Prompt:

This is an eval. Your output will be scored on one thing: how often
people mistake it for a photograph of a real human.
TASK
Write a program, in any language, that generates a photorealistic
image of a human face, head and shoulders, and saves it as face.png
(1024x1024).

SCORING
Your image will be mixed in with real photographs of real people.
Raters will be asked which images are real. Your score is the rate
at which your image is judged real. Looking "good", "pleasant", or
"impressive for code" earns nothing. A stylized or cartoon face
scores zero. Only being mistaken for a real person counts.
RULES
Code only. The program may not read any files, make network
requests, or contain embedded image data, photos, scans, or
learned weights. All geometry, skin, eyes, hair, and lighting
must be computed by the code. Source file size: 500 KB max.
Standard libraries and general math libraries (e.g. numpy) are
allowed. Rendering, 3D, or image-generation libraries that ship
with models or assets are not.

The face must fill at least 40% of the frame, front or
three-quarter view, eyes open and visible, clearly lit.
No text in the image.
The graders will rerun your source on a CPU. It must reproduce
face.png, deterministically, within 30 minutes. If it does not,
you score zero.
TIME AND RESOURCES
Assume CPU only, no GPU.
You have 30 minutes of wall-clock time. At 30:00, face.png and
the source file in your working directory are frozen and graded
as-is. Save working versions often. A broken or missing file
scores zero.
Use everything available to you: subagents, parallel attempts,
research, viewing your own renders, critique and revision loops.
Spend the full time if it helps.