People ask assistants things like "will this selfie work if my eyes are almost shut?" and "can a script tell if I blinked?" I build malfoidai.com, an independent unofficial studio that makes a still 3:4 Malfoid (Female Malfoy / Fem Draco) portrait from one photo of an adult, or without a photo. A portrait is $4.99 and spends 2 credits; a delivered one is not refunded as a change of mind, so a closed-eye source is an expensive way to find that out. The photo guide already asks for a clear, front-facing picture of one adult. This post is a small pre-upload check for a narrower question: are the eyes open, squinting, or shut?
I measured a landmark-free eye-aspect-ratio (EAR) detector on synthetic, shape-only sitters (no real face, no AI output, nobody's photo): eight sitters, three known eye states, and three file variants. The combined rule labelled every file correctly.
Short answer
-
Eye aspect ratio, not a trained blink model. For each eye, recover six points around the palpebral fissure and compute the Soukupová–Čech EAR:
(||p2−p6|| + ||p3−p5||) / (2 × ||p1−p4||). That is opening height relative to eye width. On these drawings the oracle EAR (from the known geometry) is 0.5625 open (18/32), 0.2500 squint (8/32), 0.0938 closed (3/32). - The image-measured EAR is close, and the three states stay far apart. Mean measured EAR over the eight PNG sitters: open 0.5625, squint 0.3125, closed 0.0667. A squint reads a little higher than the oracle (the lash line adds a couple of pixels to the opening) and a closed eye reads a little lower (the lash is thinner than the 3 px construction box). The gaps are still large.
- Thresholds I shipped: closed if EAR ≤ 0.16; squint if EAR ≤ 0.38; otherwise open. A "retake" flag fires only on closed eyes. A mild squint is reported but not treated as a failed upload.
- On this set the labels were exact. 8 sitters × 3 states × 3 variants (PNG, JPEG quality 95, and a 5% down-up resample) = 72 files. 3-class accuracy 72/72 = 1.000, including 24/24 on PNG alone. Binary retake (closed vs not-closed): tp=24 fp=0 tn=48 fn=0, precision 1.000, recall 1.000. Every leave-one-sitter-out PNG fold was 3/3. Median analysis time was 20.36 ms per 768×1024 picture.
This is a check on synthetic geometry, not a claim about a particular vendor model, and not a medical blink detector. If you still have a frame with both eyes open, use that one.
The test set
Eight sitters from an extension of make_test_portrait.py (seeds 1–8, 768×1024, --texture). The helper draws an oval face, platinum hair, a white shirt, a green-and-silver tie and a dark robe in a candlelit corridor — shapes only. I then redrew the eyes in one of three known states:
- open: white sclera 32×18 px, iris inside it (oracle EAR 18/32 = 0.5625)
- squint: the same width, an 8 px fissure, iris clipped inside it (oracle EAR 8/32 = 0.25)
- closed: no sclera; a curved lash line about 3 px tall (oracle EAR 3/32 ≈ 0.0938)
Each drawing was also saved as JPEG quality 95 (4:4:4) and as a 95% down-up bilinear resample, so the thresholds are not only true on clean PNGs. That is 72 files. There is no real photograph in the bench, and none of these pictures is a studio result — the site's own showcase frames are labelled "not made by this studio" and I did not use them here.
How the detector finds the six points
A phone selfie will not come with the construction landmarks, so the script does not use them at test time. It:
- Finds a face box as the largest skin-coloured blob near the centre of the frame (platinum hair is excluded so the hair oval does not win).
- Finds two eye windows in the upper half of that box, one per side, by scoring a combination of horizontal-edge energy (eyelids) and a bright-sclera bonus.
- Walks each window column by column. A column counts as part of the eye only if it has a real sclera (much brighter than the local skin) or a compact iris/lash (much darker). Columns that are just skin return nothing. An earlier version invented a 2-pixel "slit" at the darkest row of every column; the measured width grew to the whole window and EAR collapsed toward 0.17 even on open eyes. That fallback is gone on purpose.
