Long Exposures - Creating Average Frames From Movies

By on   3 comments 700 words, read ~7,953 times.

I read a Guardian article about an artist who set up an analogue camera in front of their TV, set it to long exposure, and set a movie playing. The result was a rather wonderful collection of images.

You can see more of Jason Shulman's works

Is there a way to automate this process? Yes!

Here's my attempt at a "Long Exposure" of "Taxi Driver".

A chaotic swirl of lights

It's not the same as Shulmans's process, but I think it is rather charming. Here's a timelapse of how it was constructed.

I experimented with several different different methods (code at the bottom of the page) - here are some of my other experimental results.

Taxi Driver

The Muppet Movie

I also managed to create something similar to the originals.

That's a mid-point of the movie, heavily blended with previous frames.



There's a couple of different ways to do this. These example work on Linux. Doing a digital average of all frames produces a different effect to Shulman's analogue process - but they are just as pleasing to my eye.

Naive and Slow

First, let's extract all the frames from a movie:

avconv -i movie.mkv -r 25 -ss 00:00:06 -t 01:23:45 %06d.png
  • -r 25 this extracts at 25 frames per second (for PAL movies). Use avconv -i movie.mkv to see which framerate you should use.
  • -ss 00:00:06 starts at 6 seconds in. This avoids the studio's logo.
  • -t 01:23:45 captures 1 hour, 23 minutes, 45 seconds. Useful to avoid the credits at the end.
  • %06d.png the extracted images will be numbered sequentially starting at 000000.png

To convert all those images to an average image, we use ImageMagick:

convert -limit memory 200MiB -limit map 400MiB *.png -average average.jpg

I've artificially limited the memory that the process can use. You can remove or adjust -limit memory 200MiB -limit map 400MiB depending on the speed of your system.

Fast and Pythony

We can use OpenCV to extract frames from the video. We can use PIL to average the difference between frames.

I tried several different options, but this produced the most pleasing results.

from PIL import Image
from PIL import ImageChops
from PIL import ImageEnhance
import cv2

#   Video to read
vidcap = cv2.VideoCapture('test.mkv')

#   For saved images
filename = "test"

#   Which frame to start from, how many frames to go through
start_frame = 600
frames = 180000

#   Counters
count = 0
save_seq = 0
first = True

while True:
   #   Read a frame
   success,image = vidcap.read()
   if not success:
   if count > start_frame+frames:
   if count >= start_frame:
      if first:
         #   Extract the frame and convert to image
         average_image = cv2.cvtColor(image,cv2.COLOR_BGR2RGBA)
         average_image = Image.fromarray(average_image)
         old_image = average_image
         first = False
      if (count%100 == 0):
         #   Every 100 frames (4 seconds @ 25fps)

         #   Extract the frame and convert to image
         image = cv2.cvtColor(image,cv2.COLOR_BGR2RGBA)
         image = Image.fromarray(image)

         #   Calculate the difference between this frame and the last
         diff = ImageChops.difference(image, old_image)

         #   Store the image for use in the next itteration
         old_image = image

         #   Convert to greyscale and use that as the alpha channel
         gray_image = diff.convert('L')

         #   Pick one!
         #average_image = Image.blend(average_image,image,0.1)
         #average_image = Image.alpha_composite(average_image,diff)
         average_image = ImageChops.lighter(average_image,diff)

         if (count%2500 == 0):
            #   Every 100 seconds (assuming 25fps)
            print("saving "+str(count))

            #   Darken the image slightly to prevent it getting washed out
            average_image = average_image.point(lambda p: p * 0.9)

            #   Show a preview of the image

            #   Save Image
            save_seq += 1

         if count == frames + start_frame:
         count += 1

#   Save the very last generated image

On my cheap and crappy laptop, a 90 minute movie took around 15 minutes to render.

To create a timelapse of the images, with two frames per second.

avconv -r 2 -i example%03d.png example.mp4

To crossfade the images in the timelapse

ffmpeg -framerate 2 -i example%03d.png -vf "framerate=fps=30:interp_start=64:interp_end=192:scene=100" example.mp4

Got a better / faster / more beautiful way to do it? Let me know in the comments!

Share this post onโ€ฆ

3 thoughts on “Long Exposures - Creating Average Frames From Movies

  1. mike says:

    I am definitely going to try this.

    You don't give a time for the Naive and Slow method. I assume it was slower than the Fast and Pythony method, but how much slower? ( I'm sort of hoping the Python method is a lot faster, to compensate for the time it took to write all that Python vs a couple of one liners ๐Ÿ˜‰ )

    Which method are the examples you show produced with?

    Is there any noticeable difference between the result of each method?

    I wonder, would the result if using only 1 image per second be perceptibly different to using 25 images for every second? Or using 2 images per second? Or N where N<25 and significantly reduces the time taken to render the image.

    1. Terence Eden says:

      Writing the movie to disk and then using ImageMagick was several times slower - it also took significant disk space and memory. Which my old laptop couldn't cope with.

      I used several different methods - blends, composites, averaging - try them all and see what you like the look of best.

      There wasn't any significant difference using every frame rather than every 25th - and it ran a lot quicker. Can't wait to see what you create with it ๐Ÿ™‚

  2. Thoc says:

    Hey Terence !

    Super cool stuff here ๐Ÿ™‚
    I'm currently trying and it look promising.

    (btw an indentation problem on the line count += 1 at the end ?)


What are your reckons?

All comments are moderated and may not be published immediately. Your email address will not be published.