GPNN

Generative Patch-Nearest-Neighbor

Pytorch implementation of the paper: "Drop the GAN: In Defense of Patches Nearest Neighbors as Single Image Generative Models"

Random Sample from a Single Image

With GPNN, you can generate a random sample from a given image in few seconds. For example:

GPNN's Other Applications

GPNN can perform many other tasks, such as image generation, conditional inpainting, structural analogies, image retargeting, collage, and more. Currently, this implementation supports only the first three tasks.

Code

Install Dependencies

python -m pip install -r requirements.txt

This code will run on cuda gpu if available. Running on cpu is by specifying '--not_cuda'.

Memory and Speed

If running this code on your machine is too exhausting, or you want to get quick results, you may:

• Generate smaller image by specifying '--out_size <int>'
• Use very fast approximate-nearest-neighbor method (faiss), by specifying '--faiss'. Install faiss by the following:

python -m pip install faiss-gpu

Notice that this method is different from the normalized distance matrix presented in the original paper.

Random Sample

To generate a random sample, run:

python random_sample.py -in <image_path>

You may control the variation degree of the new sample by adjusting the noise level '--sigma'. Default is 0.75.

Structural Analogies

To generate a new image where the content of one image is constructed into the structure of another image, run:

python structural_analogies.py -a <first_image_path> -b <second_image_path>

where the first image is the content and the second is the structure.

Inpainting

To generate color-guided inpainting recovery, use an inpainted image where the inpainted area color is similar to your target recovery color (in the same image):

python inpainting.py -in <image_path> -m <mask_path>

the mask is in the following format - ones in the pixels where the inpainted area is, and zeros elsewhere.

Generated images will be saved in './output' folder. To change that, specify '-out <dir_path>'.

In each of the tasks, there are many parameters to adjust. A full list may be obtained by specifying '--help'.

This repository borrows the images database from the SinGAN repository.