As often with artificial intelligence, intelligence is more interesting than artifice. A new tool, Lexica.art, a massive database of images generated by Stable Diffusion gives a new example. Because it proposes, in addition to the images, the prompt that produce them — i.e. the human part of the creation.
The quest for the perfect tag
Browsing through the prompts, one can only be amazed at the sophistication of the texts submitted to the machine. There is, of course, the basic idea — for instance, a portrait of the late Queen of England as Ozzy Osbourne (something like Janis Joplin if she had lived more than 27 years).
Better: the commands used to refine the request :
— framing and composition (stunning environment, wide-angle, aerial view, massive scale, street level view, landscape, panoramic, lush vegetation, idyllic)
— style and sharpness (art deco, rococo, art nouveau, baroque, wallpaper, poster, sharp focus, hyperrealism, insanely detailed, lush detail, filigree, intricate, crystalline, perfectionism, max detail, 4k uhd, spirals, tendrils, ornate, HQ, angelic, decorations, embellishments, masterpiece, hard edge)
— lighting (god rays, hard shadows, studio lighting, soft lighting, diffused lighting, rim lighting, volumetric lighting, specular lighting, cinematic lighting, luminescence, translucency, subsurface scattering, global illumination, indirect light, radiant light rays, bioluminescent details, ektachrome, glowing, shimmering light, halo, iridescent)
— colors (vibrant, muted colors, vivid color, post-processing, colorgrading, tone mapping, lush, low contrast, vintage, aesthetic, psychedelic
Hidden gem artstyles)
— and so on.
Best of all, when I try to search for drawings by a little-known illustrator (Tomer Hanuka — he did sign a few memorable covers for The New Yorker magazine), I come across images that resemble his style but for which there are no prompts with his name. Instead, other illustrators, such as Victor Ngai, Takato Yamamoto, Pascal Blanché, Atey Ghailan or Sachin Teng, whose existence I didn’t know until then — and some I wouldn’t have thought of as Tomer Hanuka, such as Mike Mignola, Moebius, Alphonse Mucha or Edward Hopper. Stable Diffusion seems to think that Hanuka’s art lies at the intersection of these.
In short: let’s get back to the intelligent part — the expertise that users had to develop to, trial and error, find the right terms and references to get the desired image from Stable Diffusion. The art is there.
But is it art (as we knew it) ?
You may have heard about the Colorado State Fair in late August where Jason Allen won first prize with his work “Théâtre d’opéra spatial”… generated by the Midjourney AI. No matter the scandal, the will to preserve “human art” and other corporatist reflexes*. The most interesting thing about this case is that Allen insisted on keeping secret the prompt used to make the image. This haiku is art (and incidentally, the choice among the many images created by Midjourney).
But is it real art ? Does the artist create his work emerge from nowhere or at least from his only subjectivity? Or is he “crossed” by the images produced by the society in which he lives? If so, what’s the difference with Dall-E, Stable Diffusion, or Midjourney?
“In our complex contemporary societies, we are first of all filters, explains Yves Citton. There are so many things that pass through us: we block what seems irrelevant or dangerous to us, we let pass what moves us or moves us, but by trying to inflect it. Everyone recombines, reorganizes, transforms.”
The artist freezes momentarily and in an original way thoughts, inspirations, sources, references conscious or not, which are the common material of a society.
The birth of a new textual art
The same question was asked about photography: if it is enough to press a button to produce the image, is it still art? Of course, we know that it is necessary to frame, to anticipate the effects of light, speed, depth of field and to wait, as Henri Cartier-Bresson said, for the decisive moment. These are new skills, a new art.
Knowing how to be in the right place at the right time at the right distance and in the right relationship to the subject then supplants the knowledge of the composition of the significant components of painting (including to the benefit of painting with the rise of currents that do not claim to imitate reality, which photography does — too — perfectly)
We can imagine that the current appearance of AI allowing prompt-to-image gives birth to a textual art: the art of composing these short texts that allow to generate the right image.
Let’s call it “iconography”.
Further reading
* On september, 5th, the popular furry art community Fur Affinity announced that AI-generated art was not allowed because it “lacked artistic merit.” Before that, the animation portal Newgrounds banned images made with Artbreeder: “Bottom line: We want to keep the focus on art made by people and not have the Art Portal flooded with computer-generated art.”
