When I was an art student in Toronto, my BFA thesis advisor, April Hickox, loaned me her copy of Science Is Fiction: 23 Films by Jean Painlevé. At the time, I’d been noodling around with macro photography, making atmospheric photographs of small-scale phenomena that looked like vast seascapes of glaciers and shattered ice floes. (This work later became the subject of my BFA thesis; Terra Nova.)
“Here,” she said. “I think you’d like this.”
I watched the DVD that evening.
“It was life-changing!” I exclaimed—with the febrile hyperbole of a sleep-deprived art student—when I returned it a few days later.
Dunno if I’d go that far today, but it is a damn good DVD. (I think it’s out of print?? So if you ever see a copy, grab it. Get one for me, while you’re at it.)
Jean Painlevé (1902-1989) created hundreds of short films during his long and varied career. The son of a French Prime Minister, Painlevé became one of the youngest researchers ever to deliver a paper to the Académie des sciences. In 1924, he graduated from the Sorbonne with a degree in physics, chemistry and biology.
Through his partner, Geneviève Hamon, Painlevé became friends with avant-garde artists and writers such as Alexander and Louisa Calder, Max Jacob and Man Ray. Occasionally, Painlevé slipped away to the cinema, enjoying films by Mack Sennett and Georges Méliès.
An abiding love of the sea informed his filmography; which began with L'œuf d'épinoche: de la fécondation à l'éclosion (1927), and ended with 1986’s Les pigeons du square.
Painlevé wrote Ten Commandments for the filmmaker: abandon every special effect that is not justified (number five); do not substitute words for images in any way (number nine); do not influence the audience by unfair means (number three).
AI and text-to-image generators arguably violate a few of these rules—particularly the one about substituting words for images (but at least you get an image in return for your words). Then there are deepfakes, which are easier than ever to make and proliferate. And what is AI but a special effect? Hollywood has been using AI for years.
As a documentarian, Painlevé would have probably hated AI. Making a film was more than capturing a scene and adding narration. For Painlevé, filmmaking was the ultimate expression of his scientific practice. His work was inventive—but not invented. It was research.
Let’s see how well Stable Diffusion renders Jean Painlevé’s inimitable (imitable?) style, shall we?
These images, above, were generated with Stable Diffusion 1, using prompts including Painlevé’s name, and keywords such as “coral”, “seahorse” and “jellyfish”. (Don’t you love the starfish?? It’s waving!) Compare this to the real thing, below:
After making these images, I decided to test Stable Diffusion 2 and Stable Diffusion 1 with an identical prompt: “Jean Painlevé films”.
Spot the difference? Something has clearly changed on the back end, either with the AI’s training data or its programming. Stable Diffusion 2 rendered a random guy in an avant-garde tableau, while its predecessor showed me images that look like they came from a Painlevé film.
I have to wonder whether Painlevé’s estate or Criterion Films (which released the DVD I referenced above) had a word with Stable Diffusion’s creators, because the second version of Stable Diffusion is not as precise as its predecessor. Has Painlevé’s imagery been removed from the datasets powering Stable Diffusion 2?
It could also point to something I noted in an earlier post. When Stable Diffusion 2 was released, I found, with dismay, that it produced less photo-realistic images, in favour of more illustrative or cartoonish images—like those in fashion with the NFT crowd. This is epitomized by the examples below.
Here’s “Jean Painlevé jellyfish sea”, by Stable Diffusion 2 (on the left) and 1 (on the right). A noticeable difference!
How about a longer prompt? Here’s what happened when I typed “Jean Painlevé diatom film color 1970s realistic documentary” into Stable Diffusion 2 (left) and Stable Diffusion 1 (right). Stable Diffusion 2 is clearly world-building here (I want to see that film, don’t you?), but it’s not even close!
We think of AI (and technology in general) as progressing in quality from one model to the next. The more we use AI, the more it learns, calibrates and improves. So how do you explain such a clear regression? Like HAL 9000 losing his marbles at the end of 2001: A Space Odyssey, Stable Diffusion 2 said, “Jean Painlevé? I don’t know her.”