How AI photo restyling actually works

A filter vs a restyle

A classic filter is a lookup table: every pixel of a given colour becomes another colour, the same way every time. It cannot add hair that was not there or change a t-shirt into a blazer. An AI restyle regenerates the image, so it can.

The diffusion process, briefly

Modern image models are diffusion models. Training teaches them to take a noisy image and predict a slightly cleaner version. To edit your photo, the model:

1. Adds controlled noise to your image, partly dissolving it.
2. Denoises step by step back toward a real photo.
3. At every step, checks the result against a guidance signal — here, "the same person, photographed in the mid-1980s".

Because the guidance pushes toward the 1980s, the parts of the image that are free to move — hair, clothes, lighting, backdrop — drift that way. Because the guidance also says "same person", the model keeps pulling the face back toward your original.

Why the face survives

The edit starts from your pixels, not from scratch, and the denoiser is conditioned on the original. Faces are also the thing these models are most heavily trained to keep coherent. That is why a good restyle looks like you in a costume, not like a different person.

Why results vary

Each run injects fresh random noise, so two generations from the same photo differ. Input quality matters: a sharp, well-lit face gives the model a strong anchor; a dark or blurry one leaves more to invention. This is also why the same striking result cannot always be reproduced exactly.

What it is not doing

It is not scanning your face for recognition, building a faceprint, or matching you to a database. It reads pixels, generates pixels, and — in 80s Filter's case — discards your upload afterward. See the face-data guide.