Does AI understand Depth of Field ?

Simple answer, to a certain degree – yes.

AI Prompt Midjourney v5 – A telephone with a shallow depth of field, photograph, 1970s, kodachrome – 3 out 4 frames in the grid had shallow depths of field, not bad.

What is Depth of Field?

Depth of field is how much remains in focus from your focal point. Depth is the space between foreground and background. For example if you set your focal point (focus) on something 1 meter away from you, then everything at that distance stays in focus. A shallow depth of field is where nothing much more, in front or behind, that focal point is in focus, a deep depth of field is where objects, both close and far from your focal point, remain in focus.

What creates Depth of Field?

In traditional photography a photographer needs to control light. Too much light and an image is blank (over-exposed), too little and the detail never appears (under-exposed). The point in the middle is called perfect exposure. The photographer has two tools to control exposure. The first is shutter speed, or the exposure time, the longer the light shines on the frame (negative) the more exposure the film receives. The second tool is aperture.

Aperture and f stops

f64 ! Shutter speed and Aperture settings on my Copal No.3 lens.

In the early days of photography lenses did not have an adjustable aperture. Reaction (exposure) times were so slow at the birth of photography in 1839, that lenses were left wide open to allow maximum light and reduce the exposure time (shutter speed) to avoid movement blurs.

As chemistry improved, the increased reaction time (film speed) meant that photographers had less and less time to stop (the shutter speed , putting the lens cap back on) at the perfect exposure. This meant they needed to reduce the light by using aperture plates made of wood or metal with a circle cut in the middle. This not only obscured (reduced) the amount of light entering the camera but it also increased the focal ratio of the lens. This is where the f stop number comes from: f = focal ratio which is measured by dividing the focal length of a lens (e.g 35mm , 50mm, 400mm) and dividing it by the effective aperture (the diameter of the entrance hole). Apertures are standardized by f-stops, based on halving the amount of light with each click e.g. | f1.4 | f2 | f2.8 | f4 | f5.6 | f8 | f11 | f16 | f22 |. A wide aperture e.g. f1.4 , which lets in a lot of light , will create a shallow depth of field, a narrow aperture e.g. f22 , which lets in little light , will create a deep depth of field.

AI Prompt Midjourney v5: A telephone taken at f1. 4, photograph, 1970s, kodachrome. – none of the 4 generated had a shallow depth of field, this image did have beautiful composition though. To have a shallow depth of field only the telephone dial, for example, should be sharp.

I see loads of AI prompters adding in aperture numbers e.g. f1.4 , lens focal lengths, e.g. 50mm , and cameras e.g. Canon R5. The theory would be that professionals often caption images with all these details so to replicate the style all these keywords are relevant. In practice I don’t believe AI is trained for this technical input, nor is it reading the exif metadata of images which could be a vast data resource (watch this space)

What AI prompt works best for Depth of Field?

AI Prompt Midjourney v5: A telephone with an out of focus background, photograph, 1970s, kodachrome

Unsurprisingly, at least with MidJourney 5, natural language is working best. A telephone with an out of focus background got 4/4

Following on from my last post I tried to add an element to that background. The AI then got me 2/4 where a horse was out of focus in the background. The added concept is clearly getting confusing for the AI to follow but it is doing much better in a single sentence without separation (commas) than it did in my previous post about positioning.

AI Prompt Midjourney v5: A telephone with an out of focus background which has a horse, photograph, 1970s, kodachrome – interesting point of note is the slant of the horizon, which I find to be quite uncharacteristic for the AI images I have received so far.