📋 Reference note. This blog post covers the scoring formula from a narrative angle. For the current, authoritative technical reference, see How PSL Scoring Works — the up-to-date canonical page with the full formula and dimension breakdown.
How PSL scoring actually works
A fair question: if an AI rates a photo, how do you know the number is not just made up? The answer is in how PSL scoring is structured. The model describes the photo; the backend computes the score. The two jobs are kept separate on purpose.
Step 1: The model reads visible dimensions
When you upload a photo, the vision model does not output a final score. Instead, it reads six visible dimensions — things that can be seen in the image:
- Lighting — how the light falls on the face.
- Angle — the camera angle relative to the face.
- Expression — how relaxed or forced the expression reads.
- Image quality — sharpness, resolution, and compression.
- Face visibility — how much of the face is unobstructed.
- Harmony impression — how the features read together in this photo.
Each dimension is rated on a 0–10 scale. These are observations about the photo, not a verdict on the person.
Step 2: The backend computes the score with a fixed formula
The final score is never produced by the model. The backend applies a fixed, weighted formula to the dimensions:
harmony_impressioncarries the most weight (0.45),angleis next (0.15),- the remaining dimensions (lighting, expression, image quality, face visibility) each carry a smaller weight (0.10).
The weighted raw score is then mapped to the /8 scale and clamped to the 1.0–8.0 range. Because the formula is fixed, the same dimensions always produce the same score. The model cannot "invent" a number.
Why this split matters
If the model produced the score directly, the number would drift with the model's mood, the prompt, and the photo. Two runs of the same photo could give different scores for no clear reason.
By splitting the work — model reads dimensions, backend computes score — the scoring stays consistent and repeatable. You can compare two photos and trust that a difference in score comes from a difference in the dimensions, not from randomness.
What this means for single vs. multi-photo
The fixed formula keeps the math consistent, but it cannot fix a noisy photo. A single photo still has lighting, angle, and expression baked in. That is why PSL Range uses the same consistent formula across multiple photos to produce a stable range:
- Each photo is scored with the same dimensions and formula.
- The range shows how the face reads across conditions.
- The best photo is the one where the dimensions read most cleanly.
- The variance driver points to the dimension that moves the most across the set.
How to read the score safely
A PSL-style score is an AI-generated estimate of how a photo reads. It is not a factual statement about a person, their worth, or their value. The dimensions and the score are tools for comparing your own photos — not for ranking people.
Try the scoring yourself
See the dimensions behind your score, and get a stable range across photos.