Style Transfer for Time Lapse Photography

By Robert Ellison. Updated on Saturday, February 19, 2022.

TensorFlow Hub has a great sample for transferring the style of one image to another. You might have seen Munch's The Scream applied to a turtle, or Hokusai's The Great Wave off Kanagawa to the Golden Gate Bridge. It's fun to play with and I wondered how it would work for a timelapse video. I just posted my first attempt, four different shots of San Francisco and I think it turned out pretty well.

The four sequences were all shot on an A7C, one second exposure, ND3 filter and aperture/ISO as needed to hit one thousandth of a second before fitting the filter. Here's an example shot:

I didn't want The Scream or The Wave, so I asked my kids to pick two random pieces of art each so I could have a different style for each sequence:

The style transfer network wants a 256x256 style image so I cropped and resized the art as needed.

The sample code pulls images from URLs. I modified it to connect to Google Drive, iterate through a source folder of images and write the transformed images to a different folder. I'm running this in Google Colab which has the advantage that you get to use Google's GPUs and the disadvantage that it will disconnect, timeout, run out of memory etc. To work around this the modified code can be run as many times as needed to get through all the files and will only process input images that don't already exist in the output folder. Here's a gist of my adapted colab notebook:

{ "nbformat": 4, "nbformat_minor": 0, "metadata": { "colab": { "name": "Style Transfer for Timelapse.ipynb", "provenance": [], "authorship_tag": "ABX9TyNSFwNMHz6Mtpw3zLReIFU+", "include_colab_link": true }, "kernelspec": { "name": "python3", "display_name": "Python 3" }, "accelerator": "GPU" }, "cells": [ { "cell_type": "markdown", "metadata": { "id": "view-in-github", "colab_type": "text" }, "source": [ "<a href=\"https://colab.research.google.com/gist/abfo/c2fe8af9976ae09a324f2cc410a86008/style-transfer-for-timelapse.ipynb\" target=\"_parent\"><img src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/></a>" ] }, { "cell_type": "markdown", "metadata": { "id": "HEHPFmEEV7Yk" }, "source": [ "Apply style transfer to a set of images in Google Drive. Based on the style transfer guide [here](https://www.tensorflow.org/hub/tutorials/tf2_arbitrary_image_stylization) and the drive example [here](https://colab.research.google.com/notebooks/io.ipynb#scrollTo=RWSJpsyKqHjH). This code expects a folder in the root of your drive called stylegan. The folder should contain two subfolders, input (images to style), output and a style image called style.jpg. This should be 256x256 pixels. The sample code works with square images and will take the center of the source image if in a different aspect ratio. For my purposes I'm using 4k frames and then setting the output size to 1920x1920 in order to crop a 1920x1080 HD frame from the processed impages.:" ] }, { "cell_type": "code", "metadata": { "id": "CUGH5d1o5jPh" }, "source": [ "content_img_size = (1920, 1920)\r\n", "style_img_size = (256, 256) " ], "execution_count": null, "outputs": [] }, { "cell_type": "markdown", "metadata": { "id": "rcj_QQd75oMH" }, "source": [ "Set up tensorflow and some support functions:" ] }, { "cell_type": "code", "metadata": { "id": "4Qtbh0rXFcCd" }, "source": [ "# Copyright 2019 The TensorFlow Hub Authors. All Rights Reserved.\r\n", "#\r\n", "# Licensed under the Apache License, Version 2.0 (the \"License\");\r\n", "# you may not use this file except in compliance with the License.\r\n", "# You may obtain a copy of the License at\r\n", "#\r\n", "# http://www.apache.org/licenses/LICENSE-2.0\r\n", "#\r\n", "# Unless required by applicable law or agreed to in writing, software\r\n", "# distributed under the License is distributed on an \"AS IS\" BASIS,\r\n", "# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\r\n", "# See the License for the specific language governing permissions and\r\n", "# limitations under the License.\r\n", "# ==============================================================================\r\n", "\r\n", "from google.colab import drive\r\n", "from google.colab import files\r\n", "import functools\r\n", "import os\r\n", "from matplotlib import gridspec\r\n", "import matplotlib.pylab as plt\r\n", "import numpy as np\r\n", "import tensorflow as tf\r\n", "import tensorflow_hub as hub\r\n", "import os\r\n", "import PIL.Image\r\n", "\r\n", "print(\"TF Version: \", tf.__version__)\r\n", "print(\"TF-Hub version: \", hub.