ICAO 9303 face and eye-line bands: checking AI passport photos

Automated passport-photo verifiers enforce ICAO 9303 face-height and eye-line thresholds with vendor-specific tolerance bands that can reject compliant portraits.

This guide explains the 70–80% face-height and 50–69% eye-line requirements and how to verify AI-generated headshots before submission.

How ICAO 930-3 checkers measure faces

Passport Photo Snap’s biometric-photo explainer describes an ICAO 930-3 image as a portrait whose details are fixed by numbers: outer dimensions, face height from chin to crown, eye position measured from the bottom edge, background color, and gaze direction. Because every requirement is numeric, a compliance checker can treat the photo as a measurable object rather than an aesthetic judgment.

The geometric measurement pipeline that every ICAO 930-3 checker runs is landmark detection, then face-height ratio, then eye-line ratio. First, the software locates facial landmarks—typically the chin point, a crown line at the top of the head, and the pupil centers. It then divides the chin-to-crown distance by the full image height to produce a face-height percentage, and measures the distance from the bottom edge to the eye line as a percentage of image height. Those two ratios are compared with the configured pass/fail thresholds.

In practice, the denominator is always the complete image height, including any margin, so cropping changes both ratios at once. The chin point is the lowest facial contour, the crown line is the top of the head rather than the top of the hair, and the eye line is usually derived from pupil centers rather than eyelids. Small shifts in any landmark can move a ratio enough to change a result, especially near a tolerance boundary.

The comparison stage is where vendor differences appear: the same landmark set can be tested against different tolerance bands or rounding rules, so a photo that passes one checker may fail another even when the underlying geometry is unchanged. This section does not restate those bands; it identifies the measurement stage where such disagreements originate.

As detailed in the sections below, run any AI-generated or retouched portrait through at least two independent ICAO 930-3 compliance checkers before submission. If the two checkers disagree, inspect their landmark overlays, adjust crop or alignment, and rerun both rather than editing facial features.

A useful verification habit is to keep the original uncropped file and compare each checker’s chin, crown, and pupil markers against the same image. When the markers are stable but the ratios differ, the difference is in threshold handling, not in the portrait; when the markers drift, the difference is in landmark detection and the crop should be corrected first.

The evidence: thresholds and tolerances

According to passportphotosnap.com, which defines biometric photos as photos measured by software per ICAO 930-3, the two ratios most automated checkers enforce first are face height (chin to crown) occupying 70–80% of image height, and eye position falling 50–69% up from the bottom edge. These are the specific ICAO 930-3 numeric bands, and the country-format table below is the only reference that anchors every threshold in this guide.

While those ratios remain constant, the outer frame varies by document. As noted by passportphotosnap.com, Turkey uses 50×60 mm, Schengen uses 35×45 mm, and the US uses 2×2 inch. One compliant frame can be re-cropped to fit those different outer dimensions without changing the underlying portrait.

For any submission, check the intended document's outer dimensions against this table first, then confirm the crop still meets the 70–80% face height and 50–69% eye-line bands, since changing the frame size does not shift those ICAO 930-3 ratios.

Where AI headshots fail compliance

AI-generated portraits can interact with automated scrutiny in ways that ordinary photos do not. This section maps three plausible failure modes (geometry drift, texture artifacts, background non-uniformity) to specific checker stages, giving a way to diagnose a reject before resubmitting. No grounded source quantifies how often each mode occurs, so treat the mapping as a diagnostic framework, not a frequency claim.

First, geometry drift is caught at the framing stage. Generative models trained on portrait photography can default to a tighter crop that enlarges the head relative to the frame, which can push the head past the framing ceiling that the ratio check enforces. No grounded source measures how often this happens, so verify it directly: measure chin-to-crown against image height on every generated output rather than assuming the model crops consistently.

Second, texture artifacts are flagged at the image quality stage. Overzealous skin-smoothing and super-resolution hallucination can strip away natural skin texture to the point of looking synthetic. As noted by travel.state.gov, blurriness, graininess, or pixelation are explicit rejection grounds for passport photos. With AI outputs, this can happen even when the crop is correct, because the checker is looking for photographic texture, not just overall sharpness.

