Feed-Forward Style Transfer for the Impatient

What it does

This project implements a fast neural style transfer network in TensorFlow. Instead of iteratively optimizing each image from scratch, it trains a feed-forward CNN once; afterward, applying a style takes a single pass. The result is a system that can render a 1024×680 image in about 100 ms on a 2015 Titan X, or stylize entire videos frame by frame.

The interesting bit

The author tweaks the standard recipe in small but specific ways: swapping batch normalization for Ulyanov’s instance normalization, using VGG19 instead of VGG16, and pulling style features from shallower layers like relu1_1. The README claims this empirically produces larger-scale style features than Johnson’s original formulation.

Key highlights

• Single forward pass per image after training; no per-image gradient descent.
• Supports both still-image and video stylization via frame-by-frame processing.
• Training a new style takes roughly 4–6 hours on a Maxwell-era Titan X.
• Pre-trained evaluation models are available via Google Drive.
• The architecture blends work from Gatys, Johnson, and Ulyanov into one implementation.

Caveats

• Dependency stack is frozen in 2016: Python 2.7, TensorFlow 0.11.0, and specific legacy package versions.
• License is research-only by default; the author explicitly requires contact for any non-academic or commercial use.
• CPU inference is glacial at “several seconds per frame,” so a GPU is effectively mandatory.

Verdict

Worth a look if you are studying the evolution of style-transfer architectures or need a concrete, tweakable baseline from the feed-forward era. Skip it if you want a modern, production-ready pipeline with current dependencies and a permissive license.