UpscalerJS Models

Pre-trained models for use with UpscalerJS.

The models include:

Dataset	Scale	Example
DIV2K	2x
DIV2K	3x	)
DIV2K	4x	)

Sample image used for upscaling

Contributing

You'd like to contribute a new pretrained model? Awesome!

You can get a sense of the existing pretrained models by checking out the examples folder above. Each pretrained model will have an entry (for models at different scales, they'll usually have a single README since they were trained using the same parameters and dataset).

New models

To contribute a new pretrained model, you'll first need a model that runs in Javascript. Generally, that means that:

1. the model be trained in Tensorflow or be tensorflow-compatible
2. the model can be converted using the TFJS Converter (this means avoiding things like custom layers).
3. the model be quantized, if doing so does not lead to a drastic change in accuracy. Quantization helps performance in the browser. Try for the maximum quantization you can (8-bit).

Once you've converted a model, you'll want to follow this checklist:

• Open a PR against this library. Make sure to include:
  • Your model's model.json and weights in a folder within the models folder.
  • An update to this README's models table including your model and its scale.
  • A config.json file that has a description of your model. Some helpful things to include are how you trained your model, what dataset you used, and any hyperparameters you used.
  • An entry in the examples folder, copying the description from above and also including a sample image output for evaluation purposes.
  • Bump the version of package.json. Do a minor bump (aka, 0.1.1 -> 0.1.2)
  • Special kudos if you provide a one-click Colab or Dockerfile for reproducing your results.
• Open a PR against UpscalerJS
  • Add your model to the MODELS object so it can be loaded successfully.

Credits

All models are trained using image-super-resolution, an implementation of ESRGAN by @idealo.