How AI age progression actually works — and what it cannot know

The short version

An age-progression model is an image model conditioned on a target age. You give it a face; it produces a new image of that face rendered with the visual signature of the age you asked for. The identity is preserved as far as the model can manage; the age markers are synthesised.

Nothing in that pipeline is predictive. The model has never met you, has no access to your medical history, and cannot know whether you will smoke, spend two decades outdoors, gain or lose weight, or take up swimming at 60. It reproduces the average of how faces in its training data looked at the age you requested, bent towards your features.

What actually changes in a face as it ages

Facial aging is not just wrinkles. The clinical literature describes it as a combination of skeletal change and soft-tissue change, and the skeletal part surprises most people: the facial skeleton itself remodels with age, region by region — orbit, maxilla, mandible, cranial base — losing volume and projection, which is part of why an older face reads as flatter and more hollow rather than simply more lined.

On top of that, fat compartments atrophy and redistribute, collagen and elasticity fall, the epidermis thins, and gravity plus repeated muscle pull turn dynamic lines into fixed ones. There are two competing explanatory frameworks in the surgical literature — gravitational descent versus volume loss — and the honest answer is that both effects appear in real faces.

A good age model reproduces those markers in the right order: at 40 mostly texture and early nasolabial definition, at 60 volume loss and lid changes, at 80 skeletal-looking shape change. A weak one just paints wrinkles on a young face, which is why the results look like a mask.

Sources: Facial bone aging: an update and literature review (2026), Overview of current thoughts on facial volume and aging (PubMed)

How forensic age progression differs — and why it is the honest benchmark

There is a serious, non-entertainment version of this work. The National Center for Missing & Exploited Children uses forensic artists to age-progress photographs of children missing two or more years, so the public can recognise them today.

The method is instructive because of what it needs that an app does not have. Case managers collect photographs of the child's biological parents and siblings — ideally taken at the age the child would be now — and the artist merges those real family features with knowledge of how children's faces develop. One progression takes roughly eight hours and is peer-reviewed by other artists before the family sees it. NCMEC reports more than 7,500 age progressions of long-term missing children.

That is the gap in one sentence: forensic progression uses your family's actual faces as evidence, an app uses a statistical prior. If someone tells you an app predicts your future face, the forensic workflow is the counter-example that shows what real prediction would require.

Sources: NCMEC — Long-Term Missing, NCMEC — Watching Your Child Grow Up in Pictures (2022)

Why results differ so much between apps

Three things drive quality. First, identity preservation: how much of your bone structure and expression survives the transformation. Second, age fidelity: whether 70 actually looks like 70 rather than 45-with-wrinkles. Third, input handling: lighting, angle, resolution and occlusion (glasses, hair across the face, heavy filters) change the result more than most people expect.

This is also why the same app can produce a striking result for one person and an uncanny one for the next. The failure mode is not random — it is usually a hard input: a three-quarter angle, strong side lighting, or a photo already processed by a beauty filter, which removes exactly the texture the age model needs.

What to expect from a good result

• You still recognise yourself — same eye spacing, same jaw geometry, same asymmetries.
• Ageing appears across the whole face, not only as lines around the eyes and mouth.
• The change between 40 and 60 and 80 is qualitatively different, not the same effect turned up.
• Hair, skin tone and lighting stay consistent with the original photo.

Limitations — what this article cannot tell you

It cannot tell you how you will actually look. It cannot rank models by accuracy, because there is no public benchmark that scores age-progression identity preservation against real longitudinal photographs of the same people. And it cannot tell you which app suits your photo: that depends on the photo.

FutureSelf generates a separate portrait per target age rather than one filter applied harder, which is what makes the decades look different from each other.