In my previous article, titled "The author doesn't throw anything away", I looked for a visual concept to illustrate the dilemma between inflicting a weight loss treatment to a text or, on the contrary, to flesh it out. The idea quickly came to me to stage a author facing a mirror reflecting two versions of himself: one drastically thin, the other visibly overweight. Scratching his head, the author, standing next to his work table, asks: what strategy to adopt?
THE theme seemed to me well served by this allegory. So I asked my virtual friend, L'AI, so that it generates a hyperrealistic scene corresponding to this vision. We first exchanged to check that he understood my request. He drew the following written description from it:
Realistic scene to generate:
The man, in a suit, cigar in his mouth, stands in front of a large full-length mirror.
On the table next to it: handwritten sheets of paper and a quill pen in an inkwell.
In the mirror:
• Left: A slimmer version of himself in a suit (slim silhouette).
• In the center: a more corpulent version in costume (larger silhouette).
Himself (right, outside the mirror) : medium build, observing the scene thoughtfully, scratching his head.
All three men have the same face (suggesting that they are one and the same person in different states).
All three are holding a lit cigar, with visible smoke.
No lie, I battled for hours with theAI to obtain a picture corresponding to my vision. And yet, the end result isn't exactly what I had in mind. Between responses like "I can't generate this image because it violates our content policies" and the misinterpretations, it's clear that it's particularly difficult to obtain a scene that seemed simple to understand.
So I am making serious doubts on the ability of the tool to faithfully respond to requests. Perhaps I did not go about it in the best way, or perhaps another LLM (generation model ofimages) would have been more suitable. It's possible. But given the billions of dollars invested in my virtual friend, I doubt its competitors can do much better.
Moreover, this difficulty in carrying out this visual project is not an isolated case: I have encountered it many times, at least in the field of image generation. On the other hand, I have no doubt that in the long term - probably soon - these imperfections will be corrected.
Below I offer you a selection of a dozenimages, among the thirty that I asked him to generate to get closer to my objective. Small note: when the image generation deviated too much, I started from scratch in a new conversation session, without any history to avoid interference with previous exchanges and images produced.
In conclusion : from my point of view, the graphic designers still have some good days ahead of them. But who knows how quickly AI will take over - and, why not, transform a novel in comic strip…(my dream).
