To enrich theuniverse of my workauthor and give him a visual consistency, I wanted to test the performance of the AI image generators. I must admit that the result, at first glance, is astonishing. quality and of realism, and that he approaches the scenes that I have in mind. But as soon as you dig deeper, it becomes difficult to get to the level of precision exactly matching the scene described in words.
L'optimization of a picture by theAI is no small matter. In the“Excerpt” article of my work in progresswriting (https://laurentchani.com/2025/07/29/extrait-description-de-la-salle-de-conference/), I describe a conference room which constitutes a key location demon narrativeThe key points are:
• A vaulted representing a changing sky.
• Three concentric levels connected by stairs, each including twelve tables spaced, equipped with three leather armchairs and ofcomputers for each workstation.
• A central arena displaying, on request, digital scenes.
• A character placed in the center of thearena.
I have a idea very precise about what this room should look like. I asked theAI to progress step by step in generating a hyperrealistic image of type photographic.
It turns out that it is not easy, at the moment, to obtain exactly what one wants with this type oftool. I produced a good thirtyimages before arriving at a result which I deemed acceptable. I present to you here four images only, corresponding to the key steps of design process.
The main thing issue of AI generator that I have used—though considered one of the best—is that it works best in the early hours of the morning, when theNorth America sleeps. At such times, the available resources are more important and the time limit shorter production times.
I use the same one conversation thread so that theAI keeps them information issues of previous requests. However, the more we get into the details, the more complicated it becomes to obtain a optimization satisfactory. For example, my scene comprises three concentric rows of tables. When I ask to place a character on thearena, the result is roughly correct. But if I want to move this character towards the exact center of thearena, L'generated image suddenly finds himself with four rows of tables !
This type ofinconsistency reproduces regularly. We quickly obtain a suitable version of thepicture desired, but his refining becomes a real puzzle. You have to constantly juggle with the formulations to make it clear exactly what is expected. In this specific case, between the twentieth and thirtieth iteration, none notable progress was not obtained.
It probably also depends on the method used to guide theAI, but I remain convinced that this technology is still in its infancy. I have no doubt that within ten yearsyears, his performances will be much higher - especially when we know that the GAFAM invest hundreds of billion dollars in AI of all kinds.
Conference Room 02
Conference Room 03
Conference Room 04
