Auto-generated Biomics
We’re currently working on a project for the National Trust for Scotland’s Robert Burns Birthplace Museum. During our recent ML-based hack day, I decided it was important that more people should know about the life of this pioneering poet, and that ML tools could help educate people in an accessible way — a simple comic strip.
I first asked Chat GPT to write the script for an eight-panel comic strip. This produced something that seemed very plausible. Then I asked Open AI’s image generator to create an image for each panel, given this script.
This worked quite well, and some of the illustrations were pleasingly odd.
The only problem with this ludicrously simple idea was that I found it quite addictive. I kept thinking of other people whose life stories deserved to be told in this way (a literary form that I’ve decided to name “biomics”).
After a bit of tuning, I made a simple script that yoked together the two tools, and generated a PDF of the results. The input to the script is just a person’s name. The script invokes ChatGPT with the prompt: Please describe the life of %s, in the form of the descriptions of each panel for an eight-panel comic.
Then each panel description is fed to the image generator, which is asked to provide two images in each case. The results are then output as an eight-page PDF, each page having the script for that panel and the two alternative images created in response to it.
As well as Robert Burns, my collection of educational material now includes graphic biographies of such luminaries as Tracey Emin, Caravaggio, President Putin, Maradona, Elton John, and Liz Truss.
The results in each case are impressive — although much of the interest comes from the anomalies. When requesting the images, the script generated a prompt including not only the description of the panel, but also that this was to be a panel in a comic about that person. Sometimes this was effective: I was impressed when a panel for Caravaggio was rendered in chiaroscuro. And the style of many of the panels for Elton John was also well chosen.
Sometimes however the image generator didn’t take the hint. For example the script for Caravaggio establishes in its first panel that his first name was Michelangelo; and varies the references to him, in a nice bit of literary style. For panel 3 however, which simply refers to him as “Michelangelo”, the image generator appears to be depicting the more famous artist of that name.
And what is happening with the alternative image for the first panel of Elton John’s biomic, I can’t explain.
My favourite misunderstanding is Panel 5 of the Liz Truss story. The script includes: “Liz Truss, now a Member of Parliament, is seen passionately speaking on the floor of the House of Commons.”
The image generator took this literally…
Similarly the final panel for Jeremy Corbyn which “… shows Corbyn stepping down as Labour Party leader” really does.
Although both the tools I used are from OpenAI, the image generator is more nervous about giving offence than ChatGPT. The latter wrote a script for one panel mentioning Caravaggio’s violent behaviour, and specifying “the panel also includes a glimpse of his controversial painting, Judith Beheading Holofernes”. The image generator refused to illustrate this panel, presumably because of the reference to beheading.
And in a particular act of cowardice — or censorship? — the image generator refused to illustrate any panels of the life of President Putin, rendering this comic quite disappointing.
Doing this too much does bring some diminishing returns, as one recognises recurring tropes.
And while ChatGPT is happy to mention beheadings which the image generator refuses to picture, it is still noticeably anodyne. Possibly as a reaction to some of the recent court cases brought against it, ChatGPT seems reluctant to say anything very critical of any of our subjects.
Ben is a Director at Cogapp, a digital agency specialising in the cultural sector. Please get in touch if you’d like to hear more about our work.
If you’re interested in joining us for a hack day, there’s more information on our website.
We’re on Twitter or you can contact us via our website.
