Junhao Zhuang, ♦Yanhong Zeng, Wenran Liu, †Chun Yuan, †Kai Chen
(♦project lead, †corresponding author)
Your star means a lot for us to develop this project! ⭐
PowerPaint is a high-quality versatile image inpainting model that supports text-guided object inpainting, object removal, shape-guided object insertion, and outpainting at the same time. We achieve this by learning with tailored task prompts for different inpainting tasks.
May 22, 2024🔥
- We have open-sourced the model weights for PowerPaint v2-1, rectifying some existing issues that were present during the training process of version 2.
April 7, 2024🔥
- We open source the model weights and code for PowerPaint v2.
April 6, 2024:
- We have retrained a new PowerPaint, taking inspiration from Brushnet. The Online Demo has been updated accordingly. We plan to release the model weights and code as open source in the next few days.
- Tips: We preserve the cross-attention layer that was deleted by BrushNet for the task prompts input.
| Object insertion | Object Removal | Shape-guided Object Insertion | Outpainting | |
|---|---|---|---|---|
| Original Image |  |
 |
 |
 |
| Output |  |
 |
 |
 |
December 22, 2023🔧
- The logical error in loading ControlNet has been rectified. The
gradio_PowerPaint.pyfile and Online Demo have also been updated.
December 18, 2023
Enhanced PowerPaint Model
- We are delighted to announce the release of more stable model weights. These refined weights can now be accessed on Hugging Face. The
gradio_PowerPaint.pyfile and Online Demo have also been updated as part of this release.
Recommend Environment: cuda 11.8 + python 3.9
# Clone the Repository
git clone [email protected]:open-mmlab/PowerPaint.git
# Create Virtual Environment with Conda
conda create --name ppt python=3.9
conda activate ppt
# Install Dependencies
pip install -r requirements/requirements.txtOr you can construct a conda environment from scratch by running the following command:
conda env create -f requirements/ppt.yaml
conda activate pptYou can launch the Gradio interface for PowerPaint by running the following command:
# Set up Git LFS
conda install git-lfs
git lfs install
# Clone PowerPaint Model
git lfs clone https://huggingface.co/JunhaoZhuang/PowerPaint-v1/ ./checkpoints/ppt-v1
python app.py --shareWe suggest PowerPaint-V2 that is built upon BrushNet with RealisticVision as the base model, which exhibits higher visual quality. You can run the following command:
# Clone PowerPaint Model
git lfs clone https://huggingface.co/JunhaoZhuang/PowerPaint_v2/ ./checkpoints/ppt-v2
python app.py --share --version ppt-v2 --checkpoint_dir checkpoints/ppt-v2Specifically, if you have downloaded the weights and want to skip the step of cloning the model, you can skip that step by enabling --local_files_only.
After launching the Gradio interface, you can insert objects into images by uploading your image, drawing the mask, selecting the tab of Text-guided object inpainting and inputting the text prompt. The model will then generate the output image.
| Input | Output |
|---|---|
 |
 |
Fortunately, PowerPaint is compatible with ControlNet. Therefore, users can generate object with a control image.
| Input | Condition | Control Image | Output |
|---|---|---|---|
 |
Canny |  |
 |
|
Depth |  |
 |
|
HED |  |
 |
 |
Pose |  |
 |
For object removal, you need to select the tab of Object removal inpainting and you don't need to input any prompts. PowerPaint is able to fill in the masked region according to context background.
We remain the text box for inputing prompt, allowing users to further suppress object generation by using negative prompts. Specifically, we recommend to use 10 or higher value for Guidance Scale. If undesired objects appear in the masked area, you can address this by specifically increasing the Guidance Scale.
| Input | Output |
|---|---|
 |
 |
For image outpainting, you don't need to input any text prompt. You can simply select the tab of Image outpainting and adjust the slider for horizontal expansion ratio and vertical expansion ratio, then PowerPaint will extend the image for you.
| Input | Output |
|---|---|
 |
 |
PowerPaint also supports shape-guided object inpainting, which allows users to control the fitting degree of the generated objects to the shape of masks. You can select the tab of Shape-guided object inpainting and input the text prompt. Then, you can adjust the slider of fitting degree to control the shape of generated object.
Taking the following cases as example, you can draw a square mask and use a high fitting degree, e.g., 0.95, to generate a bread to fit in the mask shape. For the same mask, you can also use a low fitting degree, e.g., 0.55, to generate a reasonable result for rabbit. However, if you use a high fitting degree for the 'square rabit', the result may look funny.
Basically, we recommend to use 0.5-0.6 for fitting degree when you want to generate objects that are not constrained by the mask shape. If you want to generate objects that fit the mask shape, you can use 0.8-0.95 for fitting degree.
| Prompt | Fitting Degree | Input | Output |
|---|---|---|---|
| a bread | 0.95 |  |
 |
| a rabbit | 0.55 |  |
 |
| a rabbit | 0.95 |  |
 |
| a rabbit | 0.95 |  |
 |
-
Prepare training data. You may need to rewrite [
Datasets](./powerpaint/datasets/init.py）per your need (e.g., data and storage formats). Here, we use petreloss to read training dataset from cloud storages. Besides, the recipe of datasets for training a versatile model can be tricky but intuitive. -
Start training. We suggest using PowerPaint-V2 version, which is built upon BrushNet and requires smaller batch size for training. You can train it with the following command,
# running on a single node
accelerate launch --config_file configs/acc.yaml train_ppt2_bn.py --config configs/ppt2_bn.yaml --output_dir runs/ppt1_sd15
# running on one node by slurm, e.g., 1 nodes with 8 gpus in total
python submit.py --job-name ppt2_bn --gpus 8 train_ppt2_bn.py --config configs/ppt2_bn.yaml --output_dir runs/ppt2_bnwhere configs/acc.yaml is the configuration file for using accelerate, and configs/ppt2_bn.yaml is the configuration file for training PowerPaint-V2.
PowerPaint-V1 version often requires much larger training batch size to converge (e.g., 1024). You can train it with the following command,
# running on a single node
accelerate launch --config_file configs/acc.yaml train_ppt1_sd15.py --config configs/ppt1_sd15.yaml --output_dir runs/ppt1_sd15 --gradient_accumulation_steps 2 --train_batch_size 64
# running on two nodes by slurm, e.g., 2 nodes with 8 gpus in total
python submit.py --job-name ppt1_sd15 --gpus 16 train_ppt1_sd15.py --config configs/ppt1_sd15.yaml --output_dir runs/ppt1_sd15 --train_batch_size 64where configs/acc.yaml is the configuration file for using accelerate, and configs/ppt1_sd15.yaml is the configuration file for training PowerPaint-V1.
Junhao Zhuang: [email protected]
Yanhong Zeng: [email protected]
@misc{zhuang2023task,
title={A Task is Worth One Word: Learning with Task Prompts for High-Quality Versatile Image Inpainting},
author={Junhao Zhuang and Yanhong Zeng and Wenran Liu and Chun Yuan and Kai Chen},
year={2023},
eprint={2312.03594},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
