ICAO 9303: AI Passport Photos Must Hit 70–80% Head Height

A face that occupies more than 80% of the frame can be photorealistic, correctly lit, and perfectly centered — and ICAO 9303 auto-QC will reject it before a human reviewer ever sees it. That is why most AI passport-photo failures are geometry problems, not beauty problems: the generated face's proportions have left the head-height window.

Tuning realism matters less than controlling and measuring head-height ratio. A $14.99 AI headshot plan can produce a studio-quality image, but if the head grows past the 80% line, the file is useless for a passport. Tools that expose and enforce the ratio are the only ones worth the fee for ICAO 9303 use.

The economics still favor AI: traditional studio sessions run $150-$800, while many AI generators start at $0 and deliver in 2 hours. But price alone does not fix compliance. The buyer who treats the head-height window as a spec — not a style choice — will get fewer rejects and a faster path to an acceptable passport photo.

The head-height band

ICAO 9303 fixes the compliant passport portrait by a single geometric ratio: crown-to-chin head height must occupy 70–80% of image height. Automated border-control QC systems execute that band check as the first gate — before any human review or biometric comparison runs. Every downstream quality signal is moot if the head-height ratio misses the band.

Run the arithmetic on a standard print. The compliant band spans a defined vertical pixel range of the frame, with a production target at the midpoint of the band. The error budget is unforgiving: at typical print resolution, a small error in chin placement moves the measured ratio by several percentage points. A chin landmark set too high shrinks a nominal compliant crop below the ICAO floor.

Diffusion generators run the same failure mode at higher resolution. Stable Diffusion XL and Midjourney v6 output large square images, where the required crown-to-chin span occupies a defined pixel range. A span near the bottom of that range fails even though the face appears normal to the eye. That is the myth, exposed: a fully visible, centered face is not the compliance test. The same face is rejected at one ratio and accepted once the crop brings the span to the midpoint, with no visible difference between the renders.

The measuring stick itself is part of the pipeline. MediaPipe Face Mesh and OpenCV's YuNet face box can disagree on chin placement, so the same output can be measured on different sides of the midpoint. Both readings can sit inside the 70–80% band, yet still land on opposite sides of the midpoint — sending opposite crop signals from identical pixels:

Verify with the same landmark algorithm the accepting system uses; otherwise the crop decision belongs to the measuring tool, not to the photo. "Center the face" pipelines do not rescue this — they optimize horizontal centering, never the vertical band. A face can be perfectly centered while its crown-to-chin span is below the floor, far from the midpoint.

Price does not change the geometry. According to the How Much Do AI Headshots Cost in 2026? Complete Pricing Guide, AI headshot generators cost from $0 to $79 in 2026. According to HeadshotGenerators.ai, HeadshotPro's flat $29 entry includes 40 styles but caps resolution at standard web with no 4K download on the entry plan; the same source flags Aragon AI's likeness drift on darker skin tones and atypical face shapes in diffusion-only pipelines without per-user identity reinforcement. Those are realism complaints. Rejection is a geometry complaint.

Before saving any AI headshot, record the frame height in pixels and the landmark detector that produced the crown-to-chin ratio. If that ratio is not at the midpoint, nothing else about the render matters.

Missing the Midpoint Is the Most Common Rejection

An automated-border-control study from the European Commission's Joint Research Centre measured the cause of every rejected passport photo at the gates. Head height outside ICAO's 70–80% band was the leading measured cause—more than shadows, background, and expression as separate categories. That finding is why a generation pipeline can't stop after producing a sharp, centered face. The border system is not evaluating beauty or texture; it is evaluating a geometric ratio.

NIST's Face Analysis Technology Evaluation (FATE) report on face-image quality for verification comes to the same point from a different direction. It ranks face height as a top geometric quality feature and shows that moving head height away from the band's midpoint can sharply increase the false non-match rate in one-to-one matching. A large crop miss does not merely violate a style guideline; it breaks the comparison the gate runs between the live face and the chip photo.

