Practical overview
When it fits
Use EdgeConnect when you need a locally run, edge-guided workflow for reconstructing missing image regions and can provide images, masks and the documented machine-learning setup.
Who it is for
- Developers building image-restoration or photo-editing workflows.
- Researchers and practitioners who want to test or train an edge-guided image-inpainting model.
Documented uses
- Fill masked or damaged regions in an image.
- Test edge prediction, inpainting or joint model stages on images and masks.
- Train the model using image datasets and masks.
- Evaluate saved inpainting results with documented image-quality metrics.
Requirements & limits
- Testing requires an input image and a grayscale mask that covers the entire region to be filled.
- The README lists Python 3, PyTorch 1.0, an NVIDIA GPU and CUDA cuDNN.
- Pre-trained checkpoints are listed for Places2, CelebA and Paris Street-View.
- The public repository metadata does not provide a license.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Test image inpainting with a pre-trained checkpoint
Run the documented test workflow on images and corresponding grayscale masks.
Before you start
- Python 3, PyTorch 1.0, an NVIDIA GPU and CUDA cuDNN.
- The repository and its Python requirements installed.
- A configuration file in the checkpoint directory and a downloaded pre-trained checkpoint.
- Input image files and grayscale mask files that cover the complete missing regions.
Steps
Clone the repository and install PyTorch dependencies and the packages listed in requirements.txt.
Download a pre-trained model and place it under the checkpoints directory.
Create a config.yaml based on the example and place it in the checkpoint directory.
Run test.py with the checkpoint, input, mask and output paths; choose a model stage if needed.
Expected outcome
The model saves inpainted results in the specified output directory.
Guide limitations
- The README requires each mask to cover the entire masked region in its input image.
- The instructions describe local testing and do not establish a hosted service.
Train the edge model on Places2
Follow the documented setup and command for training the edge-model stage using Places2.
Before you start
- Python 3, PyTorch 1.0, an NVIDIA GPU and CUDA cuDNN.
- The repository, PyTorch dependencies and Python requirements installed.
- A Places2 training dataset and a configuration file based on the example.
Steps
Download the Places2 training images from the dataset's official website.
Use scripts/flist.py to generate a training file list for the downloaded images.
Create a config.yaml based on the example configuration and place it under the checkpoints directory.
Run the documented training command with model stage 1 and the Places2 checkpoint directory.
Expected outcome
The training command starts the edge-model stage using the specified checkpoint directory and dataset configuration.
Guide limitations
- The README says convergence time varies by dataset.
- Training the full model involves additional inpainting and joint stages not covered by this single-stage recipe.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 17 October 2026
- Original website title
- knazeri/edge-connect · Official homepage title · retained in its source language