Why can't you 'enhance' a blurry photo?

The full story

The episode's narration, word for word. Numbers link to the sources below; timestamps play that moment.

Enhance 0:00

You know the scene. Detectives crowd around a screen, a grainy security-camera frame, and somebody says the magic word: enhance ¹. They zoom in again and again, until they can read a number plate, or see the killer reflected in the victim's eye ¹. In real life, zoom in on a security frame and you get bigger squares. So why can't you enhance a blurry photo?

Deckard's photograph 0:27

One of the most famous versions is in Blade Runner, from 1982 ². Harrison Ford's Deckard feeds a photograph from a suspect's hotel room into a machine called an Esper and talks to it: enhance two-two-four to one-seven-six ^2,3. The machine glides deeper into the picture, around corners, until a reflection reveals a woman, Zhora ³. It treats a flat photo like a room you can walk through, which, as one writer put it, breaks the laws of physics ³. So what does a real photo actually hold?

Buckets of light 1:06

A digital camera's sensor is a grid of tiny light detectors, the pixels ⁴. Each one reports a single number: how much light landed on its patch ⁴. It doesn't record where on the patch the light fell. So anything smaller than a pixel is averaged into one shade of grey ⁴. The spacing of that grid sets the finest detail the camera can ever capture ⁴.

A face at ten metres 1:32

A common two-megapixel dome camera has a two-point-eight millimetre lens and a sensor nineteen hundred and twenty pixels wide ⁵. The datasheet doesn't name its sensor, so we'll assume a common one of that size, with pixels two-point-nine thousandths of a millimetre apart ⁶. Near the middle of the picture, ten metres away, each pixel covers about ten millimetres of the scene, by our calculation ^5,6. A face is about sixteen centimetres wide, so it lands on about fifteen pixels ⁷. At twenty metres, about eight ⁷.

What the installers ask for 2:10

The security industry has a standard for this, IEC sixty-two six seven six, part four, built on how many pixels land across a face ⁷. To identify someone with high certainty, it asks for fifteen hundred pixels per metre ⁷. Our face at ten metres gets about ninety-seven per metre ⁷. That's enough to see that a person is there, not who it is ⁷. The camera's own datasheet says it can identify people out to about four metres ⁵. That uses the older, looser scale of two hundred and fifty pixels per metre, and our sum agrees: three-point-nine metres ^5,7,8. For the standard's highest level of certainty, you'd need to be within about sixty-four centimetres, by the same sum ⁷.

Eleven million greys 3:00

Here's the core problem. Imagine one pixel that saw four tiny details, each one of the camera's two hundred and fifty-six shades from black to white, and averaged them to a mid-grey ⁴. How many different arrangements give exactly that grey? We counted: over eleven million. Every one of them becomes the same pixel ⁴. Going backwards, from the grey to the one arrangement that was really there, isn't hard; it's impossible, because nothing in the pixel says which ⁹.

Stripes that lie 3:35

To capture a pattern of light and dark, you need at least two samples for every light-and-dark pair ⁴. Finer than that, and the pattern doesn't just blur; it can turn into a false, coarser pattern that was never there ⁴. That's called aliasing, and in a single frame, nothing in the data tells it apart from a real pattern ⁴.

Then buy more pixels 3:56

So why not just buy more pixels? Because the lens has a limit of its own ¹⁰. Light is a wave, and when it squeezes through the opening of a lens, it spreads ¹⁰. Even a perfect lens turns a point of light into a small bright disc with faint rings around it ¹⁰. When every part of a camera is perfect, that spreading is what limits sharpness ¹¹.

Lord Rayleigh's rule 4:20

In the nineteenth century, the physicist Lord Rayleigh set the rule still used today ¹⁰. Two points of light can just be told apart when the centre of one blurry disc sits on the first dark ring of the other ¹⁰. Much closer, and they merge into one blob ¹⁰. The angle comes out as one-point-two-two times the wavelength of the light, divided by the width of the lens opening ¹⁰.

