I Built an AI-Powered Face Rating Tool Around the PSL Scale
I've always found it interesting how difficult it is to objectively describe facial attractiveness.
People can look at the same face and have completely different opinions.
One person notices symmetry.
Another focuses on the eyes.
Someone else cares more about jawline, facial proportions, skin, or overall harmony.
And when people try to put all of those things into a single number, the conversation usually becomes even more subjective.
That led me to build PSL Scale.
PSL Scale is an AI-powered facial attractiveness evaluation tool that analyzes a face from a photo and produces a PSL score on a 0–8 scale, together with a breakdown of different facial dimensions.
The goal isn't to tell someone whether they're “good” or “bad looking.”
The more interesting question is:
Can we turn a subjective visual judgment into a structured, understandable analysis?
What is PSL?
PSL is a term commonly used in online looks-rating and looksmaxxing communities.
It's typically represented as a compressed 0–8 facial attractiveness scale, where around 4 is roughly average, 5–6 is above average, and 7+ represents increasingly rare, model-level territory.
The problem is that the scale is usually discussed informally.
Different people can use different criteria.
Some focus heavily on bone structure.
Others focus on eyes, skin, facial harmony, or overall presentation.
I wanted to build something that could make the process more consistent and transparent.
Instead of returning:
“You look like a 6.”
PSL Scale tries to answer:
“Why?”
From one photo to multiple dimensions
The basic experience is deliberately simple.
Upload a clear, front-facing photo.
The system analyzes the face and returns an overall score along with individual dimensions.
The current scoring system includes:
Symmetry
Harmony
Proportions
Skin quality
Facial structure
Averageness
Sexual dimorphism
Memorable features
This makes the result more useful than a single number because users can see which areas contributed to the estimate.
For example, someone might receive a relatively strong overall score but discover that their facial structure scores higher than their skin quality.
That changes the conversation from:
“What is my score?”
to:
“What actually influences my score?”
That's a much more interesting product experience.
Why use a 0–8 scale?
A normal 1–10 rating sounds intuitive, but it can create a lot of noise.
People tend to treat 7, 8, and 9 as very different categories even when the underlying difference is small.
The PSL scale is more compressed.
A score around 4 represents the middle.
Scores above 6 become progressively rarer.
Scores around 7+ are intended to represent exceptional facial harmony and structure rather than simply “pretty good.”
That compression makes the scale interesting for experimentation because relatively small movements in the middle of the distribution can feel meaningful.
Of course, this is still an estimate from a photograph, not an objective measurement of human worth or attractiveness.
Lighting, camera angle, lens distortion, expression, hairstyle, grooming, and photo quality can all change the result.
I didn't want it to be just another AI prompt
One of the interesting technical decisions was separating geometric analysis from the broader AI interpretation.
The free fast analysis runs facial landmark measurements directly in the browser.
It looks at things such as symmetry, proportions, facial structure, and harmony from facial landmarks. Because the calculation happens locally, the basic analysis can be performed without sending the photo to a server.
Then there is the deeper AI analysis.
The deeper analysis combines facial geometry with an AI model that can evaluate additional visual characteristics and produce a written report.
That report can include:
An overall rating
Category scores
Visible strengths
Areas to improve
Personalized suggestions
Photo and presentation advice
This creates two different experiences.
Basic analysis: fast, deterministic, local.
Deep analysis: more comprehensive, AI-assisted, and interpretive.
I think this separation is important because AI doesn't need to be involved in every part of a product.
Sometimes a simple mathematical measurement is better than asking an LLM to guess.
Privacy became an important part of the product
Face photos are obviously sensitive.
So I didn't want the product to work like a traditional social platform where users upload a photo and it becomes part of some permanent database.
PSL Scale is designed around a different model.
The basic photo analysis happens locally in the browser, and the site states that uploaded photos are not stored as account history. The camera version similarly performs landmark processing on-device.
For deeper AI analysis, the selected image is sent to the AI service for that particular request.
