Costs
Part of What a photograph hides in a lookalike comparison: distance, light, expression
Ordinary image errors sit behind most lookalike comparison mistakes
Lookalike comparison mistakes start in the image handling: equal-width scaling, silent stretching, mirroring, retouching and upscaled detail.
Most errors in reading a comparison happen before anyone looks at a face. They happen in the handling of the two image files: how they were cropped, scaled, sharpened, graded and saved. These are boring, mechanical steps, and each one leaves a signature that a viewer reads as anatomy.
What to take away
- Scaling two images to the same width makes two different head sizes look like one, which is the single most common setup error.
- A non-uniform stretch to fill a layout changes face width without changing anything else, and it is invisible unless you check the aspect ratio.
- Mirroring one side of a comparison is a real intervention, because faces are not symmetric and the flipped version reads as a different person.
- Heavy compression and upscaling both invent detail, and invented detail tends toward the average face rather than either original.
Mistake one: matching on width instead of on the head
Two portraits pasted into a template are usually scaled to the same width. If one was shot tighter than the other, the two heads now occupy the frame identically while being wildly different sizes in life. Every proportion the viewer compares is then a proportion within a frame, not within a head.
Equal width vs pupil scaling
Equal width
- Head size
- Frame-identical
- Proportions
- Within frame
- Real size
- Erased
Pupil distance
- Head size
- Life-accurate
- Proportions
- Within head
- Real size
- Preserved
The fix is to scale on something anatomical and stable, such as the distance between the pupils, and to state that you did. Almost no circulating comparison does.
Mistake two: ignoring the aspect ratio
Layouts have fixed slots, and images get stretched to fill them. A five percent horizontal stretch will not look wrong to anyone, and it widens a face by five percent. Two images stretched by different amounts differ by the sum, and the difference sits precisely in the dimension people use to judge face shape.
Five percent stretch
5%
Horizontal stretch widens a face by five percent
Different stretches sum their errors
Distortion sits in face-shape judgment
Check the ratio of the original file against the ratio as displayed. If they disagree, the comparison is over before it starts.
Mistake three: treating a mirrored image as the same face
Faces are not symmetric, and the mirrored version of a familiar face reads as subtly wrong to people who know it well. Front-facing cameras, some scanning workflows and casual editing all flip images without announcing it. A comparison in which one side is flipped is comparing one real face to a face that has never existed.
The tell is text: any writing, badge or parting in the image will be reversed. Where there is no text, there is no tell.
Mistake four: forgetting the lens before any of this
All of the above sits on top of the largest factor: where the camera stood. The change in apparent proportion with viewing distance reshapes a face far more than any editing step. Close in, the nose enlarges and the ears recede. Far off, the face flattens and broadens.
Neither the distance nor the focal length is recorded in a shared image.
That is why the camera work is treated as the first question rather than a technicality on this site. A comparison that has not addressed it has not started.
Mistake five: reading retouching as skin
Commercial portraiture is retouched as a matter of course. The practice of altering a photographic image after capture is as old as photography itself.
Modern retouching smooths texture, evens tone, narrows the jaw, and enlarges the eyes. It moves toward a shared house preference, so two heavily retouched portraits approach each other. Both are pulled toward the same target.
The practical consequence is that the most polished images in any comparison are the least usable, and they are usually the ones chosen, because they are the ones that look good in a layout.
Mistake six: trusting an upscaled image
Enlargement software does not recover detail. It generates plausible detail from what it has learned, and what it has learned is the general shape of faces. Feed it two degraded images and it will fill both with the same statistical guesses, which is a machine for producing resemblance.
| Handling step | Direction of the error |
|---|---|
| Scale to equal width | Erases real size difference |
| Stretch to fill a slot | Widens or narrows the face silently |
| Mirror one side | Produces a face that never existed |
| Retouch | Pulls both faces toward one house ideal |
| Heavy compression | Removes fine texture, keeps coarse shape |
| Upscale or restore | Adds invented detail drawn from an average |
Every row points the same way: toward agreement. None of these steps can make two faces look less alike, which is worth sitting with. The pipeline is biased, and a quiz built from images that went through it inherits the bias.
The check that covers most of it
Ask for the originals. If the two source files cannot be produced with their dimensions and dates, the comparison cannot be assessed. The honest description is that somebody made a picture of two people, not a comparison of two faces.
That is why nothing is published here as a pairing, and why the format itself does most of the persuading.
Common questions
Does any of this apply to a comparison I make privately?
Yes, and it is easy to test on your own photographs, where you already know the answer.
Is a screenshot of a video frame usable?
It carries compression artifacts, motion blur and whatever the broadcast grading did. It is the weakest kind of source and the most common one.
How can I tell if an image was upscaled?
Often you cannot, which is the problem. Suspiciously clean detail in an otherwise old or small image is the usual hint.