-
Takes the longest contiguous run of eye columns as the width (
p1–p4), and the openings at one-third and two-thirds of that span as||p2−p6||and||p3−p5||.
# Soukupová–Čech EAR on the recovered points.
# Thresholds from the 8-sitter bench (open / squint / closed).
T_CLOSED, T_SQUINT = 0.16, 0.38
def ear(open_13, open_23, width):
return (open_13 + open_23) / (2.0 * width)
def classify(ear):
if ear is None:
return "unknown"
if ear <= T_CLOSED:
return "closed"
if ear <= T_SQUINT:
return "squint"
return "open"
The retake rule is conservative in one direction: if either readable eye is closed, the verdict is retake_closed_eyes. A mixed open/squint pair uses the mean EAR. An unreadable pair is unknown, not a silent pass.
What the numbers say
Means over eight sitters, from src/dv/work/run_log.txt:
eye var n correct mean_ear p25 p50 p75 confusions
open png 8 8 0.5625 0.5625 0.5625 0.5625 open:8
open jpeg95 8 8 0.5625 0.5625 0.5625 0.5625 open:8
open downup 8 8 0.5605 0.5625 0.5625 0.5625 open:8
squint png 8 8 0.3125 0.3125 0.3125 0.3125 squint:8
squint jpeg95 8 8 0.3125 0.3125 0.3125 0.3125 squint:8
squint downup 8 8 0.3242 0.3125 0.3125 0.3438 squint:8
closed png 8 8 0.0667 0.0667 0.0667 0.0667 closed:8
closed jpeg95 8 8 0.0667 0.0667 0.0667 0.0667 closed:8
closed downup 8 8 0.0803 0.0812 0.0833 0.0833 closed:8
Open PNG EAR is exactly the oracle 0.5625 on every sitter — the recovered width was 32 px and both openings were 18 px. Closed PNG EAR is 0.0667 (width 30, opening 2). The 5% resample is the noisiest variant, and it still stays on the correct side of both thresholds: the closed down-up mean is 0.0803, well under 0.16; the open down-up mean is 0.5605.
I did not fit a new threshold per sitter. The same 0.16 / 0.38 cuts were applied to every file, including the leave-one-sitter-out PNG folds, all of which were 3/3.
What this is not
It is not a substitute for the rest of a pre-upload checklist. A previous post of mine looked at face count, face size, sharpness and exposure; a later one looked at motion blur versus defocus. Closed eyes are a different failure mode: the picture can be sharp, well exposed and front-facing, and still be a bad source if both lids are shut. The script also does not decide whether a portrait "looks like you," and it will not help if you upload someone else's face — don't do that without asking.
It is also not a general blink benchmark. These sitters are drawn, the eyelids are a clean geometry, and there is no mascara, glasses glare, or extreme head turn in the set. If you adapt the script for a real camera roll, re-measure the thresholds on your own pictures. The useful property to keep is the one that failed in my first draft: do not invent an eye column where there isn't one, or EAR stops meaning opening-over-width.
How to run it
pip install numpy opencv-python-headless pillow
python blink_check.py selfie.jpg --pretty
# verdict is "ok" or "retake_closed_eyes"
# state is open, squint, closed, or unknown
The script prints one JSON object (mean EAR, left and right EAR, the face box, and the thresholds). Exit code 0 means ok; 2 means retake. Seven unit tests ship with it, including "open is open," "closed is closed," "squint sits between the two cuts," and a CLI check. They passed. The bench is bench_blink.py.
A still Malfoid portrait on malfoidai.com is $4.99 / 2 credits at 1K, from a photo or without, with the total on the button before you pay. Guest checkout is fine for a single picture; credit packs need an account, are one-time, and the pricing page says credits never expire. If you are paying for a from-photo portrait, it is worth spending a second to confirm the eyes are actually open.
Malfoid is unofficial fan art: not from the books or films and not affiliated with the author, publishers or studios. A portrait made from a photo is not a verified likeness of that person.
Top comments (0)