心地とインターフェース

Around 2023, I took up something I called AI ikebana.

It was just as image-generating AI such as Midjourney was rapidly becoming familiar. All you had to do was type in some text, and one image after another appeared, unlike anything you had seen before. The experience itself still felt new, and simply looking through the images was fascinating.

I made up my own definition of AI ikebana: “Using AI to arrange real or imaginary flowers in any style you like. By combining spaces and materials that seem as though they could exist, but do not, you can enjoy flowers in a new way. (The kana school.)” It was something I did purely for fun.

Some of the plants took forms that could never occur in nature. Things appeared that were neither quite flowers nor leaves, and at times even the boundary between the vessel and the plants became uncertain. I imagine this was the result of Midjourney at the time vaguely interpreting the names of flowers it did not yet know. Many of the arrangements felt like something that had appeared in a dream, its form left indistinct.

The AI ikebana I made then is still on Instagram. Looking back at it after a long while, I find it rather charming. Compared with today’s image generators, the details are rough. Look closely and the forms are strange, the textures slightly off. Yet seen now, that somewhat lo-fi quality feels appealing in its own way.

Since then, I have tried making AI ikebana several times with newer image generators. But everything they produce is simply beautiful, and almost all of the unease in those earlier images has disappeared. The petals are finely rendered, the light is natural, and the vessel has a lovely texture. As a single image, the newer work is probably more complete.

And yet, somehow, it is not as interesting as it used to be.

The uncertainty in those 2023 arrangements—of not quite knowing whether something was a plant at all—and the strange atmosphere drifting somewhere behind the image were gone. The expressive power of image-generating AI has improved astonishingly over the past few years. But perhaps something was lost along the way.

There may be something similar in photography. Film has its own atmosphere and texture, but lately I have also heard that more people are deliberately using old compact digital cameras and early iPhones for their particular appeal. Older digital photographs have low resolution, noise in dark areas, and colors that can look a little strange. What was once merely a limitation of performance begins, with time, to look like an expression of its era.

While thinking about this, I came across Camera Looks, a feature announced for the Google Pixel 11 in August 2026.

Among the options is a Look called “Digi,” inspired by inexpensive digital cameras from the early 2000s. Rather than applying an effect to the photograph afterward, as an ordinary filter would, it changes the way the image data from the sensor is processed. It seems to work by reducing the amount of processing from the moment the photograph is taken.

Smartphone cameras have evolved to brighten dark scenes, prevent blown highlights and camera shake, and make it difficult for anyone to take a failed photograph. That is useful, of course. But once every photograph is beautifully corrected, there are times when we want to step away from that uniform way of seeing. I was intrigued that the newest smartphone was looking back to an old digital camera.

Come to think of it, every image-generation model has its own quirks. What does it judge to be beautiful, and how does it draw it? What kind of hand does it give to letters? When it encounters something it does not understand, how does it try to interpret it? As models improve, perhaps those quirks gradually become less visible. Yet I have a feeling they are precisely what, in retrospect, will come to be called “the way AI looked back then.”

Just as people choose film, use old lenses, or carry an old digital camera, might there come a time when we deliberately choose an older generative model?

We might work not with the newest model, but with something older, like Midjourney V4. We might resist raising the resolution too far. We might choose an old model to make images with that particular air of the time—a little unsettling, indistinct, and dreamlike. It would be interesting if a notion like “AI retro” were to emerge.

As a rule, newer technology performs better. But when we think of technology as a means of expression, the new does not always replace the old entirely.

As the expressions made by AI become increasingly polished, perhaps a day will come when we feel nostalgic for the time when it was still a little awkward.