On Exposure

The numbers explain the failure before any technique does. Photographers rate scene brightness in exposure values, where each step is a doubling of light. A home interior sits around EV 5 to 7 at ISO 100; an office is 7 to 8; a heavily overcast sky outside is EV 12; a sunny scene is EV 15, the value the old “sunny 16” rule encodes^[1]. The room and the view through the window are five to ten doublings apart, and that is using table averages: the sunlit sill is brighter than the table’s sky and the shadow under the couch is darker than its room.

A current full-frame sensor records roughly 13 to 15 stops of range at its base ISO, and about three of those stops at the bottom are too noisy to print, which leaves 11 or 12 usable; an older DSLR starts from 11 or 12 and keeps 8 or 9^[2][3]. So a dim room with a bright window is a scene at or past the edge of what one frame can hold. No dial on the camera changes that, because the dials do not change the scene.

Exposure, in the physical sense, is light per unit area arriving at the sensor over the time the shutter is open, measured in lux-seconds^[10]. It has two controls. Aperture sets how much light per second gets through the lens. Shutter speed sets how many seconds. That is the whole list.

The aperture is written as an f-number, the focal length divided by the diameter of the opening: a 100 mm lens with a 50 mm opening is f/2^[4]. The sequence on the ring looks arbitrary (1.4, 2, 2.8, 4, 5.6, 8, 11, 16) until you see that each step multiplies the diameter by the square root of two, which doubles the area, which doubles the light^[4]. Shutter speeds do the same thing without the disguise: 1/60 is twice 1/125. Both dials run on the same arithmetic, which is why photographers count in stops and treat them as interchangeable: close the aperture one stop, open the shutter one stop, same light^[1].

The combination collapses into one number, EV = log₂(N²/t)^[1]. Run it backwards for the living room. At f/8 and ISO 100, a room at EV 6 needs N²/2^EV = 64/64, one full second. The window, at EV 12 on an overcast day, needs 1/64 of a second in the same frame. The distance between those two shutter speeds is the photograph’s problem, stated exactly.

Canon’s beginner page on the triangle teaches that changing the ISO lets you “change the sensitivity of the sensor”^[5]; Nikon’s is titled “Understanding ISO Sensitivity”^[6]. The language is inherited from film, where a faster emulsion really was more sensitive. For a digital sensor it is wrong in a way that matters. “You cannot adjust the ‘sensitivity’ of a sensor,” Richard Butler wrote in DPReview’s explainer; “only changing your exposure can adjust how much light you capture”^[7]. The standard that defines the number, ISO 12232, does not describe a property of the sensor at all. It describes the response of the camera’s whole processing pipeline, relating exposure to the lightness of the finished JPEG^[7][8].

ISO, in other words, is an instruction about how bright to render the photons that were already counted. It looks like a third leg of the triangle because doubling it doubles the lightness of the file, the same as a stop of aperture or shutter would. But doubling it adds no light; it tells the camera to render the same light brighter and to clip sooner at the top, which is why dynamic range falls a stop for every stop of ISO^[12]. People who build imaging software have been saying this for years. Mike Dixon at Fstoppers: “ISO isn’t a component of the exposure. It’s simply amplifying the sensor values,” like turning up a radio, “static and all”^[9]. LibRaw, the company behind RawDigger: “ISO cannot be an element of exposure; it’s determined through the exposure … and is ‘applied’ after exposure”^[10].

This matters because it predicts the second frame. Pushing the ISO brightened the couch and its noise together, in proportion, and did nothing to the window, because the window was already delivering more light than the sensor could count.

The noise in the second frame had a cause, and it was not the ISO setting. Light arrives as photons, and photons arrive at random. The count at a pixel follows a Poisson distribution, whose standard deviation is the square root of its mean, so a pixel that collects 100 photons carries a noise of about 10, a signal-to-noise ratio of 10; one that collects 10,000 reaches 100; one that collects 4 has a ratio of 2^[11]. Noise does not get added to a dark image. A dark image, seen honestly, has the noise built in. The corner under the couch delivered few photons in a hundredth of a second regardless of what number was set, and brightening the result afterward brightened the uncertainty along with it. “It’s the reduction in exposure that increases the noise, because it means you’re capturing less light”^[7].

There is one genuine reason to set a higher ISO in the camera rather than brightening later, and it is about electronics rather than light. On many sensors the ISO setting applies analog gain before the signal is digitized, and “if you double the scale of the signal, then any noise introduced later has half the impact”^[12]. On sensors that are nearly ISO-invariant (Butler names the Panasonic S1R II, Sony’s FX cinema bodies, and Nikon’s ZR in its raw mode) that benefit has been engineered out, and raising ISO gains nothing you could not do in software while still spending a stop of highlight room per stop^[12]. Either way, the fix for the dark couch was more photons: more seconds, or a wider opening. Only the two real dials can supply them.

