottom.yml
Stage 3 & 4: Train the sampler with mixture-of-experts. To train the sampler, we first need to train a model to tokenize the parsing maps. You can access our pretrained parsing maps here.
python train_parsing_token.py -opt ./configs/parsing_token.yml
With the parsing tokenization model, the sampler is trained by:
python train_sampler.py -opt ./configs/sampler.yml
Our pretrained sampler is provided here.
Stage 5: Train the index prediction network. We provide our pretrained index prediction network here. It is trained by:
python train_index_prediction.py -opt ./configs/index_pred_net.yml
Remarks: In the config files, we use the path to our models as the required pretrained models. If you want to train the models from scratch, please replace the path to your own one. We set the numbers of the training epochs as large numbers and you can choose the best epoch for each model. For your reference, our pretrained parsing generation network is trained for 50 epochs, top-level VQVAE is trained for 135 epochs, bottom-level VQVAE is trained for 70 epochs, parsing tokenization network is trained for 20 epochs, sampler is trained for 95 epochs, and the index prediction network is trained for 70 epochs.
Please visit our Project Page to view more results.
You can select the attribtues to customize the desired human images.
In this work, we also propose DeepFashion-MultiModal, a large-scale high-quality human dataset with rich multi-modal annotations. It has the following properties:
- It contains 44,096 high-resolution human images, including 12,701 full body human images.
- For each full body images, we manually annotate the human parsing labels of 24 classes.
- For each full body images, we manually annotate the keypoints.
- We extract DensePose for each human image.
- Each image is manually annotated with attributes for both clothes shapes and textures.
- We provide a textual description for each image.
Please refer to this repo for more details about our proposed dataset.
If you find this work useful for your research, please consider citing our paper:
@article{jiang2022text2human,
title={Text2Human: Text-Driven Controllable Human Image Generation},
author={Jiang, Yuming and Yang, Shuai and Qiu, Haonan and Wu, Wayne and Loy, Chen Change and Liu, Ziwei},
journal={ACM Transactions on Graphics (TOG)},
volume={41},
number={4},
articleno={162},
pages={1--11},
year={2022},
publisher={ACM New York, NY, USA},
doi={10.1145/3528223.3530104},
}
Part of the code is borrowed from unleashing-transformers, taming-transformers and mmsegmentation.