Practical Prompting 002: Predicting Our Baby’s Future Eye Colour with AI

Christoph Götz -

Scope

This post continues my “Practical Prompting” series: experiments conceived and executed on the couch that turn a casual question into a hands-on AI demo.

Today’s challenge grew out of new-parent curiosity: What colour will our eleven-week-old’s eyes be once they settle? Baby eyes often start slate-gray; genetics plus melanin eventually decide where they land on the brown–blue spectrum. I wondered how far a household-grade AI workflow, camera-phone photos + a large language model, could take that prediction.

It's a genetic mess in which brown mostly wins
A playful exploration of computer vision, quick-and-dirty colour science and probabilistic genetics. Just enough rigor to entertain at the dinner table.

Today’s Problem: Forecasting Iris Pigment Maturation

Input

Extract the eye colours of Dad (gray-blue), Mom (hazel-brown) and Baby (11 weeks) from these images [give images].
1. Analyse iris pixels for each.
2. Plot them on a radar diagram (Blue, Green, Hazel, Brown, Gray).
3. Predict Baby’s adult eye colour; show current vs. projected with error bars.
4. Summarise in plain language.

Output

1 · Iris composition plots

State-of-the-art
Future Prediction

2 · Numeric breakdown

Component Dad Mom Baby (now) Baby (likely adult)
Brown 22 % 87 % 59 % 48 %
Gray 78 % 5 % 25 % 25 %
Hazel 0 % 6 % 13 % 12 %
Blue ≈0 % 2 % 3 % 10 %
Green ≈0 % 0 % 0 % 5 %

3 · Plain-English verdict

There’s roughly a 50 % chance Baby ends up medium-brown with a cool gray rim,
30 – 40 % chance of lighter gray-green,
and a 10 % tail-risk of full hazel.

Melanin deposit peaks by 18 – 24 months; check back then!

Discussion

The exercise shows how far you can go with:

  1. A phone camera (good macro mode ≈ 5 cm)
  2. A 30-line Python snippet to mask the pupil/sclera and bucket HSV hues
  3. A conversational model that both consumes images, codes and narrates

Where it shines

  • Quantifiable: Just get the facts not the guesswork
  • Visual storytelling: auto-formatted radar charts to convince your wife of her genetic imperialism
  • Modeling uncertainty: or in this case not really uncertainty. Brown has won as excpeted from a genetic standpoint

Where to stay humble

  • Lighting bias: This was very quick and dirty. A standardized acquisition setup, RAW images or at least locked WB would get us closer to real science
  • Polygenic reality: 115 + loci; any single-family estimate is noisy (see this paper)

Still, as a living-room demo it beats the old wives’ tales.

Genetics Upgrade: How to expedite time to accurate prediction with forensic grade evidence?

I guess theres no way around sequencing everybody at home and comparing the 115 responsible Loci that encode eye color and try again using Genes + Image to arrive at a more accurate prediction faster.

In short, those 115 genotypes capture about half of what makes irises light or dark, and, unlike pixel colours, they’re immune to bad bathroom lighting. Because: Who has time to wait 18-24 months to find out the real eye color?!

What actually changes in the pipeline?

Stage Image-only workflow 115-SNP workflow
Data in JPEG → HSV pixels CSV/VCF → allele dosages (0,1,2)
Model Heuristic colour buckets Polygenic score
Prediction “Looks ~60 % brown, maybe hazel” Logistic regression → P(Brown)=0.48, P(Blue)=0.10…
Parent → child Visual resemblance only Mendelian transmission + PS distribution
Uncertainty Wide (lighting, subjectivity) Narrower (47 % variance unmodelled)
Explainability “More gray flecks from Dad” “Daddy gave you rs12913832 AA which lowers melanin by 4.2 units → lighter eyes”

TL;DR

Give the AI your 115-locus panel and the household debate upgrades from
“He probably keeps Mom’s hazel” → "Momy gave you rs12913832 GG which increases melanin by 4.2 units" meaning quantified odds with forensic-grade confidence.