HumanSD

---

This repository contains the implementation of the ICCV2023 paper:

HumanSD: A Native Skeleton-Guided Diffusion Model for Human Image Generation[Project Page][Paper][Code][Video][Data]
Xuan Ju^∗12, Ailing Zeng^∗1, Chenchen Zhao^∗2, Jianan Wang¹, Lei Zhang¹, Qiang Xu^2
∗ Equal contribution ¹International Digital Economy Academy ²The Chinese University of Hong Kong

In this work, we propose a native skeleton-guided diffusion model for controllable HIG called HumanSD. Instead of performing image editing with dual-branch diffusion, we fine-tune the original SD model using a novel heatmap-guided denoising loss. This strategy effectively and efficiently strengthens the given skeleton condition during model training while mitigating the catastrophic forgetting effects. HumanSD is fine-tuned on the assembly of three large-scale human-centric datasets with text-imagepose information, two of which are established in this work.

---

• (a) a generation by the pre-trained pose-less text-guided stable diffusion (SD)
• (b) pose skeleton images as the condition to ControlNet and our proposed HumanSD
• (c) a generation by ControlNet
• (d) a generation by HumanSD (ours). ControlNet and HumanSD receive both text and pose conditions.

HumanSD shows its superiorities in terms of (I) challenging poses, (II) accurate painting styles, (III) pose control capability, (IV) multi-person scenarios, and (V) delicate details.

Table of Contents

• HumanSD
  • TODO
  • Model Overview
  • Getting Started
    • Environment Requirement
    • Model and Checkpoints
    • Quick Demo
    • Dataset
  • Training
  • Quantitative Results
  • Qualitative Results
    • Natural Scene
    • Sketch Scene
    • Shadow Play Scene
    • Children Drawing Scene
    • Oil Painting Scene
    • Watercolor Scene
    • Digital Art Scene
    • Relief Scene
    • Sculpture Scene
  • Cite Us
  • Acknowledgement

TODO

News!! Our paper have been accepted by ICCV2023! Training code is released.

• Release inference code and pretrained models
• Release Gradio UI demo
• Public training data (LAION-Human)
• Release training code

Model Overview

Getting Started

Environment Requirement

HumanSD has been implemented and tested on Pytorch 1.12.1 with python 3.9.

Clone the repo:

git clone git@github.com:IDEA-Research/HumanSD.git

We recommend you first install pytorch following official instructions. For example:

# conda
conda install pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1 cudatoolkit=11.3 -c pytorch

Then, you can install required packages thourgh:

pip install -r requirements.txt

You also need to install MMPose following here. Noted that you only need to install MMPose as a python package. PS: Because of the update of MMPose, we recommend you to install 0.29.0 version of MMPose.

Model and Checkpoints

Download necessary checkpoints of HumanSD, which can be found here. The data structure should be like:

|-- humansd_data
|-- checkpoints
|-- higherhrnet_w48_humanart_512x512_udp.pth
|-- v2-1_512-ema-pruned.ckpt
|-- humansd-v1.ckpt

Noted that v2-1_512-ema-pruned.ckpt should be download from Stable Diffusion.

Quick Demo

You can run demo either through command line or gradio.

You can run demo through command line with:

python scripts/pose2img.py --prompt "oil painting of girls dancing on the stage" --pose_file assets/pose/demo.npz

You can also run demo compared with ControlNet and T2I-Adapter:

python scripts/pose2img.py --prompt "oil painting of girls dancing on the stage" --pose_file assets/pose/demo.npz --controlnet --t2i

You can run gradio demo through:

python scripts/gradio/pose2img.py

We have also provided the comparison of ControlNet and T2I-Adapter, you can run all these methods in one demo. But you need to download corresponding model and checkpoints following:

To compare ControlNet, and T2I-Adpater's results.

(1) You need to initialize ControlNet and T2I-Adapter as submodule using

git submodule init
git submodule update

(2) Then download checkpoints from: a. T2I-Adapter b. ControlNet. And put them into humansd_data/checkpoints

Then, run:

python scripts/gradio/pose2img.py --controlnet --t2i

Noted that you may have to modify some code in T2I-Adapter due to the path conflict.

e.g., use

from comparison_models.T2IAdapter.ldm.models.diffusion.ddim import DDIMSampler

instead of

from T2IAdapter.ldm.models.diffusion.ddim import DDIMSampler

Dataset

You may refer to the code here for loading the data.

Laion-Human

You may apply for access of Laion-Human here. Noted that we have provide the pose annotations, images' .parquet file and mapping file, please download the images according to .parquet. The key in .parquet is the corresponding image index. For example, image with key=338717 in 00033.parquet is corresponding to images/00000/000338717.jpg.