Third, background non-uniformity is caught at the backdrop stage. AI background replacement often relies on inpainting, which can leave behind subtle gradients, faint color shifts, or shadow halos around the hair and shoulders. These remnants are not always obvious to the eye, but automated backdrop checks look for a uniform field, so they can trigger a reject. Unlike a physical backdrop, these flaws are stitched in, making them far more common to AI-generated portraits than to a properly lit photo.

Checker options compared

MergeImages.net advertises the broadest grounded coverage: upload a selfie, choose a country, and its AI crops, aligns, and checks compliance for 149+ countries, returning a print-ready file at 300 DPI in under 60 seconds with no account required. That breadth is useful for country-format comparison, but the same workflow also applies one vendor's crop and alignment logic.

PhotoGov states a 30-second turnaround, shorter than MergeImages.net's under-60-second claim, using smart auto-crop and background removal that it markets as meeting "100% official requirements." A maker that also performs the check can hide tolerance variance, because its own crop may sit near an edge of an acceptable band while another checker measures the same file differently.

Passport-Photo.online is strongest at capture-stage guidance: it supports taking a photo with a phone or camera or uploading from a gallery, then uses on-screen instructions and "acceptance guaranteed" framing. That guidance can reduce framing errors before automated checking, but it does not replace an independent measurement of the finished file.

The practical comparison is therefore not just speed or country count. As detailed under the numeric target below, export the finished image and test it independently, keeping the version whose measured face height sits in the conservative middle of the acceptable band rather than near either edge.

Numbers that decide pass or fail

For the purpose of passing automated checks in 2026, the only numeric target to hold for face height is 74–76% of image height. This is not an expansion of the known band, but a deliberate midpoint. By placing the chin-to-crown measurement near the center, the distance to both edges is maximized. With vendor-specific tolerance bands, a shift between checkers can be enough to cross an edge, so locking to 74–76% is the narrow point that gives the most room against that split without relying on any single checker’s leniency.

The cost of aiming at the band edges is asymmetrical. Aiming high toward 78–80% leaves less margin below a stricter upper ceiling and can trigger a false-reject from a checker that enforces a tighter top bound, forcing a resubmission cycle. Aiming low toward 70–72% leaves less margin above a stricter lower floor and can trigger a false-reject from a checker that enforces a tighter bottom bound, also forcing a resubmission cycle. In both directions, the action is the same: re-crop and re-test, so the resubmission cycle falls on the submitter regardless of which edge is missed.

Before locking any crop to that 74–76% target, verify the pixel count against the final print size at 300 DPI. For a 2×2 inch US print, that is 600×600 px minimum. For a 35×45 mm Schengen print, 35 mm divided by 25.4 mm per inch equals about 1.378 inches, multiplied by 300 DPI equals about 413 px, and 45 mm divided by 25.4 equals about 1.772 inches, multiplied by 300 DPI equals about 531 px, so at least 413×531 px at 300 DPI for that size. Treat these as a resolution floor for the measured ratios to be read cleanly. From there, run the exact cropped file through at least two independent ICAO 930-3 compliance checkers and require the reported face height to land at 74–76% before submission.

What the evidence does not establish

This section catalogs only what the evidence does not establish: the unmeasured variables behind automated passport-photo checks—vendor tolerance bands, AI-model geometry bias rates, and acceptance outcomes—are listed once in the edge-case table below. It does not convert any checker result into a guarantee, and it does not assign a frequency to a failure mode that no source has measured.

No grounded source publishes actual vendor tolerance bands or measured false-reject rates for any named checker. That means a pass/fail boundary cannot be quoted from marketing copy; it has to be recovered by a binary-search method. Use one base image, change a single geometric variable in small increments, run the same file through two independent checkers, and record the first value that flips either result. Repeat for the other ratio, and keep the two checkers' readings side by side rather than averaging them.

The hair-volume edge case shows why the boundary can be ambiguous even when the rule is clear. Crown-line detection is unreliable when hair is voluminous or absent, so the 70–80% face-height band can be measured against different top points by different software. The rule still wins if you measure to the visible crown line consistently across both checkers, because consistency removes the checker-to-checker disagreement even when the physical crown line is hard to define.

Heavy retouching creates a second ambiguity: edits can move the apparent chin, crown, or eye positions that the checkers' landmarks depend on. A portrait that measured inside the target before retouching may not measure there after skin smoothing, jaw reshaping, or hairline changes. The safe response is procedural rather than numerical—re-run both checkers after every edit and keep the measured face height at 74–76% of image height before submission.

