lc-upscaler

lc-upscaler an upscaling model trained for the purpose of upscaling digital comics drawn in the style of ligne claire. It was trained on a custom dataset of pictures in that style. There is also an experimental denoiser that removes compression artifacts from the upscaled image. lc-upscaler was trained for a very specific use case (making digital comics look better on screens with a high pixel density), which is why it's not a good choice for a general-purpose comic upscaler (for that, see waifu2x). It works best with clean lines and simple colors (i.e. no gradients or textures).

The upscaling model is based on the ESPCN architecture by Shi et al. The denoiser is based on the U-Net architecture by Ronneberger et al.

See the notebooks in the notebooks directory for a starting point on how to train your own upscaler or denoiser.

Installation

Create a new Conda environment with the required dependencies:

$ conda create --name upscaler --file requirements.txt

Activate the environment

$ conda activate upscaler

If you do not want to create a Conda environment, you can install the following required packages manually (see version numbers in requirements.txt):

• click
• opencv-python
• pillow
• numpy
• tensorflow

Usage

$ python lc-upscaler.py <OPTIONS> image.jpg

With no additional options, the command above will output a 2 × upscaled PNG image named image_upscaled.png.

The following options are available:

• --scale FLOAT: The scale factor to use when upscaling the image. Valid range is 1.0-2.0. Default value is 2.0, i.e. 2 × upscaling.
• --compress: Compress the output using JPEG compression. If not specified, the output will be saved as a PNG image.
• --quality INT: The quality of the JPEG compression. Valid range is 1-100. Default value is 90. --compress must be specified for this option to have any effect.
• --denoise: Apply a denoising filter to the input image. This is an experimental feature and may not work well with all images.
• --suffix STRING: The suffix to use for the output filename. Default value is "_upscaled", i.e. image.png will be saved as image_upscaled.png.

Note The upscaling process may take a while, depending on the size of the input image. As an example, a 1500×2000 page takes about 15 seconds to upscale on my machine. If denoising is enabled, the process will take even longer. The same 1500×2000 page takes about 45 seconds to upscale with denoising enabled. Denoising is not always required, in fact, sometimes it may worsen the output quality.