Before DeepCreamPy can be used, the user must color regions in the artwork using green color with an image editing program (e.g. GIMP, Photoshop). DeepCreamPy takes the images with green colored regions as input, and a neural network automatically fills in the highlighted regions.

You can download the latest release for Windows 64-bit here.

For users interested in compiling DeepCreamPy themselves, DeepCreamPy can run on Windows, Mac, and Linux.

Please before you open a new issue check closed issues and check the table of contents.

Features

Reconstructing images of any size
Reconstruction of ANY shaped censor (e.g. black lines, pink hearts, etc.)
Decensoring of mosaic censors
Limited support for black and white/monochrome images
Generate multiple variations of reconstructions from the same image

Limitations

The reconstruction is mainly for anime-style human figures that have minor to moderate redactions. If an organ (e.g. arms, legs) is completely deleted, reconstruction will fail.

It does NOT work with:

Screentones (e.g. printed material)
Real life material
Reconstruction of nipples
Reconstruction of lower orifice of the alimentary canal
Animated gifs and videos

Setup:

Usage:

Miscellaneous:

To do

Moving to PyTorch or newer versions of Tensorflow
Improving UI
Error logging

Contributions

We're open for contributions as long as your contribution complies with GNU Affero General Public License v3.0 and be advised of GitHub inbound=outbound rule.

For contributions you used to sign Contributor License Agreement (the "CLA") but it's no longer the case for now.

This project was initially created by deeppomf and all credit goes to them. Special thanks to ccppoo, IAmTheRedSpy, 0xb8, deniszh, Smethan, harjitmoe, itsVale, StartleStars, SoftArmpit and everyone else for their contributions!

License

Source code and official releases/binaries are distributed under the GNU Affero General Public License v3.0.

Acknowledgements

Example mermaid image by Shurajo & AVALANCHE Game Studio under CC BY 3.0 License. The example image is modified from the original, which can be found here.

Neural network code is modified from Forty-lock's project PEPSI, which is the official implementation of the paper PEPSI : Fast Image Inpainting With Parallel Decoding Network. PEPSI is licensed under the MIT license.

Training data is modified from gwern's project Danbooru2017: A Large-Scale Crowdsourced and Tagged Anime Illustration Dataset and other sources.

See ACKNOWLEDGEMENTS.md for full license text of these projects.

README.md

DeepCreamPy