An AI project: Die Silhouette, by Anna Corsep

OVER THE YEARS I’ve collected most books which have ever been published on the subject of portrait silhouettes. Among them are a few I cannot read. One such is a slim volume published in German, in 1908, by the silhouettist Anna Corsep.

Die Silhouette

A rather grubby copy of Corsep’s book, alarmingly titled “Die Silhouette”, has been languishing on my shelves for a number of years. It’s a short book, just 47 pages, and is printed in an old-fashioned German fraktur font, complete with long s’s and other typographical oddities. This makes it difficult to scan with current OCR (Optical Character Recognition) technology. Since I speak no German, and could not really justify paying for a professional translation, I had given up much hope of ever reading it, so simply contented myself with looking at the illustrations.

The German title translates as “The Silhouette”. Inside the book are 36 silhouettes which are a mixture of silhouette portraits and black, paper-cut designs. A few are by Anna Corsep herself, but most are by other German silhouettists of the nineteenth century. Some of these were artists I’d never heard of, such as Paul Konewka. I was eager to learn more, so earlier this year, after listening to some of the current talk about AI (Artificial Intelligence), I wondered if ChatGPT might be able to read it for me. So I asked it.

“Absolutely, Charles” came the instantaneous reply
“I’d be glad to help with that”.

Using AI as a Translation Tool

ChatGPT asked me to scan the book and upload it as a PDF. This turned out not to be so easy and there were a number of false starts. Eventually, I worked out that if I scanned each page as a separate image, uploading them one by one, it could have a go at reading them. ChatGPT then provided me, page by page, with both a German transcript and a sub-optimal, modern-English translation. This took time, as file upload limits meant I could only send two or three pages per day.

At first, my aim was simply to read the book, but as the days passed I became increasingly fascinated. Imperfect though the translated pages obviously were, it became clear to me that Anna Corsep was a kindred spirit. She was just as obsessively fascinated by silhouettes as I am today! I began to think about making a proper English translation, since none seemed to exist.

Instinctively, I felt I should follow the same layout as Corsep’s original, keeping all the illustrations in the same place, with text flowing around them in the same way. I also felt the book should be printed in an appropriate font: one which might have been used in the 1900s had the book been printed in English. I settled on using Caslon, which is both easy to read and dates from the same period.

When I discussed these ideas with ChatGPT it quickly suggested it could both translate the book and also do the layout, following my ideas about how to the present the finished book. This sounded like a good plan, so I asked it to make a trial layout for page 1, which it had already translated.

The results were shockingly bad and a real lesson in the limits of AI technology. About the only thing it got right was the colour of the paper! The English text (which it had previously translated quite well) was now completely garbled. ChatGPT even took it upon itself to redraw Corsep’s small, silhouette vignette, with which she opens the book. This, of course, was completely inappropriate and totally astounded me. Why do that? ChatGPT had sort of done what I asked it to, but it appeared to lack any sense of judgment at all. I realised that AI has no common sense.

When I challenged ChatGPT about this, it disarmingly replied:

“Yes, it’s a crash course in AI quirks and limits!”

Getting along with ChatGPT

AI works better when you give it really simple instructions, breaking things down, step by step. Combining the different stages of book translation was clearly a mistake, so I created a workflow, trying to work out what ChatGPT could usefully help with, as follows:

1. Scan each page as a greyscale image (Photoshop)
2. Transcribe into modern German (ChatGPT)
3. Line-by line, visual check of the transcription (me)
4. English translation (two versions: ChatGPT and DeepL)
5. English editing, using my knowledge of the subject (me)
6. Layout text and images to match original (me)

Once I broke the process down like this, ChatPGT and I began to get along quite well. Each morning I uploaded a couple more pages, asking first for a transcription, then a translation. I then displayed the transcription in a Fraktur font, to make it appear similar to the scanned pages from the book.

My idea was to make a visual comparison, looking for errors. Unlike OCR, which tends to get individual characters wrong, ChatGPT gets whole words or even sentences wrong. Strangely, it does this in a misguided attempt to write grammatically-correct text, but it makes errors far easier to spot! I have a very visual mind, with a good eye for shape, so words which were the “wrong shape” began to jump out at me, letting me know where to retype. The fact that I only understood a handful of the words I was reading no longer seemed relevant; after a few pages the process became quick and efficient. I even stopped reading a long ‘s’ as ‘f’!