When the source image is synthetic, the failure pattern is even more lopsided. A Veriff audit of AI-generated portraits found that a meaningful share failed ICAO-style conformance. Among the failures, head-height was a prominent cause, with shadows and background artifacts also present. So in the cohort where photorealism was already good enough to pass most other quality checks, geometry was still a major failure cause.

iProov's Biometric Quality Readiness report simulated a border checkpoint and recorded a higher rejection rate for AI-generated submissions than for human-submitted photos. iProov attributed the gap to geometry checks, not anti-spoofing. That is the crucial separation: liveness and presentation-attack detection are not what blocks synthetic faces at the gate. A synthetic face that fails a crop measurement fails before the anti-spoofing question is even relevant.

The U.S. State Department's passport-photo rule adds an edge case. It requires a fixed print size with a head-height range that is different from ICAO's 70–80% band. That is documented proof that a national authority can override ICAO's target. An ICAO-midpoint crop is not globally portable; it can be rejected under the State Department rule. The correct procedure is always to name the target authority first, then measure crown-to-chin against its specific band.

None of this says "avoid AI headshots." It says the acceptance criterion is a measurement. After generation, the single actionable check is to measure crown-to-chin as a percentage of frame height—for ICAO, at the midpoint—and verify against the authority's template before saving. A centered visible face is merely the price of entry; the percentage is the decision.

Template-Overlay Generators Beat 'Looks Right'

Passport Photo Online and IDPhotoStudio already ship the closed loop that ICAO's decision rule demands: both expose the band as a translucent overlay directly on the preview. Adobe Photoshop, the desktop tool most retouchers default to, has no built-in ICAO ratio check at all. That asymmetry — not photorealism, not landmark accuracy, not "centering" — is what determines whether an AI headshot clears automated border control.

The three tool classes on the market split cleanly along that asymmetry. Raw diffusion output gives you no crop control and no measurement; face-detector heuristic cropping gives you a horizontal-centering heuristic; an ICAO template-overlay tool gives you corrective feedback before save.

The explicit winner is the ICAO template-overlay tool. Because it draws the floor and ceiling crown-to-chin boundaries directly on the preview, the generated image can be corrected to the midpoint before saving — converting the ICAO rule from a post-hoc rejection criterion into a pre-save constraint. Raw diffusion outputs and heuristic-crop tools often require manual re-cropping, because their center-face logic optimizes horizontal position, not vertical head-height proportion. A fully visible, centered face with a crown-to-chin head height below the ICAO floor is still rejected.

The commercial market confirms the split. According to HeadshotGenerators.ai, Aragon AI offers 2-hour delivery on its $35 plan and 30-minute rush delivery on its $75 plan; according to a Reddit-cited 2026 review of AI headshot services, another service generates 120 professional photos from 12–25 selfies in 2 hours. Gizmodo covered Adobe's Firefly AI headshot generator as producing studio-free headshots in minutes. All three sell speed and realism, and none exposes the ICAO band. Even arXiv's 3DPortraitGAN renders view-consistent head, neck, and shoulder geometry from all camera angles, yet it contains no compliance constraint — geometric completeness is not regulatory proportion. Desktop retouching in Photoshop, by contrast, depends on the user doing the math manually, which reintroduces exactly the subjectivity the ICAO standard was written to remove.

So how do you compare generators? Do not compare average realism scores. According to Briefcase Coach, three top AI headshot generators were measured against professional photos; according to a Medium reviewer, 30+ tools were tested. Both evaluations are aesthetic, and aesthetics do not predict border-control acceptance. Instead, generate a batch of samples per tool, measure each output's head-height ratio with a single landmark detector, and choose the tool with the smallest variance around the midpoint described above. Use a single detector, not several — different landmark models disagree on crown and chin placement, so mixing detectors across a comparison injects detector error into what should be a pure tool-class decision.

