The Fujifilm patent that chooses its autofocus method by frame texture

Fujifilm published a patent on October 6, 2026 that describes a change in how focus is calculated. Before measuring, the camera analyses what kind of detail sits in the focus area and picks a calculation method. It is not a new sensor or a different system. It is another way of deciding which maths to use.

The difference looks small, but it moves autofocus from one category to another. A classic system measures a difference and calculates a distance. This one classifies what it is looking at first.

From measuring a distance to classifying a scene

Phase-detection autofocus works with pairs of pixels that see the scene from slightly different angles. The camera compares both signals and deduces how far apart they are. The larger the separation, the further away the plane that sits in focus. The whole process is a measurement.

The patent adds a step before that. Instead of always applying the same calculation, it examines the differences between neighbouring pixels to estimate which kind of detail dominates the focus area. With that estimate it picks the method, and when mid-frequency information dominates it can apply several and combine the results.

That is the real change. If the result depends on a prior classification, autofocus behaviour stops being predictable from the scene and starts depending on how the camera read that scene. For the photographer, a focus miss is no longer explained only by a lack of contrast: it may come from the system choosing the wrong category.

It is not a criticism of the design. It is the consequence of autofocus having become a recognition problem rather than an optical one. The patent also mentions a trained subject-recognition model that helps decide which calculation applies.

Where it hurts: both extremes of detail

Classifying by frequency makes sense because the two extremes of detail are exactly where phase detection gets complicated. In a low-frequency area, such as a clean sky, a blank wall or skin without texture, there are few edges to compare and the measurement becomes ambiguous.

At the other end, a high-frequency, repetitive area, such as a fence, a woven fabric, foliage or a tile pattern, offers too many similar edges. The camera can pair the wrong pixel and confirm focus that is actually off, without warning.

Mid frequency is the territory of portraiture and everyday scenes: hair, eyes, clothing, a face at medium distance. That is where the patent proposes combining several calculations, and it makes sense. It is the most common case and the one with the most room to average out errors.

The document announces no product

A patent describes a method, not a product. The full text of the document is available to check each point. Publishing it does not mean Fujifilm has put it in a camera, or that it will. Nor does the text reference any specific model.

There is a detail about dates worth keeping in mind. The patent was published in October 2026. That date is not the date of invention: most offices make the file public about eighteen months after filing. The work may be considerably older, and it may even be in cameras sold today without anyone having pointed it out.

That is why the link to an AI-assisted autofocus system rumoured for a future model is, for now, a hypothesis. The document describes a trained model that helps with the decision, and that fits the idea, but it does not confirm it.

What does become clear is the direction. Autofocus in today's cameras is no longer played out only in motor speed or the number of points. It is played out in how well the system understands what it has in front of it, and that looks more like a classification problem than a mechanical one.

For anyone who photographs, the practical consequence is a single one. It pays to test autofocus in the usual hard scenes: smooth surfaces in low light, repeated patterns, backlight. And not to trust that a higher point count will solve the problem. The patent, if it reaches production, aims exactly there.