Which Photos Can AI Read Well? One Sign at Three Qualities

By Junaid Aslam. 6 min read. Updated .

In this test, the analyzer read the large sign text on all nine runs, at every quality. The small background names moved with the file: the original and the JPEG read most of them, and the crop read none. The JPEG also read 'MEDICAL' as 'SURGICAL' in the hospital name on all three runs.

One CC0 photo of a no-parking sign in Plaridel, Philippines, by JWilz12345. Three files, all 1200x900: the original as WebP at quality 0.85 (165,110 bytes), a JPEG at quality 0.2 (56,741 bytes), and a center crop of 60% of the width and height as WebP at quality 0.85 (94,342 bytes). Each file went through the production AI Image Analyzer on 2026-11-16, Ask mode, empty question box. The outputs below are trimmed to the sections named.

Original (WebP, quality 0.85)

Real output, trimmed to the Text visible and Photo quality sections:

## Text visible in the image
"NO PARKING"
"BETWEEN SIGNS"
"Hindi ka tanga,"
"Nakakaintindi ka."
"CITYSTATE CITY CASH"
"BancNet"
"VISA"
"MARCELO-PADILLA"
"RENS & MEDICAL HOSPITAL, CORP."

## Photo quality notes
The image is sharp and well-exposed, with good clarity on the text. There is minimal noise and no significant blur.

The main sign is exact. The model misread the hospital name. The sign in the photo reads '...S & MEDICAL HOSPITAL, CORP.' under 'MARCELO-PADILLA', so the model's 'RENS' is a guess at the cut-off first word. The three repeat runs read 'RENS & MEDICAL' twice and 'EYE, EAR, NOSE & THROAT' once.

Heavy JPEG (quality 0.2)

Real output, trimmed to the Text visible and Photo quality sections:

## Text visible in the image
"NO PARKING", "BETWEEN SIGNS", "Hindi ka tanga, Nakakaintindi ka.", "CITYSTATE", "CITY CASH", "VISA", "MARCELO-PADILLA", "EYE & SURGICAL HOSPITAL CORP.", "EYE AND CATARACT", "AMBULATORY UNIT", "GLAUCOMA CENTER", "TC"

## Photo quality notes
The image is sharp and clear with accurate exposure. There is minimal digital noise, and the depth of field is sufficient to keep both the sign and the immediate surroundings in focus.

The main sign is still correct. The model's background reading is wrong: 'MEDICAL' became 'SURGICAL', and all three repeat runs made the same swap. A 'TC' token appears that the original did not list. The quality notes call the file sharp and clear. The file is 56,741 bytes, about a third of the original.

Center crop (60%)

Real output, trimmed to the Text visible, Other objects, and Photo quality sections:

## Text visible in the image
"NO PARKING"
"BETWEEN SIGNS"
"Hindi ka tanga, Nakakaintindi ka."

## Other objects
In the background, there is a portion of a street with a vehicle visible in the distance. Behind the sign, the partial view of a motorcycle tire and a blue plastic chair are visible. The sign is supported by black metal poles and structural supports.

## Photo quality notes
The image is in sharp focus, particularly on the text of the sign. The exposure is balanced, though the lighting conditions are slightly overcast. There is a small amount of digital noise typical for outdoor photography.

The main sign survived the crop, even with the edges cut. The background names are gone, and the model describes the sign in isolation. The motorcycle tire and the blue chair are real, so those details are correct. The noise note is a judgment; the crop file is a clean WebP.

The three files side by side

QuestionOriginalHeavy JPEG (0.2)Center crop (60%)
Main signNO PARKING / BETWEEN SIGNSSameSame
Tagalog lineCorrectSameSame
Background namesMost names on 3/3 runs, brand name wrongMost names on 3/3 runs, 'MEDICAL' became 'SURGICAL'No names on 3/3 runs
Photo quality notesSharpSharp, no visible noiseSharp, minimal noise
File size165,110 bytes56,741 bytes94,342 bytes

Counts below are out of three repeat runs per file. 'Read' means the item appears in the Text visible section:

Text itemOriginal WebPJPEG 0.2Crop 60%
NO PARKING3/33/33/3
BETWEEN SIGNS3/33/33/3
Tagalog line3/33/33/3
CITYSTATE CITY CASH3/33/30/3
BancNet3/30/30/3
VISA3/33/30/3
MARCELO-PADILLA3/33/30/3
MEDICAL in the hospital name2/30/30/3
AMBULATORY UNIT3/33/30/3
GLAUCOMA CENTER3/33/30/3
Stray 'TC' or 'TO' after the signs0/33/30/3

The source settles the hospital name. The sign reads '...S & MEDICAL HOSPITAL, CORP.' under 'MARCELO-PADILLA', and the business is Marcelo-Padilla Children's and Medical Hospital Corp. 'SURGICAL' is wrong. The original file got 'MEDICAL' on two of three runs, and the JPEG got it on none. No run got the cut-off first word right: 'RENS' and 'EYE, EAR, NOSE & THROAT' are both guesses.

The yellow sign at the right edge is partly hidden behind a vehicle awning. Only the 'T' is clear in the photo, so the source cannot confirm the 'TC' the JPEG added.

What this means

In this test, large text was stable. All nine runs read 'NO PARKING', 'BETWEEN SIGNS', and the Tagalog line, at every quality. If you need one short, high-contrast line read from a photo, this test suggests it will survive compression and a crop.

Small text was less stable. The hospital name moved between runs at original quality and became 'SURGICAL' on every JPEG run. The correct word is 'MEDICAL'. Check a proper noun against the source before you publish it.

Compression had a measurable effect on that name in this test. The JPEG read 'SURGICAL' on 3/3 runs, while the original read 'MEDICAL' on 2/3. The same line also changed at original quality, so the file is not the only source of error. The lower-detail file gave the same wrong word every time.

The model cannot judge compression. It called the quality-0.2 JPEG sharp and clear on every run. Its quality notes describe the visible image, not the file. The model does not know the JPEG quality or the byte size. For a file quality check, use the file size and the encoder settings.

A crop changes the answer. The tight crop kept the sign and dropped every background name on all three runs. The model then described the motorcycle tire and the blue chair, which are real, but it could not name the place. Send the full frame when context matters.

For alt text, use the Alt Text Generator. For camera and GPS tags, use the EXIF Viewer.

Tested on

2026-11-16 against the production AI Image Analyzer, with the repository at c36e31f for the first runs and a7eb773 for the repeats. Nine requests in Ask mode with an empty question box, three for each file. The quotes above come from the first single run of each file; the counts in the table come from the nine repeat runs. The repeat crop was rebuilt from the same source at the same 60% setting and came out at 92,478 bytes, against 94,342 bytes for the first crop. An identical request can be answered from the analyzer's response cache for an hour, so each repeat changed the case of the image MIME type in the request. The image bytes and the prompt were identical, and the production logs show Gemini answered every repeat.

Primary sources

Photo credit

  • No-parking sign, Plaridel, Philippines: JWilz12345, CC0, source.

Try it yourself

More guides