An anonymous peer reviewer worked through a manuscript, looked at one of its figures, and recognized one of the authors. The face was a reflection in another person’s eye. Across the study’s bystanders, those extracted faces ran 27 to 36 pixels wide and covered roughly 322 pixels apiece, and that accidental recognition became the paper’s second experiment.
What Jenkins and Kerr Actually Photographed
Rob Jenkins, then at the University of York, and Christie Kerr at the University of Glasgow published Identifiable Images of Bystanders in PLOS ONE on December 26, 2013. Eight volunteers, three women and five men with a mean age of 23.5, were run in two groups of four. Each one sat once as the photographic subject and stood three times as a bystander in somebody else’s portrait.
The camera was a Hasselblad H2D 39-megapixel body with a 120mm macro lens, shot at ISO 50, f/8, and 1/250 second on manual focus, from about 1 m away. Two Bowens DX1000 heads with dish reflectors sat side by side roughly 80 cm behind the camera, aimed upward. That upward tilt was deliberate, and the paper says why: it excluded the catchlight. Two more DX1000 heads in soft boxes went behind baffles on either side of the subject, lighting the bystanders rather than the face. Three volunteers, the photographer, and Jenkins himself stood in an arc about a meter out.
The files measured 5,412 by 7,216 pixels, just over 39 million in total. The subject’s face, hair excluded, took up about 12 million of them. The iris took about 54,000, which the authors note is under 0.5% of frontal face area. Out of that iris, they cropped rectangles 27 to 36 pixels wide by 42 to 56 pixels high, framed head and shoulders like a tiny passport photo. Average face area inside those crops was 322 pixels. Pilot work put the gap between a subject’s face and a reflected bystander’s face at a factor of around 30,000.
The processing was minimal. The crops were rescaled to 400 pixels high using bicubic interpolation, then run through Auto Contrast in Photoshop. Bicubic interpolation is a resampling filter. It smooths the pixel grid and it adds no information whatsoever. Nothing in this study recovered detail that was not already sitting in the sensor’s output.
Two experiments followed. In the first, observers saw each reflection paired with a studio enrollment photo taken about 14 months earlier and judged same person or different person across 16 pairs, with chance at 50%. Foils were not random; for every target, three independent raters picked the closest lookalike out of a cohort of 108 undergraduates, matched on gender, ethnicity, build, hair, and approximate age. Sixteen observers from unrelated departments scored 71%. Sixteen classmates of the targets scored 84%. In the second experiment, ten people who had known Jenkins for a mean of 18.2 years were shown a six-face lineup of reflections and asked to name anyone they recognized. They named him 90% of the time and misidentified strangers 10% of the time.
The Popular Version and What the Paper Says
The study gets recycled every few years as proof that zoom and enhance is real, usually attached to a still from a security camera. The popular version has a reasonable foundation. Faces genuinely were recovered from an eye, they genuinely were identified, and the paper genuinely proposes forensic use, specifically for crimes in which victims are photographed and the camera is later seized.
The record is narrower than the retellings, but not as narrow as the debunkings. The enhancement added nothing. The pixels, though, did carry identity. Experiment 1 asked whether the reflections could support identity discrimination at all, and observers who had never seen the bystanders matched them to comparison photographs 71% of the time against a 50% floor, with the foils chosen as the most similar-looking person out of a cohort of 108 students. The effect was enormous, d = 1.91, and the authors describe it as confirming that the bystanders “could be reliably distinguished from similar foils.” What familiarity bought was more: 84% on the same task, and, in Experiment 2, the ability to name a face outright rather than match it against a photograph you were already holding. Jenkins and Kerr close the paper by saying so outright: “obtaining optimal viewers – those who are familiar with the faces concerned – may be more important than obtaining optimal images.”
The preconditions are stated just as plainly. Clear corneal reflections, the authors write, “require the subject’s face to be in focus, and viewed from a roughly frontal angle under good lighting. They also require high image resolution in order for bystanders’ faces to be properly resolved.” Read the methods against that list, and the shoot was engineered to satisfy every clause: studio strobes at 1/250 second, ISO 50, a macro lens at a meter, manual focus, a frontal passport pose, and separate lights aimed at the bystanders so they had something to reflect. Nowhere in the paper do the authors invoke CCTV or surveillance footage; the seized camera is the scenario throughout.
What 100 Megapixels Actually Buys You
Thirteen years of sensor progress should have moved those preconditions. The Fujifilm GFX100 II writes 11,648 by 8,736 pixels, which is 101,756,928 in total, or 2.61 times the 2013 file. Resolution scales with the square root of that, so 1.61 times linear. At identical framing, a bystander crop that ran 27 to 36 pixels wide in 2013 comes out around 44 to 58 pixels wide today, and 322 pixels of face area becomes about 839.
Put that against the floor for familiar recognition. Jenkins and Kerr cite earlier work by Yip and Sinha finding that people identify familiar faces down to resolutions equivalent to 7 by 10 pixels. The 2013 reflections sat about 2.1 times above that floor linearly. A 102-megapixel file puts you about 3.5 times above it. That is a genuine improvement and it is an improvement in margin, not in kind. It does not convert matching into naming. What separated the strangers from the classmates was familiarity and not sharpness, since both groups looked at the same pictures, and the strangers still needed a comparison photograph in hand to work against.
Diffraction eats into the rest. The GFX100 II’s pixel pitch works out to 43.8 mm across 11,648 pixels, or 3.76 micrometers. An Airy disk in green light at f/8 is 2.44 times the wavelength times the f-number, which comes to 10.7 micrometers, nearly three of those pixels wide. Shoot Jenkins and Kerr’s aperture on a modern back, and the lens’s own blur circle is already spread across a three-pixel neighborhood. Open the lens up to cash in the resolution, and you surrender depth of field on a subject whose cornea is a bulge, not a plane.
