Photo-to-Photo Generation: How Likeness Is Actually Preserved

Photo-based generation solves a problem no text description can: putting a specific person in the frame. The mechanics of that preservation are not obvious, and they explain most of the trade-offs users run into.

What is actually inherited

The model extracts a set of appearance features from the source: facial proportions, the shape of individual features, skin tone and character, hair type, distinctive marks such as moles and the shape of the hairline. That set becomes a constraint within which a new image is built.

Crucially, what is inherited are properties, not pixels. That is why the result can show you from another angle, under different light and in different clothing — but with your face.

The freedom control

The process has a parameter called different things in different interfaces: transformation strength, source influence, reference weight. Physically it sets how heavily the source image is noised before work begins.

• Light noising: the result stays close to the source, likeness is maximal, but the scene changes little.
• Medium: features are preserved while light, background and pose change — the working range for most tasks.
• Heavy: near-total freedom, only a general type survives from the source, likeness is lost.

Why you cannot have both

This is not a limitation of a particular service but a property of the method. Information about your appearance lives in the same data that has to change for stylisation. The more you allow the model to rewrite, the more it rewrites your appearance too.

The practical conclusion: decide what matters in this particular frame. For an avatar where recognition matters, take less stylisation. For a decorative illustration where expressiveness matters, take more and accept the loss of exact likeness.

What strengthens likeness at the same stylisation

1. A large, sharp face in the source — more features are extracted.
2. Soft light that shows volume: features read more accurately.
3. A style whose lighting is close to the source: less needs rewriting.
4. One person in frame: features do not blend.
5. No beauty retouching: individual markers survive.

Several photos as input

Some scenarios accept a set rather than a single shot. That improves accuracy: the model sees the appearance from several angles and under different light, and its representation becomes fuller.

But it only works when the shots agree: the same person, a close period of time, comparable quality. Mixing photographs from different years yields an averaged age rather than a better likeness.

Why results look younger or older

A frequent observation: a generated portrait reads several years different. The reason is that age is carried by skin micro-texture and fine detail — exactly what is lost first when features are extracted.

Smooth skin in the output is not the model intending to rejuvenate you but a loss of texture. Stating texture explicitly in the request compensates partly.

Checking the source before you start

It is worth thirty seconds of assessment before a run. Open the photo and crop everything but head and shoulders. If the face stays sharp and readable, the source will work. If it turns into a soft blob, any generation from it will be averaged, whatever settings you choose.

When photo mode is not needed

If no recognisable person belongs in the frame, a source only constrains the model: it consumes capacity holding on to an appearance that could have gone into the scene. For abstract illustrations, backgrounds and invented characters, generating from scratch gives a better result.