(Spoiler alert, despite the title, this post contains no obviously nude images. Sorry!)

This week I revisted a mystery that has been bugging me for several years. If you’re a regular reader of this blog you are probably familiar with my study emergent orange, a color that is produced when you make amalgam images by averaging the pixels together and then normalizing them. For example, if I smush a number of female headshots together, I get this:

I’ve written about this effect numerous times over the past decade, most recently in a paper for the 2015 Bridges conference.

When I was first experimenting with this, using Flickr images back in 2005, someone noticed the similarity of my process to that of artist Jason Salavon, who produced a number of lovely impressionistic nudes in 2002 by combining large numbers of Playboy centerfolds, producing one amalgam image for each decade from the 1960s thru the 1990s. As Salavon says, “This tracks, en masse, the evolution of this form of portraiture.”

What struck me about Jason’s images was the relative lack of orange. Based on my own experience with making amalgams of uncorrelated photos, I would have expected them to more closely resemble my headshot picture, and the others I had produced, all of which exhibit a distinct orange cast. I attempted to contact Jason a couple times over the years to ask him about this, but I never got a response back.

So this week, I was reminded of this mystery when I saw a post on BoingBoing alerting the Internet to a complete collection of Playboy centerfold images (NSFW) over on Imgur. So I wrote a little script to pull all the images, and then attempted to reproduce Salavon’s results.

Using my normal amalgam algorithm, as expected, the images are far more orange than Salavon’s.

However, by playing with the histograms, first in Photoshop, and then in code, I stumbled upon a simple technique which produces images which resemble Salavon’s, and is closer (though not identical) to what he did.

So why so little orange? Let’s use the 1980s image as an example. Here are the histograms of each color channel of my amalgamated image, which are representative of the overall colors of the complete image set, if added together.

As you can see there is a lot more energy in the red channel, about half the energy in the green channel and even less in the blue channel. This is why the image has a red/orange cast. You can see why by looking at the original centerfold images which have a lot of earth tones, a lot of warm lighting, and not a lot of blues and greens (I’ll leave that as an exercise to the reader).

I get an image much closer to Salavon’s if I individually stretch each red/green/blue channel to the limit on both ends, normalizing each channel individually; a process I’ll call “equalization” because it forces the contribution of each channel to be roughly equal. I believe this is what Salavon means by “normalization” in his own descriptions of his process.

You can see the histograms of Salavon’s actual 1980s image on the right. As you can see, he has indeed manipulated each color channel separately, doing something much closer to equalization than a true normalized image, although he’s reduced the blue levels a bit. Its certainly possible these levels were tweaked by hand in Photoshop.

You’ll also notice differences in the curve outlines (such as the prominent bump on the low end of his blues). I believe this is due to his use of a different set of scans, from equipment with different color response.

While Salavon’s channel tweaking produces a more artistically satisfying image, it does not accurately represent the colors, and particularly the hues of the component centerfold images. In particular, it greatly amplifies the contribution of blue and green, creating the impression that the backgrounds are mostly blue, or blue-green, which they simply aren’t.

Salavon’s earliest Playboy image, made in 1998 does indeed exhibit a more pronounced orange shift, and was likely not color-corrected in the way his 2002 work was. The Salavon original shown below, on the left, contains images from 1988 through 1997 (a 12 year spread) and is closer in hue to my own unequalized reconstruction, shown in the middle. If I equalize just the red channel, as shown on the right, the resulting image is somewhat closer to Salavon’s.

Using the 3-channel equalization technique, I was able to produce a full set of Salavon-style amalgam images, extending back to December 1953, when the magazine was first published, all the way up thru the month before this post, February 2016. The similarity between the 1990s and 2000s image are quite striking.

Many thanks to Jason Salavon and the Flickr Community! You’ll find sample code for making amalgam images in Python in my Github repo.

This entry was posted on Wednesday, March 2nd, 2016 at 11:56 am and is filed under Jim's Projects. You can follow any responses to this entry through the RSS 2.0 feed. Both comments and pings are currently closed.