Lens versus pixel 4:47

For our camera, the lens opening works out to one-point-seven-five millimetres wide ⁵. In green light, Rayleigh's rule says it can separate details about four millimetres apart at ten metres ^10,5. Each pixel covers ten, so for this cheap camera, the pixels are the bottleneck, not the lens ^5,6. But the lens sets a floor no number of pixels can get under ¹¹. Detail much finer than that is smeared out before it ever reaches the sensor ^10,11.

The eye in the photo 5:18

What about the killer in the victim's eye? In 2013, two researchers, Rob Jenkins and Christie Kerr, actually tried it ¹². With a thirty-nine-megapixel studio camera a metre away, the faces of bystanders reflected in volunteers' eyes came out only about three hundred pixels in all, in head-and-shoulders crops about thirty pixels wide ¹². People who knew them matched them correctly eighty-four percent of the time, and friends of one researcher could pick him out by name ¹². So the trope's physics is real; its camera isn't ¹². The reflected faces covered about a thirty-thousandth of the area of the subject's face, and on our security camera at ten metres, that's less than a tenth of a pixel ¹².

Real enhancement 6:07

Not all enhancement is fake. Forensic analysts remove noise, adjust brightness, steady shaky video and average frames together ¹³. If you know exactly how an image was blurred, you can partly undo it, because the detail was smeared, not lost ¹⁴. Some phones take a burst of shots while your hand trembles, so each frame samples the scene at slightly different points, and merging them really does add resolution ¹⁵. All of these work with measurements the camera actually made ¹⁴.

Imagining the pixels 6:40

Then came artificial intelligence ¹. In 2018, a student team at Duke University built an upscaler that, in the university's words, makes a best guess at the missing details, essentially creating new information that wasn't there before ¹.

Pulse 6:58

Two years later, Sachit Menon and colleagues at Duke published a method called PULSE ⁹. It takes a face sixteen pixels across, about our face at ten metres, and returns one over a thousand pixels across ⁹. That's four thousand and ninety-six output pixels for every one that was measured ⁹. PULSE doesn't sharpen the input; it searches for realistic faces that, shrunk back down, match it ⁹. Its authors wrote that it 'makes imaginary faces of people who do not exist', and 'will not help identify or reconstruct the original image' ⁹.

Many faces, one blur 7:37

Like our eleven million greys, a small blurry face is consistent with an enormous number of sharp faces ^9,16. As one AI artist put it, there are probably millions of possible faces that reduce to the same pixel pattern, all equally correct ¹⁶. An upscaler has to pick one ⁹. It picks one that looks like the faces it learned from, so it's a confident guess, shaped by its training, not a recovered truth ^1,16.

The weekend it went viral 8:06

In June 2020, people fed an easy-to-run version shared online a pixelated photo of Barack Obama, and it returned the face of a white man ¹⁶. The PULSE team said the bias was likely inherited from the faces the underlying model was trained on, though there could be other factors ¹⁶. Others, like the AI artist Mario Klingemann, argued that the way it searched was to blame ¹⁶. The tool showed exactly what an upscaler does: it fills the gap with a guess, and the guess has to come from somewhere ¹⁶.

The courtroom 8:42

In 2024, in a murder case near Seattle, the defence wanted to show phone video improved by AI software from Topaz Labs ¹⁷. Judge Leroy McCullogh refused ¹⁷. He wrote that the tool used opaque methods to represent what the AI model thinks should be shown ¹⁷. A forensic video analyst for the prosecution said every pixel in the enhanced video was new, giving the illusion of clarity ¹⁷.

Restore, don't invent 9:10

A 2026 review written for Interpol puts the core principle simply: enhancement must not create information that wasn't in the original ¹⁴. Methods that reverse a known blur or noise restore the signal ¹⁴. Methods that generate detail can hallucinate, adding plausible but not necessarily factual detail ¹⁴. In one study it cited, cleaning faces up with a super-resolution model actually made face recognition less reliable ¹⁴.