The cornea is the real ceiling. Ko Nishino and Shree Nayar at Columbia modeled it in 2004 in The World in an Eye, giving the average adult apex radius of curvature as 7.8 mm and the limbus radius as about 5.5 mm. That makes the front of the eye a convex mirror roughly 11 mm across, and in every eye-camera configuration Nishino and Nayar model, the cornea’s field of view exceeds and contains the roughly 120-degree human field of view they use as the benchmark. Everything in front of your subject, plus a margin they cannot themselves see, gets compressed into that disc. Sensor development samples that miniature more finely. It does not make the miniature any bigger.
Why AI Enhancement Makes the Problem Worse
Jenkins and Kerr made one forecast in the discussion, that pixel count per dollar had been doubling roughly every twelve months, and that phones would soon routinely carry cameras above 39 megapixels. The forecast landed. Samsung’s ISOCELL HP2 is a 200-megapixel sensor, and it arrives in a form that dismantles the technique rather than enabling it. Its pixels are 0.6 micrometers, about the width of a wavelength of green light, and the same Airy calculation at the fast apertures phone lenses use puts the blur circle across roughly four of them. Samsung’s own answer is Tetra²pixel binning, which merges up to 16 pixels into one and drops the output to 12.5 megapixels as the light falls. Full 200-megapixel capture is available if you ask for it in good light. The default path is not. Beyond the binning, phones run multi-frame and computational processing that varies by capture mode and by maker, and Samsung describes its own as improving sharpness and dynamic range. Whatever it does for the face, none of it is designed with a 30-pixel reflection inside the eye in mind, and its effect on detail at that scale is a property of the pipeline rather than of the sensor.
Modern denoising and upscaling tools are the sharper hazard, because they produce a confident, detailed, plausible face out of a smear. A model trained on a prior does not retrieve missing detail. It supplies what its training says should be there. The clearest statement of that comes from the researchers behind PULSE, a Duke University super-resolution method that in 2020 turned 16-by-16-pixel inputs into photorealistic 1024-pixel portraits. Their own model card says: “Not suitable for facial recognition/identification. PULSE makes imaginary faces of people who do not exist, which should not be confused for real people. It will not help identify or reconstruct the original image.” The next line goes further than Jenkins and Kerr’s own results support, arguing that the method demonstrates face recognition is not possible from low-resolution or blurry images, because it can generate visually distinct high-resolution faces that all downscale correctly to the same input.
That is why the 2013 paper’s bicubic-and-contrast pipeline is the honest one. It is forensically defensible precisely because it invents nothing, and if you point a generative upscaler at a corneal reflection you get a picture of a person who was never in the room. Where the field has genuinely moved is toward more views rather than more pixels. A 2023 preprint by Hadi Alzayer, Kevin Zhang, Brandon Feng, Christopher Metzler, and Jia-Bin Huang, published at CVPR 2024 as Seeing the World through Your Eyes, reconstructs a 3D scene beyond the camera’s line of sight from eye reflections across multiple frames of a moving person. Their listed failure mode is four or fewer images, or too little motion.
The Catchlight Is a Lighting Diagram
Jenkins and Kerr had to defeat the catchlight to see the room, tilting their main heads upward so a blown white blob would not sit on top of the information they wanted. In the studio, you run that decision the other way, and the same physics that makes bystanders faintly readable makes your lighting largely readable.
An 11 mm convex mirror facing whatever your subject faces reproduces every source in front of them at a scale that reports its apparent size, which is the source’s real size and its distance rolled into one number. A large softbox far back and a small one close in can leave the same mark. Shape narrows the modifier without naming it. A circle suggests an octabox, a parabolic, or a beauty dish, and a beauty dish often gives itself away with a dark core where the deflector sits. A vertical bar suggests a stripbox. A rectangle with dividing lines running through it is a window, and the lines are the mullions. A ring centered on the pupil is a ring light. A soft, shapeless brightening with no defined edge is a bounce or a scrim.
Count names the number of sources. Two catchlights mean two lights, or a light and a reflector, and the reflector’s is usually the dimmer one. Position inside the iris names the angle: a catchlight at eleven o’clock puts the key high and camera left, and one sitting low and centered means whoever lit it worked from underneath. Size relative to the iris names apparent source size, which does more than anything else to determine how hard or soft the light is. Every portrait you admire carries its own setup diagram in two small mirrors, and reading them is a faster education than any lighting breakdown someone else writes for you. Peter Hurley’s course on lighting the face covers this ground at length.
The pupil takes its name from pupilla, Latin for a young girl, because of the tiny reflected figure you see when you look closely into someone’s eye. Retouchers have been adding, moving, brightening, and cloning catchlights for as long as retouching has existed, which makes the catchlight both a piece of evidence and one of the most routinely fabricated features on a human face.
Shoot a headshot tomorrow at f/8 with focus placed on the near eye, then open the raw file and crop to the iris at 100%. A convex mirror throws an upright, shrunken, virtual image, so there is a small, badly lit, right-way-up photograph of you in there, holding a camera about a meter away. Badly lit is the operative phrase. Jenkins and Kerr did not leave that to chance: they put two softboxed heads behind baffles on either side of the subject purely to light the people being reflected, and aimed the main lamps upward so the catchlight would not sit on top of the thing they were trying to read. Reproduce that, and a stranger holding a picture of you would beat chance on average. Skip it, which is what an ordinary headshot does, and you are below the conditions the number came from. Someone who knows you might still say your name.
Lead image by Osmo Lundell, CC BY-SA 4.0. Source.