The table is the complete inventory of what remains unmeasured. It separates variables that are unknown from handling steps that are controllable, so the reader can see which risks require measurement and which require only discipline.

Worked Example: Run the Numbers

Illustration (hypothetical inputs, not a measured case): Scenario—one AI-generated 2×2-inch U.S. passport portrait, 600×600 px at 300 DPI; Date—March 15, 2026; Party—one adult applicant submitting for a U.S. passport.

Step 1: image height = 600 px. Step 2: reader-rule face-height window = 74–76% of 600 px = 444–456 px. Step 3: eye-line window used in this illustration = 50–69% of 600 px = 300–414 px from the bottom. Step 4: Checker 1 reports face height 448 px (74.7%) and eye line 330 px (55.0%); Checker 2 reports face height 452 px (75.3%) and eye line 324 px (54.0%).

Both readings fall inside the 444–456 px face-height window and the 300–414 px eye-line window, so the two checkers agree on pass/fail for this image.

Winner for this illustration: submit the image without recropping. Break-even trigger: if either checker reports face height below 444 px (74%) or above 456 px (76%), or an eye line outside 300–414 px, treat the image as failing and regenerate or recrop before submission.

And decision rules

Use this pre-submission checklist to operationalize the 74–76% face-height target for an AI-generated or retouched portrait. It converts the target into three measurable pass/fail checkpoints and the if/then rules that follow from each result.

Worked example: a U.S. 2×2 inch photo at 300 DPI is 600×600 px. Checkpoint 1: measure chin-to-crown face height and require 438–456 px, which is 73–76% of image height; for the stricter 74–76% target, use 444–456 px. Checkpoint 2: measure the eye line from the bottom edge and require 300–414 px, or 50–69%. Checkpoint 3: sample luminance at all four corners and require the values to fall within one narrow band. If any checkpoint fails, correct the geometry or background before submission; if all pass, submit the file only after a second independent checker agrees.

Copy-usable checklist: [1] measure face height ÷ image height and require 0.74–0.76; [2] measure eye line from bottom ÷ image height and require 0.50–0.69; [3] measure corner luminance and require one narrow band. Save the three measured values with the file name so a later edit can be compared against the original.

Decision rules: if the face-height ratio is below 0.74, enlarge or reposition the face in the frame; if it is above 0.76, reduce face scale or add margin; if the eye-line ratio is below 0.50, raise the portrait; if it is above 0.69, lower it; if the four corner samples do not share one narrow luminance band, replace or even out the background. After any change, remeasure all three checkpoints because one adjustment can move more than one ratio.

A rejected application cycle costs more time than a short self-measurement either way. Run the checklist above, then apply the two-checker verification rule described earlier in this guide and keep the stricter result.

What to do next

Frequently Asked Questions

What face-height target is recommended to reduce automated false rejects?

Target 74–76% face height to stay within vendor-specific tolerance bands and reduce automated false rejects.

What eye-line range does ICAO 9303 require from the bottom of the image?

ICAO 9303 requires the eye line to fall 50–69% from the bottom of the image.

What can happen when measurements fall near threshold boundaries?

Compliant portraits can be rejected when measurements fall near threshold boundaries.

Why might a photo pass one verifier but fail another?

A photo that passes one verifier may fail another if face height or eye line is close to the 70–80% or 50–69% limits.

What verification step is recommended before submitting AI-generated or retouched portraits?

Run every AI-generated or retouched portrait through at least two independent ICAO 9303 compliance checkers before submission.

What details does Passport Photo Snap’s explainer say are fixed by numbers?

Passport Photo Snap’s biometric-photo explainer describes an ICAO 930-3 image as a portrait whose details are fixed by numbers: outer dimensions, face height from chin to crown, eye position measured from the bottom edge, background color, and gaze direction.

Quick answers

Also worth reading: ICAO 9303: AI Passport Photos Must Hit 70–80% Head Height: ICAO 9303: AI Passport Photos · ICAO 9303 2026: AI Headshots 57% Fail Rate, Crop Fixes: ICAO 9303 2026: AI Headshots · 2026 ICAO 9303 Head-Height Rule Breaks GANs, Diffusion Passes: 2026 ICAO 9303 Head-Height Rule