This creates a useful trade-off:
Local processing whenever possible, server-side AI only when necessary.
For a product dealing with faces, I think that distinction matters.
Then I added live camera analysis
After building the photo workflow, I started wondering:
What if you didn't have to take a photo first?
That became the Camera PSL experiment.
The camera version tracks facial landmarks in real time using MediaPipe and continuously updates four geometry dimensions:
Symmetry
Harmony
Eye area
Jawline
The scan happens in the browser, and users can optionally generate a shareable result video.
This changes the interaction from:
Take photo → upload → analyze
to:
Open camera → position your face → scan → result
It feels much more like a consumer app than a traditional AI analysis website.
The product is becoming a toolkit
One thing that surprised me while building PSL Scale is how many adjacent questions naturally come from one facial analysis.
Once someone knows their PSL score, they often want to understand more.
What about my eyes?
What about my face shape?
What about symmetry?
What about the golden ratio?
What hairstyle would suit me?
How does another photo compare?
That led to additional tools around the core PSL rating experience, including Hunter Eyes, Golden Ratio, Face Shape, Face Symmetry, AI Hairstyle Changer, and PSL Rating Compare.
The product is gradually becoming less like a single “face score calculator” and more like a facial analysis toolkit.
Building the scoring system is probably the hardest part
The UI is relatively straightforward.
The difficult question is:
How should a face actually be scored?
Attractiveness isn't a single measurable variable.
You need to combine multiple dimensions.
Symmetry matters, but perfect symmetry isn't necessarily the goal.
Averageness can matter, but distinctive features can also make someone memorable.
Jaw structure matters, but so does the relationship between the jaw, cheekbones, eyes, nose, and mouth.
Skin quality matters, but a photograph can exaggerate or hide skin characteristics.
And then there are things that are difficult to quantify at all — expression, charisma, style, and how someone looks in motion.
That's why I think the score should always be treated as an estimate, with the breakdown being more useful than the number itself.
The biggest lesson: the number isn't the product
This is probably the most important thing I've learned building PSL Scale.
Initially, it is tempting to think the product is:
Upload a photo → get a number.
But the number is actually the least interesting part.
The more valuable experience is:
Upload a photo → understand the dimensions → identify strengths → identify possible improvements → experiment → compare results.
That's a much better feedback loop.
For example, if someone changes their hairstyle, improves their presentation, takes a better photo, or simply changes the camera angle, they can compare the results.
The score becomes a measurement inside a larger experimentation loop.
What I'm working on next
There are several areas I want to improve.
Better consistency
The same person should ideally receive reasonably consistent results across photos taken under similar conditions.
Reducing sensitivity to lighting, lens distortion, pose, and expression is an ongoing challenge.
Better explanations
A score without an explanation isn't particularly useful.
I want the system to become better at explaining why a particular dimension received its score.
More useful comparisons
Comparing two photos can be more meaningful than looking at one score in isolation.
I'm interested in making side-by-side comparisons more detailed and useful.
More personalized recommendations
Instead of generic advice, the system should prioritize the few changes that are most relevant to the individual image.
The goal isn't to give users a giant checklist.
It's to answer:
“If I only change two things, what should I focus on?”
The bigger idea
I don't think AI can determine whether someone is objectively attractive.
There is no universal formula that can reduce human attraction to a perfect number.
But I do think AI can make subjective visual analysis more structured.
That's the experiment behind PSL Scale.
Take something that people normally discuss with vague language:
“Good jaw.”
“Nice eyes.”
“Looks above average.”
“Something feels off.”
And try to turn those observations into measurable dimensions, scores, and explanations.
Not because the final number is the absolute truth.
But because having a structured framework can make the conversation more interesting.
That's what I'm building with PSL Scale.
If you're interested in AI, computer vision, facial analysis, looksmaxxing, or consumer AI tools, I'd be interested to hear what you think.
Would you trust an AI-generated facial analysis more if it showed you exactly how it arrived at the score?
Try it here:
