PConv-Keras VS lama

Recently I came across Partial Convolutions for Image Inpainting using Keras and spent a few days playing with the Jupyter notebooks included in the repo. I tried hard to reverse engineer everything in an effort to learn, but my limited experience with deep learning neural networks left me unable to accomplish anything.

https://www.fixmyphoto.ai/ has been deployed "on a serverless React application with AWS lambda functions handling inference" according to the github page.

You could try to use ControlNet inpaint+lama locally, but results aren't as good in my experience. Or you could try local install of lama directly, but the setup process isn't very smooth.

You may be able to remove the actor with lama. https://github.com/advimman/lama

LaMa: Resolution-robust Large Mask Inpainting with Fourier Convolutions (Apache-2.0 license) Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, Victor Lempitsky (Samsung Research and EPFL)

You could try using LaMA: https://github.com/saic-mdal/lama .

Task 1: 30$ Write a script to prepare data for https://github.com/saic-mdal/lama. You should be able to prove on video that your script works properly after i give you some data and you train and there's some acceptable output (not perfect, since i know how ml works).

I have some interest in image inpainting, but the recent paper introduced by Samsung - https://github.com/saic-mdal/lama seems like it reached a stage where there isn't much room that we can add for novelty. I am quite interested but not confident if I can come up with a novel idea that can do better than the proposed model, especially if they trained the model on multiple GPUs for days, which is something that I don't have.

Code: https://github.com/saic-mdal/lama