Models

This guide discusses how models work in UpscalerJS.

Open example in Stackblitz.

Overview

UpscalerJS upscales images using neural networks trained on specific scales, configured in a variety of architecture sizes

info

By default, UpscalerJS uses default-model, which is available here.

UpscalerJS offers a number of models for download and installation, depending on the use case.

Installing

In this example, we'll use esrgan-thick, the most performant model. Install the model with:

npm install @upscalerjs/esrgan-thick

Code

We'll need to decide what scale model we wish to use. The larger the scale, generally the less accurate the resulting upscaled image will be.

View non-upscaled versions

View images all at same size

Original	2x Upscaled	3x Upscaled	4x Upscaled

We'll use the 2x scale model. We can import the specific model with:

import x2 from '@upscalerjs/esrgan-thick/2x'

And we can then pass the model as an argument to our upscaler:

import Upscaler from 'upscaler'
const upscaler = new Upscaler({
  model: x2,
})

The resulting image will be upscaled using the esrgan-thick model.

How Models are Loaded in the Browser

Tensorflow.js requires that models be available via a public URL. Therefore, UpscalerJS will attempt to load model files via a public CDN. The CDNs UpscalerJS loads from, by order of preference, include:

1. jsdelivr
2. unpkg

tip

If you want to avoid a CDN and self-host a model instead(or you've got a custom model) check out the guide on self hosting.