lengstrom/fast-style-transfer

Combines perceptual loss optimization with instance normalization to enable real-time stylization at 100ms per frame on GPU, surpassing earlier neural style transfer methods. Supports both single images and video frame sequences through a trainable feed-forward transformation network, eliminating the need for iterative optimization per input. Built on TensorFlow with pre-trained style models available, integrating VGG19 features for content preservation while capturing artistic style characteristics.

10,975 stars. No commits in the last 6 months.