__version__)\r\n", "print(\"Eager mode enabled: \", tf.executing_eagerly())\r\n", "print(\"GPU available: \", tf.test.is_gpu_available())\r\n", "\r\n", "def crop_center(image):\r\n", " \"\"\"Returns a cropped square image.\"\"\"\r\n", " shape = image.shape\r\n", " new_shape = min(shape[1], shape[2])\r\n", " offset_y = max(shape[1] - shape[2], 0) // 2\r\n", " offset_x = max(shape[2] - shape[1], 0) // 2\r\n", " image = tf.image.crop_to_bounding_box(\r\n", " image, offset_y, offset_x, new_shape, new_shape)\r\n", " return image\r\n", "\r\n", "@functools.lru_cache(maxsize=None)\r\n", "def load_image(image_url, image_size=(256, 256), preserve_aspect_ratio=True):\r\n", " \"\"\"Loads and preprocesses images.\"\"\"\r\n", " # Cache image file locally.\r\n", " #image_path = tf.keras.utils.get_file(os.path.basename(image_url)[-128:], image_url)\r\n", " image_path = image_url\r\n", " # Load and convert to float32 numpy array, add batch dimension, and normalize to range [0, 1].\r\n", " img = plt.imread(image_path).astype(np.float32)[np.newaxis, ...]\r\n", " if img.max() > 1.0:\r\n", " img = img / 255.\r\n", " if len(img.shape) == 3:\r\n", " img = tf.stack([img, img, img], axis=-1)\r\n", " img = crop_center(img)\r\n", " img = tf.image.resize(img, image_size, preserve_aspect_ratio=True)\r\n", " return img\r\n", "\r\n", "def show_n(images, titles=('',)):\r\n", " n = len(images)\r\n", " image_sizes = [image.shape[1] for image in images]\r\n", " w = (image_sizes[0] * 6) // 320\r\n", " plt.figure(figsize=(w * n, w))\r\n", " gs = gridspec.GridSpec(1, n, width_ratios=image_sizes)\r\n", " for i in range(n):\r\n", " plt.subplot(gs[i])\r\n", " plt.imshow(images[i][0], aspect='equal')\r\n", " plt.axis('off')\r\n", " plt.title(titles[i] if len(titles) > i else '')\r\n", " plt.show()\r\n", "\r\n", "def tensor_to_image(tensor):\r\n", " tensor = tensor*255\r\n", " tensor = np.array(tensor, dtype=np.uint8)\r\n", " if np.ndim(tensor)>3:\r\n", " assert tensor.shape[0] == 1\r\n", " tensor = tensor[0]\r\n", " return PIL.Image.fromarray(tensor)" ], "execution_count": null, "outputs": [] }, { "cell_type": "markdown", "metadata": { "id": "bZwqFmPM43tQ" }, "source": [ "Load the hub module, this will take a few minutes" ] }, { "cell_type": "code", "metadata": { "id": "V5KZbbG-47hk" }, "source": [ "hub_handle = 'https://tfhub.dev/google/magenta/arbitrary-image-stylization-v1-256/2'\r\n", "hub_module = hub.load(hub_handle)" ], "execution_count": null, "outputs": [] }, { "cell_type": "markdown", "metadata": { "id": "NM-z5lKXGBQf" }, "source": [ "Open drive (you'll need to allow access each session, and load the style image)" ] }, { "cell_type": "code", "metadata": { "id": "T9759bbNIp8l" }, "source": [ "drive.mount('/content/drive', force_remount=True)" ], "execution_count": null, "outputs": [] }, { "cell_type": "code", "metadata": { "id": "7QWqAgulWd3Z" }, "source": [ "style_image = load_image('/content/drive/MyDrive/stylegan/style.jpg', style_img_size)\r\n", "style_image = tf.nn.avg_pool(style_image, ksize=[3,3], strides=[1,1], padding='SAME')\r\n", "show_n([style_image], ['Style image'])" ], "execution_count": null, "outputs": [] }, { "cell_type": "markdown", "metadata": { "id": "GAyE0dNiD44E" }, "source": [ "Process all the files... This might take a while and sometimes fails so it won't process if the output file alredy exists. You can run this cell as many times as needed to process all the input files (if you get out of memory errors try restarting the runtime and running the cells above again). This will run much faster if you can grab a GPU instance. If you change anything (like the style image) you'll want to clear the output folder and run again. " ] }, { "cell_type": "code", "metadata": { "id": "GK9ifHT4D_s5" }, "source": [ "directory = '/content/drive/MyDrive/stylegan/input'\r\n", "outdirectory = '/content/drive/MyDrive/stylegan/output'\r\n", "for file in os.listdir(os.fsencode(directory)):\r\n", " filename = os.fsdecode(file)\r\n", " outputfile = os.path.join(outdirectory, filename)\r\n", " if (os.path.isfile(outputfile)):\r\n", " continue\r\n", " print(filename)\r\n", " content_image = load_image(os.path.join(directory, filename), content_img_size, preserve_aspect_ratio=True)\r\n", " outputs = hub_module(tf.constant(content_image), tf.constant(style_image))\r\n", " stylized_image = outputs[0]\r\n", " tensor_to_image(stylized_image).save(outputfile)\r\n", "\r\n", "#make sure files are saved when we're done\r\n", "print('Done, disconnecting from drive...')\r\n", "drive.flush_and_unmount()" ], "execution_count": null, "outputs": [] } ] }

