Order of Multitudes

Pixelating William Blake

Sarah T. Weston is a PhD candidate in the English and History of Art Departments at Yale University. She specializes in literature and art of the eighteenth and nineteenth centuries, with a particular interest in William Blake, the history of science, and media studies. She has held a Digital Humanities Fellowship and serves as primary investigator on the Yale DHLab’s BlakeTint project.

Over the past few years, the Yale DHLab and I have been collaborating on a digital humanities project entitled BlakeTint—an intensive study of the English Romantic poet and painter William Blake (1757-1827). The dynamic website allows users to examine the intricacies of Blake’s approach to color in his magnificent “illuminated books.” Blake invented a wonderfully intricate artisanal process (which he dubbed his “infernal method”), whereby he illustrated his poetry and etched his hybrid designs on copper plates. He printed and reprinted these books throughout his lifetime, hand-coloring each copy—imbuing each iteration of the book with a new aesthetic in each re-print. Blake’s unique corpus provides us with a record of how one man altered his opinions, his politics, his convictions, his personal mythologies over the course of a lifetime: returning to the same image, the same text, and making it new each time.

Many artists and philosophers systematized color in the Romantic period, yet Blake—perhaps one of the most interesting color-theorists—remains the least studied. BlakeTint allows us to see Blake’s system-building from several novel perspectives and to trace a greater systematic approach within Blake’s methodology.

The project also asks what the digital sphere can do to offer new modes of looking, organizing, and understanding art. So, too, must we consider what damage algorithms do when processing images at large scales, like this—what minute details and beauties are lost in the computer’s cold eye? How do we best test the precision of an algorithm and find one that most accurately captures the original work of art? How accurate is computer vision? BlakeTint offers a method of information management that allows us to see Blake’s works from an entirely new register. The project’s methodology can also be replicated with other datasets, and we hope other scholars will use it to analyze their own collections of art.

BlakeTint brings together multiple resources across campus—from the Yale Center for British Art’s vast collections of works by Blake (one of the foremost repositories of Blake’s works in the United States) to the Yale DHLab’s brilliant staff, who created this gorgeous tool. Resting as it does at a crux between book history and computer science, humanities and chemistry, BlakeTint showcases Yale’s diverse holdings and specialties, networking the inimitable collections and exceptional resources singular to Yale.

The tool studies the progression of Blake’s color theory from his earliest copies (1789) to those made in the year of his death (1827). Using a process called “color quantization,” we used three different algorithms (K-Means, Median Cut, and Octree) to extract the top twenty colors used in each of Blake’s plates—rendering color palettes for each image. We had each algorithm reconstruct the original image, using the color palettes it had extracted. In rigorously combing through hundreds of reconstructions, we were able to spot which of the three algorithms most precisely extracted colors from the original. To illustrate the comparative work we were doing, the following image places, side-by-side, the three algorithmic reconstructions of Blake’s print, “Infant Joy,” from a copy of Songs of Innocence and of Experience (Copy L, 1795) housed at Yale.

Median Cut registers the pink in the blossom on the right and the blue on the left side of the print—moments which the other two algorithms fail to perceive. The “K-Means” algorithm (middle), however, perhaps renders the red of the main blossom a little truer to the original.

For its certain failings at times, Median Cut did steadily seem to provide the greatest accuracy—at times even picking up on subtleties of Blake’s brushwork. The following image zooms in on a small detail of “Infant Joy,” contrasting the original image with the three algorithmic reconstructions:

Blake’s original brushwork is on the top left. Through close observation, Median Cut (bottom left) is most able to capture Blake’s brushwork, where K-Means and Octree barely register this small moment, finding little to no tonal variation.

Once we had settled on Median Cut as our algorithm of choice, we created two color palettes for each of Blake’s plates: one “equal” and one “proportional.” The following image shows the original image (the poem, “The Tyger” from the same copy of Songs) converted into two color bars.

On top, the equal color bar displays the twenty extracted colors, equally. Right below, the proportional color bar displays the ratio of each color on the plate—showing not only which colors are on the plate, but how much there is of each.

The proportional color bars are collated into a vertical array—each book, represented as a row of bars. Below, the color bar for “The Tyger” can be seen in its place within Songs Copy L.

BlakeTint allows visitors to sort Blake’s books “by year” or “by book,” studying his approach to color from whichever angles feel most fruitful to their eye. The project also allows users to inspect the algorithmic reconstructions of each plate—glimpsing Blake’s works as the computer sees them.

The experience of translating Blake’s “copy theory” into the digital world has been fascinating and difficult. One concern we have repeatedly had to negotiate with is the idea of the “digital copy”—how the inaccuracies of digital reproductions of works of art remediate the original work of art. Many of the digitized images available of Blake’s works are not color-corrected, were photographed using different color-scales, or are a decade old. So, for all of our fretting about the accuracies of the algorithms we were using (whether they truthfully captured the original image), the original images we were feeding the algorithms were, themselves, flawed and inaccurate.

Working on BlakeTint made me urgently interested in what gets “lost in translation” in digital reproductions. With a seed grant from the Order of Multitudes Sawyer Seminar, I was able to complete hands-on archival research to check the accuracy of our digital humanities tool—analyzing Blake’s color theory not only from a digital distance but also in person. I spent time in several collections in the United States and United Kingdom, assessing how true-to-life the baseline digital images that we fed into our algorithm were.

