Awesome

ColorizationWithRegion

The code for pg2021 paper "Line Art Colorization Based on Explicit Region Segmentation"

This is a simple implementation for comparison with Tag2Pix (code / paper).

Overall project refactoring and further optimization may be later.

Usage

1. Build the environment and dataset according to Tag2Pix.

2. Use python code/skeleton/line_art2skeleton.py <line art folder> to create skeleton maps.

   See DanbooRegion for environment.

   Some code and pretrained model are from DanbooRegion.

   For each line art folders, e.g., keras_train, xdog_train, keras_test or others, create a corresponding folder to place skeleton maps, like keras_train_skeleton and others.

3. Replace loader/dataloader.py of Tag2Pix with code/loader/dataloader.py to load skeleton maps.

   We also remove the random_jitter for visible test results while training.

4. For dual-branch, replace network.py and tag2pix.py of Tag2Pix with files in code/dual_branch.

5. For direct concatenation, replace network.py and tag2pix.py of Tag2Pix with files in code/direct.

6. Train the model as Tag2Pix.