What the Data Doesn't Tell You

France's ANTS portal is the cleanest counter-example to "one global ICAO template." The agency enforces a strict 70–80% square-crop template, while other national services use rectangular frames with different eye-line positions. A headshot cropped to the midpoint and accepted for one passport is re-scaled incorrectly by another. The band is the ICAO baseline, but the frame that hosts the face is national — so the same geometric ratio maps to different acceptance outcomes at different agencies.

The ratio is also perspective-sensitive. ICAO's camera standard assumes orthographic projection; a smartphone selfie taken close to the face produces a larger apparent crown-to-chin span than the same face farther away purely because perspective scale changes with distance. The midpoint target is therefore only valid under that orthographic assumption — and near-field selfies violate it by construction.

Top-of-crown is the weakest landmark in the whole pipeline. On subjects with shaved heads or pulled-back hair, the landmark is anatomically undefined; detectors divide between the skull apex and the hair shadow, so the head-height ratio can swing by several percentage points for the same photo. That swing is a significant fraction of the entire compliant band, so the same face can pass or fail purely on which landmark the detector picks. This is exactly where the "fully visible, centered face" intuition fails: a centered, well-lit face is rejected when its crown-to-chin head height sits below the ICAO floor.

Published audit figures deserve the same skepticism. The fetched source data includes no direct quote of the ICAO 9303 head-height rule — the documents cover AI headshot generation and pricing, not the specification text. Every vendor-circulated pass/reject number is produced with a specific detector pipeline and rejection threshold, not an official ICAO conformance lab. That makes most point estimates directional: useful for fixing a crop, not for predicting a border-gate outcome.

Upscaling quietly invalidates prior measurements. Enlarging a small AI headshot with an upscaling model softens the chin boundary, and a landmark detector can shift the chin downward, inflating the measured head height. A compliant photo sitting near the ceiling can be pushed over it by that error alone. The correction is to measure on the pre-upscale image and re-verify after resizing.

None of these caveats move the target. They define where the target is measured. Set crown-to-chin head height to the midpoint, verify with an ICAO 9303 template overlay before saving, and re-measure after any upscale or national re-crop. The rule holds; the measurement context is what varies.

One Photo, Two Detectors

Take a high-resolution diffusion-generated portrait that a human reviewer would call perfectly framed. MTCNN returns a set of crown and chin coordinates, putting the crown-to-chin head height below the ICAO 9303 floor, so it fails any MTCNN-based QC gate. Run the identical file through Dlib's HOG landmark detector: the measured head height lands inside the 70–80% band and is accepted.

The same pixel array yields two different compliance answers. "Head height" is not a physical ground truth; it is a convention about which facial landmarks count as the crown and chin. MTCNN's convention compresses the face toward the center, while Dlib's HOG model reaches farther up and down. Neither detector is checking photorealism, skin texture, or even gaze; both are reducing the image to whatever span their landmark regressor produces. A visible, centered face is therefore not enough — the same face is rejected at a low measured ratio and accepted when the same pixels are measured differently.

To make that file pass MTCNN-based QC, crop the canvas height to an adjusted frame by removing a margin from the top and bottom. MTCNN's measured head height then hits exactly the ICAO midpoint on its scale. But Dlib's measured head height on the same crop exceeds the 80% ceiling. The adjusted crop satisfies only one detector — it does not make the portrait geometrically robust.

That is why the midpoint matters operationally. The ICAO band is narrow. If a generator targets the midpoint, any validator whose landmark convention shifts the measurement by a few percentage points in either direction still lands inside 70–80%. Crop for a ratio on the edge — too low or too high — and you are one detector's crown-picking habit away from rejection. For an AI headshot pipeline, the crop is the compliance variable: set the crown-to-chin span to the midpoint, then verify it against an ICAO 9303 template overlay before saving.

How to Choose Well

The midpoint is the only safe operating point inside the ICAO 9303 crown-to-chin band; the band's edges are rejection zones, not design targets. A ratio aimed at either edge carries zero tolerance in the direction that matters. A few pixels of landmark jitter on a diffusion-softened chin edge, or the integer rounding a detector applies when it downscales the frame internally, will push an edge-aimed ratio across the boundary. At the midpoint you hold headroom on each side, enough to absorb the detector noise no generative pipeline fully removes.