If closing the aperture buys depth, the instinct is to close it all the way. Physics charges for that too. A lens cannot focus a point of light to a point; it focuses it to a disk, the Airy disk, whose diameter is 2.44 times the wavelength times the f-number^[13]. For green light at 550 nm that is about 1.3 micrometres per unit of f-number^[14]: at f/8 the smallest spot the lens can draw is roughly 11 µm wide, at f/16 about 21 µm. Once that spot is wider than about two of the sensor’s pixels, the pixels are no longer the limit; making them smaller “would not significantly increase the captured image resolution”^[13].

The pixels are a known size. A 24-megapixel full-frame sensor spreads 6,000 pixels across 36 mm, so each is 6.0 µm; the 45-megapixel EOS R5 fits 8,192 across the same width, 4.4 µm each^[15]; the iPhone’s main sensor has 1.22 µm pixels that it bins into 2.44 µm quads for most photos^[16]. Two-pixel widths of 12, 8.8 and 4.9 µm are reached near f/9, f/6.5 and f/3.6. That is why phone cameras do not have an aperture ring. It is also why the real-estate guides land on f/8 to f/11^[20]: on a 24-megapixel body, that is the last stop before diffraction starts taking back the sharpness the depth was supposed to buy.

Aperture earns its place as a creative dial because it decides how much of the room is in focus, and the formulas for that are old and public^[22]. With a 24 mm lens on a full-frame sensor focused on a chair three metres away, and the usual 0.03 mm circle of confusion, f/2.8 holds sharpness from about 2.1 to 5.3 m; f/8 holds it from 1.3 m to infinity. A living room is five or six metres deep. At f/2.8 the bookshelf on the far wall is soft and the coffee table in front of you is soft; at f/8 everything the room contains is sharp, at the cost of three stops of light, which the shutter now has to pay back in seconds.

Every trade so far ends at the shutter. The room needs f/8 for depth; f/8 at ISO 100 needs a second of light; ISO cannot shorten that second without making the noise visible and spending highlight room. A second is far past what hands can hold. The usual rule for handheld sharpness is a shutter speed of at least one over the focal length, 1/24 for a 24 mm lens^[17]. Stabilization moves the line: Canon rates the EOS R5 with its 24–105 at eight stops by the CIPA standard^[15], which on paper turns 1/100 into about two and a half seconds at the long end, though the rating is for camera shake and says nothing about the person who walks into the frame. The real-estate guides say the obvious thing: put it on a tripod, keep ISO at 100 to 400, and let the shutter run from 1/60 down to half a second or longer^[20]. A tripod is the least expensive object in photography that changes the result this much. It converts time, which an empty room has in unlimited supply, into light.

The window still does not fit. The same guides bracket, taking frames two stops under and over and merging them^[20]. That is the oldest answer to a scene wider than a sensor, and it is what the phone has been doing without asking. Google’s HDR+ pipeline, described in a 2016 paper, captures a burst of frames at a constant exposure “low enough to avoid blowing out highlights,” then aligns and merges them so the shadows accumulate photons across frames while the window is protected in each one^[18]. Apple’s Deep Fusion shoots a short underexposed frame, three normal ones and a long one, then merges them with neural networks tuned to the sensor’s noise^[19]; since 2022 the Photonic Engine applies that merge earlier in the pipeline^[16]. When my phone photographed the same room and got the window and the couch in one picture, it was not because the phone has a better sensor. Its pixels are a fifth the size. It took the bracket for me and never mentioned it.

Say that plainly, because the gear conversation around photography is mostly a conversation about money. The thing that fixed this room was a tripod, which costs less than a dinner out, and an understanding of which dial does what, which costs nothing. The phone solves the same problem for free by spending computation instead of time. The camera solves it by spending time instead of computation. Neither is more honest than the other. One of them tells you what it did.

Dixon’s objection to the triangle is that “it was simply three settings placed in a triangle for no other reason than a triangle has three sides”^[9]. That is fair, and it is also why the triangle survives: it is a map of the three dials a camera has, and a beginner needs a map of the dials. It fails as a map of the physics. Two of the dials move light and trade against depth and motion. The third moves a number and trades against noise and highlights. None of them touches the gap between a room and a window, which is set by the sun and the sensor and closed only with time, or with more than one frame.

In Chicago the sun on the shortest day reaches 24.7 degrees above the horizon at noon and is gone by 4:23 in the afternoon, after nine hours and eight minutes^[21]. Winter light comes in low and sideways, the room gets darker, the view stays bright, and the gap widens. The album will have winter in it. So the plan for the next frame is the one the arithmetic wrote: tripod, f/8, ISO 100, a second for the room, a sixtieth for the window, and a merge. Or the phone, which will do the same thing in a second and tell me nothing, which is fine for an album and not fine for learning what happened.