A Beauty Pass That Does Not Move Your Features
Beautify is not a technical instruction. There is no measurement a model can take that says a face has become more beautiful, so what it actually does is move the face towards whatever its training data labeled beautiful — and that label came from human choices, at scale, with all their preferences in it. The direction of drift is well documented and specific, and this page is built to resist it rather than to pretend it does not exist.
Use this without the search next time. Prathom Workbench puts Prathom's tools in your toolbar.
Add to Chrome — freeDrop an image here, or click to browse
Up to 25 MB. Uploaded temporarily, deleted after processing.
What it does
- Tone evened and shine reduced only
- Nose, jaw, eye, and lip proportions locked
- Skin color held rather than lightened
- Explicit guards against feature drift
How to use AI Face Beautifier
- 1
Note your own proportions first
Look at the width of the nose and the line of the jaw in your original. These are the two measurements that beauty passes move most often.
- 2
Run the pass
Tone evening and shine control are applied while the instruction holds every proportion and the natural color of the skin.
- 3
Flip between the two files
Switch back and forth rather than looking at each separately. A drift you cannot see side by side is obvious when the images alternate in place.
How it works
Your portrait is scaled to about one and a fifth megapixels and repainted at a low denoise with a fresh seed each run. The positive instruction is deliberately narrow: even the skin tone, soften harsh shine, and hold the exact proportions of the nose, jaw, eyes, lips, and cheekbones along with the natural color of the skin.
The negative prompt is where most of the design sits. It names the drifts directly — lightened skin, narrowed nose, enlarged eyes, slimmed jaw, reshaped face, altered ethnicity, doll face — because a model steered away from a specific outcome behaves better than one merely asked to be careful.
Where the drift comes from
This is not a conspiracy and it is not a bug in any one product. It falls out of how the tools are built.
To make a model that beautifies, someone has to define beautiful in data. In practice that means large collections of faces with ratings attached, or before-and-after pairs from commercial retouching, or engagement signals from platforms. Every one of those sources encodes the preferences of the people and the markets that produced it.
Learn from that and you learn a direction: towards the mean of what those sources rewarded. Applied to a face, a direction becomes a set of edits. Noses narrow. Eyes grow. Jaws slim. Skin lightens. Faces converge on each other.
The effect is strongest for people furthest from the training mean, which is exactly backwards from what anyone would design on purpose. A person whose features already match the mean sees a subtle polish. A person whose features do not sees themselves edited into somebody else.
Restricting the operation
The way to reduce the problem is to stop asking for beauty at all.
Beauty is undefined, so a model given that instruction has to fill in the definition itself, and it fills it in from the data. Tone evenness and shine reduction are defined: one is variance in skin color across a region, the other is a specular highlight from a light source. A model asked for those has much less room to interpret.
That is why this page does what it does and refuses the rest. It is not a complete answer — a sampler with a face in it can always move the face — but the narrower the request, the smaller the drift, and the easier it is for you to check whether it happened.
Compare by flipping
The checking method matters more here than on any other portrait page, because the changes are gradual and each one looks reasonable alone.
Do not view the two files side by side. Open them in the same window and switch between them so the image alternates in place. Feature drift that is invisible in a side-by-side comparison — a jaw two percent narrower, a nose slightly thinner, skin half a stop lighter — appears immediately as movement when the pictures swap.
If something moves, you now know exactly what the pass did, and you can decide whether you wanted it.
Publication gate
This page ships once the workflow has been run against portraits across a range of skin tones and face shapes, with nose width, jaw line, eye size, and mean skin luminance measured in the input and the output and compared numerically rather than by eye.
Examples
Portrait with harsh flash shine
The intended case. Shine is a lighting artifact rather than a feature, so reducing it changes nothing about the person.
Uneven skin tone under mixed light
Evens the patchiness. Check the overall skin color against the original, since evening tone is the operation most likely to lighten it as a side effect.
Frequently asked questions
Why do beauty filters change my face shape?
Because they have no definition of beauty other than what they were shown. A model trained on images that people rated attractive learns the average of those images, and applying that average means moving individual faces towards it — narrower noses, larger eyes, smaller jaws, lighter skin. It is not an opinion the model holds, it is a mean it was handed.
Does this page do that too?
It is built to resist it. The instruction restricts the pass to tone and shine, and the negative prompt names the specific drifts as things to avoid, including lightened skin and altered ethnicity. That reduces it substantially and does not make it impossible, which is why the page asks you to compare rather than to trust.
Will it lighten my skin?
It should not, and this is worth checking every time. Skin lightening is the most consistently reported bias in this class of tool, and it often arrives disguised as evening the tone or lifting the shadows. Compare a patch of cheek in both files, and if the output is paler, discard it.
What should I use instead if I only want spots gone?
The retouch page. It removes temporary marks and does nothing else, which is a defined operation with a checkable result. Reach for a beauty pass only when the actual problem is uneven tone or harsh shine across the whole face, and skip it otherwise.
Is the AI Face Beautifier tool free, and do I need an account?
It is free and there is nothing to register for. No account is created, no address is collected, and no reduced-quality preview is withheld for a plan that does not exist. The constraints that do apply come from the workflow itself rather than from billing: 25 MB in, a shared GPU queue, and a thirty-minute window to collect the output.