Auto-Inpainting

Recognize and remove needless foreground from photos in a smart way.

Installation

1. Clone this repo to your local filesystem with dependent submodules

   (base) $ git clone --recurse-submodules https://github.com/jaredyam/auto-inpainting.git

2. cd to the project directory and create a new virtual environment with conda

   (base) $ cd <path/to/auto-inpainting>
   (base) $ conda create --name auto-inpainting python=3.8
   (base) $ conda activate auto-inpainting
   (auto-inpainting) $ pip install -r requirements.txt

Inference

Before making inference, we need first prepare pre-trained models (from the original repo):

• segmentation model:
  • U2Net: download the pre-trained model u2net.pth (176.3 MB) from GoogleDrive or Baidu Wangpan (code: pf9k) and put it into the directory ./segmentation-models/U-2-Net/saved_models/u2net/.
• inpainting model:
  • LAMA: download the folder big-lama and put it in to the directory ./inpainting-models/lama/.

Then, make inference on a single image by running

(auto-inpainting) $ bash inference.sh <path/to/input-image>

Gallery

Original	Demo