After downloading the images and pose, you need to extract zip files and make it looks like:

|-- humansd_data
|-- datasets
|-- Laion |-- Aesthetics_Human
|-- images
|-- 00000.parquet
|-- 00001.parquet
|-- ...
|-- pose
|-- 00000
|-- 000000000.npz
|-- 000000001.npz
|-- ...
|-- 00001
|-- ... |-- mapping_file_training.json 

Then, you can use python utils/download_data.py to download all images.

Then, the file data structure should be like:

|-- humansd_data
|-- datasets
|-- Laion |-- Aesthetics_Human
|-- images
|-- 00000.parquet
|-- 00001.parquet
|-- ...
|-- 00000
|-- 000000000.jpg
|-- 000000001.jpg
|-- ...
|-- 00001
|-- ...
|-- pose
|-- 00000
|-- 000000000.npz
|-- 000000001.npz
|-- ...
|-- 00001
|-- ... |-- mapping_file_training.json 

If you download the LAION-Aesthetics in tar files, which is different from our data structure, we recommend you extract the tar file through code:

importtarfiletar_file="00000.tar"# 00000.tar - 00286.tarpresent_tar_path=f"xxxxxx/{tar_file}"save_dir="humansd_data/datasets/Laion/Aesthetics_Human/images"withtarfile.open(present_tar_path, "r") astar_file:
forpresent_fileintar_file.getmembers():
ifpresent_file.name.endswith(".jpg"):
print(f" image:- {present_file.name} -")
image_save_path=os.path.join(save_dir,tar_file.replace(".tar",""),present_file.name)
present_image_fp=TarIO.TarIO(present_tar_path, present_file.name)
present_image=Image.open(present_image_fp)
present_image_numpy=cv2.cvtColor(np.array(present_image),cv2.COLOR_RGB2BGR)
ifnotos.path.exists(os.path.dirname(image_save_path)):
os.makedirs(os.path.dirname(image_save_path))
cv2.imwrite(image_save_path,present_image_numpy)

Human-Art

You may download Human-Art dataset here.

The file data structure should be like:

|-- humansd_data
|-- datasets
|-- HumanArt |-- images
|-- 2D_virtual_human
|-- cartoon
|-- 000000000007.jpg
|-- 000000000019.jpg
|-- ...
|-- digital_art
|-- ...
|-- 3D_virtual_human
|-- real_human
|-- pose
|-- 2D_virtual_human
|-- cartoon
|-- 000000000007.npz
|-- 000000000019.npz
|-- ...
|-- digital_art
|-- ...
|-- 3D_virtual_human
|-- real_human
|-- mapping_file_training.json |-- mapping_file_validation.json 

Training

Note that the datasets and checkpoints should be downloaded and prepared before training.

Run the commands below to start training:

python main.py --base configs/humansd/humansd-finetune.yaml -t --gpus 0,1 --name finetune_humansd

If you want to finetune without heat-map-guided diffusion loss for ablation, you can run the following commands:

python main.py --base configs/humansd/humansd-finetune-originalloss.yaml -t --gpus 0,1 --name finetune_humansd_original_loss

Quantitative Results

Metrics can be calculated through:

python scripts/pose2img_metrics.py --outdir outputs/metrics --config utils/metrics/metrics.yaml --ckpt path_to_ckpt

Qualitative Results

• (a) a generation by the pre-trained text-guided stable diffusion (SD)
• (b) pose skeleton images as the condition to ControlNet, T2I-Adapter and our proposed HumanSD
• (c) a generation by ControlNet
• (d) a generation by T2I-Adapter
• (e) a generation by HumanSD (ours).

ControlNet, T2I-Adapter, and HumanSD receive both text and pose conditions.

Natural Scene

Sketch Scene

Shadow Play Scene

Children Drawing Scene

Oil Painting Scene

Watercolor Scene

Digital Art Scene

Relief Scene

Sculpture Scene

Cite Us

@article{ju2023humansd,
title={Human{SD}: A Native Skeleton-Guided Diffusion Model for Human Image Generation},
author={Ju, Xuan and Zeng, Ailing and Zhao, Chenchen and Wang, Jianan and Zhang, Lei and Xu, Qiang},
booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
year={2023}
}
@inproceedings{ju2023human,
title={Human-Art: A Versatile Human-Centric Dataset Bridging Natural and Artificial Scenes},
author={Ju, Xuan and Zeng, Ailing and Wang, Jianan and Xu, Qiang and Zhang, Lei},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
year={2023},
}

Acknowledgement

• Our code is modified on the basis of Stable Diffusion, thanks to all the contributors!
• HumanSD would not be possible without LAION and their efforts to create open, large-scale datasets.
• Thanks to the DeepFloyd team at Stability AI, for creating the subset of LAION-5B dataset used to train HumanSD.
• HumanSD uses OpenCLIP, trained by Romain Beaumont.