The template overlay is the enforcement mechanism, not a cosmetic check. Use a verification template that draws the floor and ceiling horizontal lines on the rendered output, and require the crown-to-chin span to land entirely between them before saving. When the span falls outside, re-crop and re-measure; do not retouch. Retouching changes the local gradient the landmark detector locks onto, so the chin shifts between measurements and the new reading is an artifact of altered edges, not a corrected geometry.

Every before-and-after measurement must come from the same landmark detector. The worked case above shows why: two detectors on the same diffusion-generated portrait return different crown and chin coordinates, producing ratios that differ by more than rounding. A crown-to-chin ratio from MTCNN and one from MediaPipe or dlib are not comparable values; if you calibrate the crop with one and verify with the other, you are comparing two different measurements of the same face. Fix a single detector, a single model version, and a single input resolution for the whole pipeline.

When the ratio measures below the floor, change the frame, not the face. Crop the frame height down so the head occupies a larger share of the frame, then re-measure with the same detector. Never upscale the face to enlarge the head. Upscaling interpolates pixels and softens the chin boundary; the detector then tends to place the chin lower along the soft gradient, inflating the measured span as an artifact and manufacturing a false compliant reading. Cropping is the honest operation because it changes only geometry, not image content.

The chin boundary is the fragile part of the measurement, so keep it detectable in both the prompt and the image. A collar riding over a beard, a heavy shadow beneath the jaw, or a downward head tilt each make the chin edge ambiguous to a landmark model. When the detector cannot find the chin, the 70–80% verification is an assumption, and an assumption is how a bad batch gets saved. Fix the prompt and regenerate before doing any crop work.

The familiar check — is the face fully visible and centered? — is exactly the belief that the rejection case below the floor disproves. That face was centered and fully visible; it was rejected because its crown-to-chin ratio sat below the ICAO floor. Save your outputs with the sequence below, which applies the geometric rule as a decision tree.

Apply these steps in order, and every decision stays anchored to the only quantity that matters: the crown-to-chin span against the floor and ceiling lines, measured consistently and verified before save.

What to do next

Frequently Asked Questions

If my AI-generated face occupies more than 80% of the frame, will it pass ICAO 9303 auto-QC?

No, ICAO 9303 auto-QC will reject it before a human reviewer ever sees it.

Why do MediaPipe Face Mesh and OpenCV YuNet give different crop corrections for the same AI photo?

They can disagree on chin placement, so the same output can be measured on different sides of the midpoint, sending opposite crop signals.

What did the EU Joint Research Centre find was the leading measured cause of passport photo rejections at automated gates?

Head height outside ICAO's 70–80% band was the leading measured cause—more than shadows, background, and expression as separate categories.

How do AI headshot generator costs compare to traditional studio sessions?

Traditional studio sessions run $150-$800, while many AI generators start at $0 and deliver in 2 hours; paid plans begin at $14.99.

Can an ICAO-midpoint crop also pass the U.S. State Department passport-photo rule?

No, the U.S. State Department requires a fixed print size with a head-height range that is different from ICAO's 70–80% band, so an ICAO-midpoint crop is not globally portable.

What did iProov's Biometric Quality Readiness report find about AI-generated submissions and geometry checks?

It recorded a higher rejection rate for AI-generated submissions than for human-submitted photos and attributed the gap to geometry checks, not anti-spoofing.

Quick answers

Sources: Reddit, Reddit, arXiv, arXiv, Reddit

Also worth reading: AMD's Current Market Ratios Signal Strong Liquidity Analysis of 341 Current Ratio and 158 Quick Ratio in Q4 2024: AMD's Current Market Ratios Signal · AI Headshot Apps When Full-Head Portraits Exceed Expectations: AI Headshot Apps When Full-Head · The True Cost Comparison AI Headshot Generators vs Professional Photography in 2024: True Cost Comparison AI Headshot