A mirrored photo and the original hold the same picture, one in reversed left-to-right order: one is the version other people see, the other is the version you usually see yourself.
Neither version is more accurate than the other. Which one looks better is a question of which one you are used to seeing, not a question of which one is correct.
A camera records the scene as light arrives from the front of it. A mirror shows the same scene from the other side of the glass, with left and right exchanged. Put the two side by side and every feature sits in the same place on the frame, but the arrangement runs the other way.
That is the entire difference. The file on your disk is the camera's version; the file you get after a horizontal flip is the mirror's version. Neither was taken at a different moment, with a different lens, or from a different position.
Pictures of people survive the swap because a face is read as a whole. Text does not, because reading depends on the order the shapes are in. We put that to a test instead of asserting it: three pictures containing text were mirrored with the tool, then read by an optical character recogniser before and after.
| Picture | Text in it | OCR on the original | OCR on the mirrored copy |
|---|---|---|---|
| Text only | SALE ENDS FRIDAY | Read in full | YAOIRR e0h3 — no word recovered |
| Text plus logo | NORTHWIND COFFEE / ROASTED IN SMALL BATCHES | Both lines read in full | C]VIIWHTROVI, HOTA8 — no word recovered |
| House number plate | No. 42 / ELM STREET | Both lines read in full | T 39T2 M — no word recovered |
Five text tokens went in and five came back. None came back from the mirrored copies, even though the recogniser was still finding letter-shaped objects in them and returned glyph soup rather than nothing at all. Numbers are not spared: the 42 on the plate came back as 39T2.
A camera does not always store the picture the way it looked on the screen. It can store the sensor's arrangement and attach an orientation value telling the viewer how to turn it. Eight values are defined, and four of them describe a mirror rather than a rotation. The browser applies that value before the flip happens, so a photo can be mirrored twice without anyone pressing the button twice.
| Orientation value | What the viewer applies | Decoded size from a 600 × 400 file | Mirror in the chain? |
|---|---|---|---|
| 1 | Nothing | 600 × 400 | No |
| 2 | Mirror, left to right | 600 × 400 | Yes |
| 3 | Turn 180° | 600 × 400 | No |
| 4 | Mirror, top to bottom | 600 × 400 | Yes |
| 5 | Mirror plus a quarter turn | 400 × 600 | Yes |
| 6 | Quarter turn clockwise | 400 × 600 | No |
| 7 | Mirror plus a quarter turn, reversed | 400 × 600 | Yes |
| 8 | Quarter turn anticlockwise | 400 × 600 | No |
The last column is the one that matters here. Values 2, 4, 5 and 7 put a reflection in the chain before the tool does anything, so what appears on screen is already one reflection away from the file on disk.
Everything below was measured on this site's flipper in a desktop Chromium browser, with each file loaded through the page's own picker and exported as PNG so that nothing was lost to compression. No figure here was copied from another site or inferred from a format specification.
The text test used three pictures rendered at 1000 by 240, 1000 by 320 and 700 by 380 pixels, each mirrored once with Flip Horizontal and exported. Source and mirrored copy were then read by the same recogniser with the same settings, so the only variable in play was the mirror itself.
Five of five expected tokens were recovered from the originals — SALE ENDS FRIDAY, NORTHWIND COFFEE, ROASTED IN SMALL BATCHES, No. 42, ELM STREET — and zero of five from the mirrored copies. The mirrored outputs were not blank. The recogniser returned strings such as YAOIRR e0h3, C]VIIWHTROVI and 39T2, which is what mirrored letterforms amount to: the same shapes, facing the wrong way and in the wrong order, so a reader that expects order can only guess at them.
The orientation sweep used one 600 by 400 source saved eight times with values 1 through 8. Each file was loaded, the decoded pixels were captured from the page's own canvas, and the result was compared against all eight possible transforms of the stored image. Every value matched its expected transform with a mean absolute pixel difference of 0.0, and values 5 to 8 decoded to 400 by 600. The browser turns the picture rather than merely relabelling it.
Four of those eight values contain a reflection: 2 and 4 are single-axis mirrors, 5 and 7 combine a mirror with a quarter turn. That is how a photo ends up reflected before you touch it, and why a single Flip Horizontal can leave a picture looking the same way it started.
Sharpness was checked with the variance of the Laplacian, a standard measure of edge energy, on one picture before and after a mirror: 1368.6039163232124 against 1368.6039163232126, a difference of 2.3e-13. That is floating-point noise rather than image degradation, and it is what a pure relocation of pixels looks like when you measure it.
What we did not test, and will not claim: photographs taken with a phone's front camera, where the mirroring behaviour belongs to the phone and not to the file; apps that mirror a preview while leaving the saved frame alone; HEIC files, which the picker does not accept; recognisers other than the one built into Windows, which produced the strings above; and how a person reads a mirrored face, which is a question for perception research rather than for a pixel count.
The value sits in the file's EXIF block, which any EXIF reader will display as Orientation. It is metadata, not pixels, which is why a photo can look correct in one program and sideways in another: the program either honours the value or ignores it. Export the picture through a browser tool and the value is applied to the pixels on the way in, so the exported file needs no flag at all.
Mirror a photo now — free, no upload, no watermarkNeither is more real than the other. One is the version other people see, the other is the version you usually see yourself. The pixels are the same in both; only their left-to-right order differs.
Reading depends on the order the letterforms are in, and a mirror reverses that order while also flipping each shape. In our test, five text tokens were recovered from the originals and none from the mirrored copies, including the digits.
Yes. Four of the eight EXIF orientation values describe a mirror rather than a rotation, so a file can arrive at the tool already reflected. One more horizontal flip then produces a picture that has been mirrored twice.
No. A mirror relocates pixels without blending them, so no detail is averaged away. We measured edge energy before and after on the same picture and the two values agreed to within floating-point noise.
The recogniser was reading reversed glyphs and guessing at what they were, so 42 on the plate came back as 39T2. The digits themselves did not change; their shapes were reversed and then misread.