README.md

Free-Form Image Inpainting with Gated Convolution

Input

(Image from https://github.com/shepnerd/inpainting_gmcnn/blob/master/tensorflow/imgs/paris-streetview_256x256/001.png)

Shape : (1, 3, 256, 256)

Output

Left to right: input, mask, image generated by the network, ground truth

Results on Paris streetview with rectangle masks

Shape : (1, 3, 256, 256)

Results on CelebA with rectangle masks

Shape : (1, 3, 256, 256)

Results on CelebA with random strokes

Shape : (1, 3, 256, 256)

Results on Places with random strokes

Shape : (1, 3, 256, 256)

Usage

Automatically downloads the onnx and prototxt files on the first run. It is necessary to be connected to the Internet while downloading.

For the sample image,

$ python3 deepfillv2.py

If you want to specify the input image, put the image path after the --input option. You can use --savepath option to change the name of the output file to save.

$ python3 deepfillv2.py --input IMAGE_PATH --savepath SAVE_IMAGE_PATH

By adding the --model option, you can specify model type which is selected from "places", "celeba". (default is places) And it can specify mask type "rect" or "stroke" after the --mask_type option. (default is rect)

$ python3 deepfillv2.py --model places --mask_type rect

For CelebA with rectangle masks.

$ python3 deepfillv2.py --model celeba -i celebahq_256x256_001.png

For CelebA with random strokes.

$ python3 deepfillv2.py --model celeba --mask_type stroke -i celebahq_512x512_016.png

For Places with random strokes.

$ python3 deepfillv2.py --model places --mask_type stroke -i places2_001.png

Also if you want to use your own mask instead of the auto-generated one, you can specify your mask path by --mask_path.

$ python3 deepfillv2.py --model places --mask_path your_mask_path

Reference

Free-Form Image Inpainting with Gated Convolution

Framework

Pytorch

Model Format

ONNX opset=11

Netron

celeba_256x256.onnx.prototxt

places_256x256.onnx.prototxt

places_512x512.onnx.prototxt

places_1024x1024.onnx.prototxt