One final problem is that the style transfer example produces square output images. I just set the output to 1920x1920 and then cropped a HD frame out the middle of each image to get around this.

Here's a more detailed workflow for the project:

I usually shoot timelapse with a neutral density filter to get some nice motion blur. When I shot this sequence it was the first time I'd used my filter on a new camera/lens and screwing in the filter threw off the focus enough to ruin the shots. Lesson learned - on this camera I need to nail the focus after attaching the filter. As I've been meaning to try style transfer for timelapse I decided to use this slightly bad sequence as the input. Generally for timelapse I shoot manual / manual focus, fairly wide aperture and ISO 100 unless I need to bump this up a bit to get to a 1 second exposure with the filter.

After shooting I use LRTimelapse and Lightroom 6 to edit the RAW photos. LRTimelapse reduces flicker and works well for applying a pan and/or zoom during processing as well. For this project I edited before applying the style transfer. The style transfer network preserves detail very well and then goes crazy in areas like the sky. Rather than zooming into those artifacts I wanted to keep them constant which I think gives a greater sense of depth as you zoom in or out.

Once the sequence is exported from Lightroom I cancel out of the LRTimelapse render window and switch to Google Colab. Copy the rendered sequence to the input folder and the desired style image and then run the notebook to process. If it misbehaves then Runtime -> Restart and run all is your friend.

To get to video I use ffmpeg to render each sequence. For this project at 24 frames per second and cropping a 1920x1080 frame from each of the 1920x1920 style transfer images.

Then DaVinci Resolve to edit the sequences together. I added a 2 second cross dissolve between each sequence and a small fade at the very beginning and end.

Finally, music. I use Filmstro Pro and for this video I used the track Durian.

Here's the result:

Add your comment...

(All Code Posts)

(Published to the Fediverse as: Style Transfer for Time Lapse Photography #code #ml #tensorflow #drive #python #video How to apply the TensorFlow Hub style transfer to every frame in a timelapse video using Python and Google Drive. )

code, ml, tensorflow, drive, python, video