In my archival adventures, I studied Blake’s original saturation and tone against the images we used, finding quite a few copies that truly felt inaccurately rendered in digital reproduction. In person, for example, the Huntington Library’s Copy E of Songs of Innocence and of Experience was much more vibrant than it appears to be online. What looks, on the digital reproduction of the title page for “Songs of Innocence,” like a watery light yellow is actually a pulsating, almost greenish-hue in person. Other books, such as Copy H of The Marriage of Heaven and Hell (1790), housed at the Fitzwilliam Museum in Cambridge, UK, threw into stark relief the challenges of digitally rendering gold leaf. The book flutters with streaks and flakes of gold, glittering magically in one’s hands, changing its aspect with the slightest shift in angle of perspective. These transcendent effects render quite dully on the 2D digital page.

This research trip not only allowed me to check the accuracy of our project, it also sparked reflection about how seeing works of art through the “lens” of digital tools can mediate our experience of the works of art in person. After spending so much time with and analyzing the truncated color-bar versions of Blake’s books, the experience of seeing books in person was a truly interesting experience. I found my eye gravitating to aspects of his plates that I had not before—seeing his books with new eyes. When I first saw BlakeTint’s abbreviated versions of Blake’s books, I was astonished by patterns I found that I had never fully grasped in the archive: a period of gravitation to primary colors (red, yellow, blue), then a distinct period of gravitation to complementary colors (especially blue and orange). Viewing Blake’s books in person, again, I was surprised that I had not noticed these trends before. For example, while I was looking at Songs Copy E, I was preoccupied with Blake’s use of primary colors in his water-coloring (see: “Introduction” or “The Blossom”)—my eye likely missing details that might have caught my attention before, when it would not have necessarily registered anything terribly meaningful in the presence of the primary colors on the page. It was truly interesting to see the way in which raw data, algorithms, and “the machine” retools how we look at works in the archive—how seeing works of art through a computer’s vision re-tweaks our own.

BlakeTint reconfigures our existing ways of approaching art historical and museum studies, offering a way of managing information in bulk and large datasets (e.g. all of Blake’s plates across all of his illuminated books), collating many copies from multiple collections and putting them together—seeing what we can understand about Blake’s works when they are taken together as a multitude: condensed and massed together, brought into conversation across distances. Yet the experience of seeing these books in person, as well as observing their storage within the archive, raised some interesting questions about how we can perform meaningful research on museum collections stored across multiple museums. Each museum archive utilizes different methods of cataloguing, storing, and photographing works of art. Digital reproductions of works of art (even within one museum) might have differing levels quality—photographed in different eras, with different tools.

Coming away from the research trip, I was struck by how almost impossible it would be to actually truly accurately reflect the exact particulars of Blake’s original books in a digital project. Even if one were to pay each museum to retake photographs of each plate using the exact same specifications, it would still be near-impossible to ensure that the colors on the plates were exactly as Blake painted them, two centuries ago. Watercolors are among the most unstable of art materials—degrading incredibly easily with even the smallest bit of exposure to ultraviolet (UV) light. The integrity of Blake’s colors on each plate are at the whims of the curators who care for them.

Yet each museum stores its objects differently. Some museums keep plates still bound together into one book; others have unbound the books, isolating and individually mounting each print. Some copies kept in book-form insert a thin, protective tissue paper between Blake’s plates; other museums do not use these protective measures. Some individually mounted prints have tissue-paper sheets covering the image from any light exposure; some mounted prints have a sheet of opaque printer paper (offering even more protection); some mounted prints have no protective cover at all.

In one particular copy of Songs of Innocence and of Experience that I examined, several of the more “famous” plates (e.g. “The Tyger,” title pages, frontispieces, etc.) were completely fried by the exposure to sunlight. On these plates, not only were the water-colors faded thanks to the material degradation spurred along by UV exposure, but the very color of the paper Blake printed on was entirely different—a crispy yellow-brownish instead of creamy white. While I was examining this particular copy, I was particularly reflecting on how moments of error, like this, may have skewed our project. BlakeTint deletes the white background of each of Blake’s plates, isolating only the colors Blake added. The sunburned plates (where white background looked yellow-brown) could have easily been “misread” by the algorithm. The computer, mistaking accidentally-sun-bleached paper for purposeful coloration, might not have necessarily deleted those portions of the plate correctly.

With so much disparity among museum collections (from micro-level decisions such as inserting a piece of tissue-paper between pages to what camera lens to use when photographing to how often to display a work of art), and no standardization or calibration between them, BlakeTint and any projects like it are entirely mediated by (multiple sets of) conservation practices, museum administration, and bureaucracy.

Overall, my archival work sparked many thoughts on the accuracy and limits of digital resources, causing me to think about what BlakeTint’s brand of data management and organization of information both gains and loses in its implementation. What are the methods of storage and access that digital tools allow that special collections do not? What gets lost in the order of multitudes—in considering data from such a distance?

Especially in the era of Covid-19, when so many of us have been kept away from the archives and the actual works of art, tools like BlakeTint do offer glimpses into and askance glimmers of the museum. Yes, we lose the particularity of the individual encounter in the archive/museum, but perhaps where we lose certain data points, we gain others. If flawed, digital tools like BlakeTint do allow a specific kind of looking across collections, across multitudes that we could not possibly experience in the museum, itself, where the individual page or work of art becomes more precious than the eagle’s-eye totality. An aggregation for the eye. An ordered multitude.