iLook Face Analysis Stability Benchmark 2026
This page preregisters the evaluation design. Results have not yet been published because the licensed synthetic adult dataset has not yet been acquired and run.
Research question
How stable is the versioned iLook report when the exact same image is repeated, and how much do controlled changes in lighting, crop, and small viewing angle affect the result?
Planned evaluation set
The full benchmark will use 100 licensed synthetic adult portraits. Each baseline image will be analyzed five times. A 30-image subset will receive controlled lighting, crop, and small-angle transformations. Dataset licensing and provenance will be documented before the first production run.
Registered metrics
- Overall-score mean, standard deviation, and maximum repeated-run spread.
- Face-shape category agreement across repeated runs.
- Schema-validation success and malformed-output retry rate.
- Safety refusal and temporary-failure rates.
- Score and category changes under controlled photo conditions.
Non-claims
This benchmark does not measure attractiveness accuracy, identity, age, ethnicity, health, personality, disability, sexuality, or demographic characteristics. Repeatability is not the same as scientific validity.
Downloads and citation
Protocol in Markdown · Results table in CSV
Cite as: Arctan Technologies Private Limited. “iLook Face Analysis Stability Benchmark 2026: Preregistered Protocol.” iLook, 19 August 2026. https://www.ilook.fit/research/face-analysis-stability-2026
Read the broader iLook methodology and corrections policy.