Browser neural style transfer, circa 2017

What it does This repo runs Fast Neural Style Transfer entirely in the browser using Deeplearn.js (now TensorFlow.js). You pick a pre-trained style, feed it an image or webcam feed, and the network redraws your content in that style. Each style model weighs about 6.6MB and downloads to your machine; nothing gets uploaded to a server.

The interesting bit The author is admirably blunt about the project’s limits: the UI is ugly (“10 minute tolerance for tweaking HTML and CSS”), and adding a new style requires training a separate network from scratch on a Maxwell Titan X for 4–6 hours. The real value now is historical—this code was absorbed into ml5.js, and the author points to a newer repo that handles arbitrary styles without per-style retraining.

Key highlights

• Runs inference locally in the browser; no server round-trips for your images
• ~6.6MB per style model—bandwidth-conscious on mobile
• Build chain uses Bower and a custom watch script (period-appropriate, now broken)
• Includes Python utilities to convert TensorFlow checkpoints to the Deeplearn.js format
• Author explicitly deprecated this in favor of arbitrary-image-stylization-tfjs

Caveats

• Build instructions no longer work; the repo is unmaintained
• Per-style training requirement makes it inflexible compared to modern approaches

Verdict Worth a look if you’re studying the evolution of in-browser ML or maintaining legacy ml5.js integrations. Skip it if you actually need style transfer today—use the author’s newer project or TensorFlow.js directly.