Join us at TensorFlow World, Oct 28-31. Use code TF20 for 20% off select passes. Register now

Artistic Style Transfer with TensorFlow Lite

View on Run in Google Colab View source on GitHub Download notebook

One of the most exciting developments in deep learning to come out recently is artistic style transfer, or the ability to create a new image, known as a pastiche, based on two input images: one representing the artistic style and one representing the content.

Style transfer example

Using this technique, we can generate beautiful new artworks in a range of styles.

Style transfer example

This tutorial shows how to use a pre-trained TensorFlow Lite model to apply style transfer on any pair of content and style image. You can use the pre-trained model to add style transfer to your own mobile applications.

The model is open-sourced on GitHub. You can retrain the model with different parameters (e.g. increase content layers' weights to make the output image look more like the content image).

Understand the model architecture

Model Architecture

This Artistic Style Transfer model consists of two submodels: 1. Style Prediciton Model: A MobilenetV2-based neural network that takes an input style image to a 100-dimension style bottleneck vector. 1. Style Transform Model: A neural network that takes apply a style bottleneck vector to a content image and creates a stylized image.

If your app only needs to support a fixed set of style images, you can compute their style bottleneck vectors in advance, and exclude the Style Prediction Model from your app's binary.


Import dependencies.

from __future__ import absolute_import, division, print_function, unicode_literals
  # %tensorflow_version only exists in Colab.
  import tensorflow.compat.v2 as tf
except Exception:
import IPython.display as display

import matplotlib.pyplot as plt
import matplotlib as mpl
mpl.rcParams['figure.figsize'] = (12,12)
mpl.rcParams['axes.grid'] = False

import numpy as np
import time
import functools

Download the content and style images, and the pre-trained TensorFlow Lite models.

content_path = tf.keras.utils.get_file('belfry.jpg','')
style_path = tf.keras.utils.get_file('style23.jpg','')

style_predict_path = tf.keras.utils.get_file('style_predict.tflite', '')
style_transform_path = tf.keras.utils.get_file('style_transform.tflite', '')
Downloading data from
458752/458481 [==============================] - 0s 0us/step
Downloading data from
114688/108525 [===============================] - 0s 0us/step
Downloading data from
2834432/2827952 [==============================] - 0s 0us/step
Downloading data from
245760/244072 [==============================] - 0s 0us/step

Pre-process the inputs

  • The content image and the style image must be RGB images with pixel values being float32 numbers between [0..1].
  • The style image size must be (1, 256, 256, 3). We central crop the image and resize it.
  • The content image can be any size. However, as we trained the model using square-cropped data, cropping the content image to a square results in better stylized image.
# Function to load an image from a file, and add a batch dimension.
def load_img(path_to_img):
  img =
  img = tf.image.decode_image(img, channels=3)
  img = tf.image.convert_image_dtype(img, tf.float32)
  img = img[tf.newaxis, :]

  return img

# Function to pre-process style image input.
def preprocess_style_image(style_image):
  # Resize the image so that the shorter dimension becomes 256px.
  target_dim = 256
  shape = tf.cast(tf.shape(style_image)[1:-1], tf.float32)
  short_dim = min(shape)
  scale = target_dim / short_dim
  new_shape = tf.cast(shape * scale, tf.int32)
  style_image = tf.image.resize(style_image, new_shape)

  # Central crop the image.
  style_image = tf.image.resize_with_crop_or_pad(style_image, target_dim, target_dim)

  return style_image

# Function to pre-process content image input.
def preprocess_content_image(content_image):
  # Central crop the image.
  shape = tf.shape(content_image)[1:-1]
  short_dim = min(shape)
  content_image = tf.image.resize_with_crop_or_pad(content_image, short_dim, short_dim)

  return content_image

# Load the input images.
content_image = load_img(content_path)
style_image = load_img(style_path)

# Preprocess the input images.
preprocessed_content_image = preprocess_content_image(content_image)
preprocessed_style_image = preprocess_style_image(style_image)

print('Style Image Shape:', preprocessed_content_image.shape)
print('Content Image Shape:', preprocessed_style_image.shape)
Style Image Shape: (1, 849, 849, 3)
Content Image Shape: (1, 256, 256, 3)

Visualize the inputs

def imshow(image, title=None):
  if len(image.shape) > 3:
    image = tf.squeeze(image, axis=0)

  if title:

plt.subplot(1, 2, 1)
imshow(preprocessed_content_image, 'Content Image')

plt.subplot(1, 2, 2)
imshow(preprocessed_style_image, 'Style Image')


Run style transfer with TensorFlow Lite

Style prediction

# Function to run style prediction on preprocessed style image.
def run_style_predict(preprocessed_style_image):
  # Load the model.
  interpreter = tf.lite.Interpreter(model_path=style_predict_path)

  # Set model input.
  input_details = interpreter.get_input_details()
  interpreter.set_tensor(input_details[0]["index"], preprocessed_style_image)

  # Calculate style bottleneck.
  style_bottleneck = interpreter.tensor(

  return style_bottleneck

# Calculate style bottleneck for the preprocessed style image.
style_bottleneck = run_style_predict(preprocessed_style_image)
print('Style Bottleneck Shape:', style_bottleneck.shape)
Style Bottleneck Shape: (1, 1, 1, 100)

Style transform

# Run style transform on preprocessed style image
def run_style_transform(style_bottleneck, preprocessed_content_image):
  # Load the model.
  interpreter = tf.lite.Interpreter(model_path=style_transform_path)

  # Set model input.
  input_details = interpreter.get_input_details()

  # Set model inputs.
  interpreter.set_tensor(input_details[0]["index"], preprocessed_content_image)
  interpreter.set_tensor(input_details[1]["index"], style_bottleneck)

  # Transform content image.
  stylized_image = interpreter.tensor(

  return stylized_image

# Stylize the content image using the style bottleneck.
stylized_image = run_style_transform(style_bottleneck, preprocessed_content_image)

# Visualize the output.
imshow(stylized_image, 'Stylized Image')


Style blending

We can blend the style of content image into the stylized output, which in turn making the output look more like the content image.

# Calculate style bottleneck of the content image.
style_bottleneck_content = run_style_predict(
# Define content blending ratio between [0..1].
# 0.0: 0% style extracts from content image.
# 1.0: 100% style extracted from content image.
content_blending_ratio = 0.5 #@param {type:"slider", min:0, max:1, step:0.01}

# Blend the style bottleneck of style image and content image
style_bottleneck_blended = content_blending_ratio * style_bottleneck_content \
                           + (1 - content_blending_ratio) * style_bottleneck

# Stylize the content image using the style bottleneck.
stylized_image_blended = run_style_transform(style_bottleneck_blended,

# Visualize the output.
imshow(stylized_image_blended, 'Blended Stylized Image')