New setting. The same person.
Generating a portrait with attractive lighting is one task. Keeping a person recognisable in a new setting is the harder one. A face can look convincing while its eyes, mouth or head shape have quietly changed.
We developed a system for this: it organises existing images, selects suitable references, generates new scenes and checks facial similarity. AI handles the creative work and image generation. Software measures and filters; a person judges the result.
Three faces to explain the process
Mira, Jonas and Amara were invented and generated with AI for this article. Their starting images are synthetic too. They are not clients, employees or real people whose photographs we are publishing.
- Mira: Curly hair, freckles and a distinctive face shape.
- Jonas: Glasses, a short beard and greying hair.
- Amara: Short natural curls and delicately defined features.
The next series changes clothing and surroundings. Look particularly at the eyes, nose, mouth and head shape: what stays consistent, and what appears to change?

From a folder of images to a working process
For real people, the developed workflow starts with their consent and approved original photographs. What matters is which images are suitable for which situation.
| Step | What the system handles |
|---|---|
| Organise images | Catalogue head angle, expression, framing and suitability |
| Select references | Assemble facial angles and expressions suited to the requested scene |
| Generate a scene | Change setting and clothing while preserving distinctive features |
| Compare separately | Compare the result with separate scoring images from the same angle |
| Filter results | Archive measurable results below the relevant threshold and retry with other references |
| Review side by side | Display the generated image, generation references and scoring images in a comparison gallery |
Automatic reference selection excludes the scoring images. Otherwise the comparison would stay too close to the input images. Retries are limited: the system makes at most three attempts rather than generating indefinitely.
What 0.75 actually means
The code locates faces with YuNet, then aligns them for feature comparison with SFace. A similarity value is calculated against each usable scoring image; the displayed value is their mean, rounded to three decimal places.
These are configured quality thresholds in our system, not results measured for the demonstration characters shown here:
| Head angle | Minimum for the automatic filter |
|---|---|
| Frontal | 0.75 |
| Three-quarter | 0.60 |
| Profile | 0.55 |
The gallery adds another distinction: green starts at 0.75. A side view can therefore reach its lower filtering threshold and still appear yellow. 0.75 does not mean “75 per cent the same identity”. Changing pose changes the comparison, which is why angles are treated separately. If no face can be measured, there is no passed numerical check: the system flags the image for manual review.
Another angle reveals more
This series changes head position and expression. With a smile, the mouth, teeth and cheeks deserve particular attention. A technical facial comparison cannot replace a visual assessment of these details.

We do not claim a passed quality measurement for this synthetic image demonstration. It illustrates the task and the visual review. The documented filters and comparisons come from the developed system; testing against original photographs of real people would be a separate run.
What this means for your project
The useful combination is creative generation, suitable references and a traceable quality filter. It gives a more deliberate way to develop a series than generating every scene independently.
Whether a series suits a website, profile or campaign depends on the intended style, available material and results actually achieved. A suitable numerical score alone is not a promise of image quality.
Tell us about your image task or learn to use AI on your own project. Our knowledge article on AI images and rights discusses publication questions.
Sources.
- SolvraONE: entwickeltes Porträt-System, Dokumentation und Code-AbgleichOwn source · As of
- Entwicklung des Porträt-Systems seit November 2026Own source · As of
- OpenCV: Gesichtserkennung und Merkmalsvergleich mit YuNet und SFaceVendor / primary · As of