Discovering New Artists

Corsep writes in her book about many artists, some of whom she admires, while others she clearly doesn’t. The former group includes Paul Konewka and Robert Erbe, both of whom she illustrates in the book, while the latter group includes Marie Rehsener and Johanna Beckmann. The book includes many fascinating snippets of information. Did you know, for instance, that the famous German writer Johann Wolfgang von Goethe also made silhouettes? Several of his portraits are illustrated in the book.

Whenever Corsep writes about an artist of whom she doesn’t approve, she begins by writing about their life and work in quite a factual manner. She then changes tack, using a phrase like:

“Far be it from me to criticise…”

Before going on to do exactly that, at some length. Corsep has firm views about what constitutes a good silhouette. She writes about how images need to be flattened, with arms and legs (for instance) altered slightly to work better in profile. For her, each object in a composition needs to be given its own space, not overlapping with other objects. Such overlaps, which work well in a drawing or painting, create confusion in a silhouette composition, where objects all too easily merge into one another.

Her method of taking silhouettes

Later in the book Corsep illustrates a number of her own paper cuttings, and also describes in detail her method for taking portrait silhouettes. This is similar to the method used a hundred years earlier by English artists such as John Miers: making a life-sized shadow drawing and then mechanically reducing it to the desired scale. She reduces these drawings using a laborious grid method, still well-known to artists today, copying portions of the outline square by square.

This method, according to her, is the only way to create a convincing silhouette likeness and works far better than any attempt to cut the image freehand (such as you might see in beer halls, restaurants, or bathhouses for instance!)

Several of Corsep’s portraits, taken using this method, also appear in the book. These she surrounds with imaginative paper-cut borders of her own design. Far be it from me to criticise, but do these portraits quite live up to the expectations created by her detailed description of how they were made?

I wonder what Anna Corsep might have made of the way I work at events today?

Finishing the Book

The last stages in making the book involved editing the text and creating a layout similar to the original. Since ChatGPT’s English translations were clearly prone to flights of fancy, I obtained a second translation of each page from DeepL. Comparing the two, I found ChatGPT’s unreliable translations to be more engagingly written, while the DeepL translations were somewhat dry in tone, but felt more convincing and accurate. Between the two of them I felt I could understand clearly what Corsep was trying to say, so my task became to combine the best of each in order to make my own English version of the text. I did this in light of my own artistic experience and knowledge of the subject matter.

In this way I found I could make what I believe to be a fair and accurate translation of Corsep’s words, focussing on what she seemed to be trying to say about silhouettes. I also kept in mind the need to preserve her unique, and inevitably old-fashioned, turn of phrase. Corsep writes in a bright, clear voice, full of conviction about a subject she clearly cares deeply about.

At the end of the book I’ve added a series of footnotes, since I felt some of Corsep’s cultural references might be lost on an English-speaking audience from the 2020’s. These also include notes on the translation and a few comments of my own about the art and history of silhouettes. Finally, I made the book available as a print-on-demand book using Lulu. I’ve made it available at a price which I hope is not prohibitive, but which does include a small mark-up for my time in making it.

Is this book important?

Online, German references to “Die Silhouette” describe the book as formative. Reading it, I can see why. Many of the ideas Corsep expresses have since been repurposed by later writers on the same subject. The book is clearly important for anybody interested in the art and history of silhouettes, but this is admittedly quite a niche audience.

Thinking more broadly, Corsep is writing at a unique moment: just 100 years after the heyday of the silhouette in Europe and just 50 years after its demise due to competition with photography. This demise is clearly something she feels very deeply, since she writes about how silhouettes have too readily been overlooked in favour of this exciting new technology.

Corsep describes the profile line (rarely seen in photographs) as the main line of the face, which gives each of us our individual character. She points out that many animated, expressive faces, become flat and dull in photographs. Later in the book she describes the shudder of emotional resonance she feels when viewing or working on an old friend’s silhouette:

“I’ve often suddenly been reminded, as though I could hear them, of their tone of voice and even the words they spoke”.

This is a theme I also touched upon in my TEDx talk: that in an age so oversaturated in photographic images as today’s, it may be time to look again at the silhouette. At its heart, Corsep’s book is a plea for silhouettes to taken more seriously, both by the general pubic and the art establishment, written at a time when nobody was taking such an antiquated art seriously at all. It’s a plea I wholeheartedly endorse: it could not be more relevant in our modern age.