Er0manga / Er0mangaInpaint

🦙 Manga decensoring with LaMa inpainting

Er0manga censorship inpainting

This repo contains training and testing code for the manga censorship inpainting and is based on LaMa.

Quick start:

Step 0: Install anaconda3/miniconda3 and create the environment:

You can skip this step if you already have the environment from Er0mangaSeg. Both Segmentation and Inpainting repo use the same environment. If you don't have the environment, follow the instruction from Step 0 in Er0mangaSeg.

Step 1: Download the pretrained model:

Download pretrained LaMa model from https://mega.nz/file/gRYn2CBJ#HZ73lqn5noX_t2eyfIaDk7sIDfnGQ9gBwClJ6O3VdTE and put it into the pretrained directory. This model is finetuned by the author on uncensored manga dataset, so it performes better on screentoned images.

Step 2: Run the demo:

Run the model on the demo image located in demo_images:

PYTHONPATH=$PWD TORCH_HOME=$PWD python bin/uncen.py --in_dir=./demo_images/demo_input/ --mask_dir=./demo_images/demo_mask/ --out_dir=./output --debug_dir=./output_dbg --checkpoint=./pretrained/00-30-09

The result will be located in ./output and the debug output will be located in ./output_dbg.

The debug output should look like this:

Usage:

Test the model:

PYTHONPATH=$PWD TORCH_HOME=$PWD python bin/uncen.py --in_dir=<input dir with png images> --mask_dir=<dir with corresponding inpainting masks> --out_dir=<output directory> --debug_dir=<optional debug output directory> --checkpoint=<directory with checkpoint, must contain models/best.ckpt and config.yaml>

Train the model:

Modify configs/training/hconfig.yaml to match you configuration (dataset location, gpus used, batch size, checkpoint, etc...) and run:

PYTHONPATH=$PWD TORCH_HOME=$PWD python3 bin/train.py -cn hconfig

Dataset format:

Same as in Er0mangaSeg.

TODO: add dataset collection scripts