This fork ports features over from my ESRGAN-Bot repository and adds a few more. It natively allows:

• In-memory splitting/merging functionality (fully seamless, recently revamped for the third time and no longer requires tile size)
• Seamless texture preservation (both tiled and mirrored)
• Model chaining
• Transparency preservation (3 different modes)
• 1-bit transparency support (with half transparency as well)
• Both new-arch and old-arch models
• Regular and SPSR models, of any scale or internal model settings
• On-the-fly interpolation

Tile size was recently removed! It is no longer needed for split/merge functionality!

To set your textures to seamless, use the --seamless flag. For regular tiled seamless mode, use --seamless tile. For mirrored seamless mode, use --seamless mirror. You can also add pixel-replication padding using --seamless replicate and alpha padding using --seamless alpha_pad.

To chain models, simply put one model name after another with a > in between (you can also use + if using bash to avoid issues), such as 1xDeJpeg.pth>4xESRGAN.pth note: To use model chaining, model names must be the complete full name without the path included, and the models must be present in your /models folder. You can still use full model paths to upscale with a single model.

For on-the-fly interpolation, you use this syntax: <model1_name>:<##>&<model2_name>:<##>, where the model name is the path to the model and ## is the numerical percentage to interpolate by. For example, model1:50&model2:50 would interpolate model1 and model2 by 50 each. The numbers should add up to 100. If you have trouble using : or &, either try putting the interpolation string in quotes or use @ or | respectively ("model1@50|model2@50").

To use 1 bit binary alpha transparency, set the --binary-alpha flag to True. When using --binary-alpha transparency, provide the optional --alpha-threshold to specify the alpha transparency threshold. 1 bit binary transparency is useful when upscaling images that require that the end result has 1 bit transparency, e.g. PSX games. If you want to include half transparency, use --ternary-alpha instead, which allows you to set the --alpha-boundary-offset threshold.

The default alpha mode is now 0 (ignore alpha). There are also now 3 other modes to choose from:

• --alpha-mode 1: Fills the alpha channel with both white and black and extracts the difference from each result.
• --alpha-mode 2: Upscales the alpha channel by itself, as a fake 3 channel image (The IEU way) then combines with result.
• --alpha-mode 3: Shifts the channels so that it upscales the alpha channel along with other regular channels then combines with result.

To process images in reverse order, use --reverse. If needed, you can also skip existing files by using --skip-existing.

If you're upscaling images of the same size, you can do --cache-max-split-depth to only calculate the automatic tile size once to improve performance.

Examples:

• python upscale.py 4xBox.pth --seamless tile
• python upscale.py 1xSSAntiAlias9x.pth>4xBox.pth
• python upscale.py 4xBox.pth --binary-alpha --alpha-threshold .2
• python upscale.py /models/4xBox.pth
• python upscale.py "1x_model1.pth@50|1x_model2.pth@50>2x_model3.pth"

If you want a GUI for ESRGAN and if you're on Windows, check out Cupscale. It implements most of this fork's features as well as other utilities around it.