A note about our pictures 9:43

A confession, while we're here. Some of the pictures and moving shots in this video, like the desks and the empty rooms, were made by AI tools, and the voices you're hearing are synthetic too. They're plausible, not records: no camera ever saw those exact rooms. That's fine for setting a mood, and it's exactly why such pictures don't belong in evidence. The diagrams are different: every number in them is either computed or taken from a source we cite.

The photographer who got it right 10:14

The film that got it right came out in 1966: Michelangelo Antonioni's Blow-Up ^18,19. A London photographer thinks his pictures of a park may show a crime, so he enlarges them, then enlarges them again ^18,19. As the critic Roger Ebert described it, we end up looking at arrangements of light and shadow, dots and blurs, that may show what ¹⁸? Maybe a gunman, maybe a body ¹⁸. The more he enlarges, the coarser the grain gets, until the body is only a suggestion in it ^18,19.

So, can you enhance? 10:50

So can you enhance a blurry photo? You can clean it, steady it, and undo a known blur, and with many frames you can genuinely gain detail ^15,14,13. But detail that never reached the sensor is gone, and anything that appears there is a guess ^9,14. The squares are the truth; the enhance button is a story. Next time, another thing films show that light won't let you see: why can't you see a laser bolt flying past?

Sources

Every factual claim in the episode is tied to one of these. Spotted an error? Tell us.

1. Robin A. Smith, "Duke Team Attempts a Real-Life Version of CSI 'Zoom and Enhance'",… — today.duke.edu
2. Abe J. Riesman, "Blade Runner: The Final Cut", The Harvard Crimson, 15 November 2007 — thecrimson.com
3. Vishal Wagh, "Photo Reveal: The Most Misunderstood Scene in 'Blade Runner'", No Film… — nofilmschool.com
4. Tucsen, "Understanding Nyquist Sampling: Balancing Optical and Camera Resolution", 4… — tucsen.com
5. Hikvision, DS-2CD2123G2-IU 2MP AcuSense Fixed Dome Network Camera, datasheet (2024) — assets.hikvision.com
6. Sony Semiconductor Solutions, IMX307LQD/LQR product flyer, diagonal 6.46 mm (Type… — sony-semicon.com
7. Axis Communications, "Pixel density based on IEC 62676-4:2025", white paper, September… — whitepapers.axis.com
8. Axis Communications, "Perfect Pixel Count: Meeting your operational requirements",… — axis.com
9. Sachit Menon, Alexandru Damian, Shijia Hu, Nikhil Ravi and Cynthia Rudin, "PULSE:… — arxiv.org
10. OpenStax, College Physics 2e, section 27.6 "Limits of Resolution: The Rayleigh Criterion" — openstax.org
11. R. Nave, HyperPhysics, "The Rayleigh Criterion", Georgia State University — hyperphysics.gsu.edu
12. Rob Jenkins and Christie Kerr, "Identifiable Images of Bystanders Extracted from… — journals.plos.org
13. SWGDE, "Best Practices for Digital Forensic Video Analysis", 18-V-001 version 1.1, 22… — swgde.org
14. Zeno Geradts, Stijn van Lierop and Meike Kombrink, "Interpol review of forensic image… — pure.uva.nl
15. Bartlomiej Wronski et al., "Handheld Multi-Frame Super-Resolution", ACM Transactions… — arxiv.org
16. James Vincent, "What a machine learning tool that turns Obama white can (and can't)… — theverge.com
17. Tim Stelloh, "Washington state judge blocks use of AI-enhanced video as evidence in… — nbcnews.com
18. Roger Ebert, "Blow-Up" (Great Movies), RogerEbert.com, 8 November 1998, archived — web.archive.org
19. Wikipedia, "Blowup" (film: year, director, synopsis; cross-check of [10]) — en.wikipedia.org

Researched and scripted with AI assistance, fact-checked claim by claim, with synthetic narration and diagrams drawn in code. How we make episodes.