AI portraits

AI portraits redraw the guest's picture in a chosen style — as a comic, an oil painting, an astronaut. Unlike the digital background this is not computed by the booth but by a service on the internet: Replicate. That needs an account there, and the pictures leave the local network.

Everything is set up in the admin area "AI portraits" in eight menu items: Setup (switch and style selection in the flow), Model & costs (model, upscaling), Quota (limit, warning and what happens afterwards), Styles (selection, display for the guest, own wish, groups), Your own styles (own styles, ideas and recipes, example pictures), Background, Labelling and Import and export (style selection by customers, styles as a file). All of it belongs to the event; the access to the provider and the device library belong to the device.

Setting it up#

The access to the provider sits under "Device" > "Access & accounts" > "AI Provider (Replicate)". It belongs to the device, applies to every AI function and stays put when the event changes. The instructions are there under "Setup". Two steps are easily missed:

• Automatic top-up of the credit belongs switched on. If it runs out during the event, guests get their picture without a style - and nobody is standing by to notice.
• A payment method has to stay stored on the account permanently. Accounts that hold nothing but credit are throttled to six pictures per minute. In operation that looks like a fault of the booth.
• The event's privacy notice has to be adjusted before the feature is switched on. See Data and privacy.

"Check connection" names the reason in plain words — wrong key, unknown model, throttling, no credit or no connection at all.

In the menu item "Setup", the block "Ready for the guests" shows what is set up and what is still missing: key, model, enabled styles and limit, plus what affects the AI portraits from outside - privacy by default and offline mode switch them off silently, and custom usage & privacy notices in HTML have to mention the transfer to the provider themselves. Clicking an entry leads to the setting.

Intermediate step or separate mode#

Style selection in the flow decides where the guest picks the style.

• Part of the photo flow — the style selection appears before the filter selection as an intermediate step. Anyone who wants no style passes it with "Continue without AI".
• Separate capture mode — "AI portrait" sits as its own tile next to Photo, GIF and Boomerang. Whoever taps it then picks a style; the photo mode is left without a style selection.

The separate mode has its own layouts, default layout, layout selection and print permission, set under "Capture Modes & Layout Assignment". In the layout editor the modes additionally show an AI portrait box. It captures with the photo camera. When updating to this version, the mode takes over the photo layouts ticked "Use for AI portraits" and the values of the photo mode; nothing changes for the guests.

As part of the photo flow the Use for AI portraits tick in the layout editor still applies: guests do not get to see a layout deselected there once they have chosen a style.

Without a stored key and in offline operation the tile does not appear at all — instead of leading into a flow that ends in an ordinary photo anyway. With the quota used up it stays, greyed out, and its button names the reason instead of “Select”. Vanishing in the middle of the party would be worse: the guest would look for the mode they saw a moment ago.

Model and quality#

Only models the provider keeps warm are offered. A community model would spend minutes booting on its first call — unusable with guests waiting. The price list below the selection shows all models side by side with their price per portrait, in the order of the selection: per quality, plus whether and at what extra cost an own background is possible. The column of the selected quality is highlighted, as is the selected model; tapping a name selects the model. The amounts come from the provider's model pages and carry their date, as there is no interface for retrieving prices. The Background Studio shows the same kind of list for AI cut-out.

Upscaling#

The models deliver 1K to 2K. Without upscaling the portrait is then merely stretched — larger, but not sharper. The step pays off with a 1K model, on large paper and whenever the single portrait is passed on at full size: through the cloud upload of the edited single pictures or the export.

Recraft Crisp is the inexpensive choice for "Same as print" and "High"; Topaz costs more and also reaches "Very high". Both providers see the portrait and therefore belong in the privacy notices. If upscaling fails, the guest still gets the portrait — at its previous size.

Styles#

More than 220 styles in thirteen groups are supplied. Tick what the guest gets to choose from; "Deselect all" clears the list in one go. Next to each group name, "Select all" and "Deselect all" appear again and only affect the styles of that group — handy when only one or two groups suit the occasion.

Each group is a row with small preview images and the counter “12 of 16”; a click opens its tiles. Above them, a field searches all groups for a style, and “Selected only” shows just what the guest gets to choose from.

The groups themselves can be adjusted: rename them, add your own and sort them with the arrows. The order applies to the guest as well — whatever is at the top is what they see first. Renaming applies to the language the admin area is currently in; the other languages keep the factory name. A group without released styles is not shown to the guest.

"Allow combining two styles" lets the guest put two styles on one image, such as "Viking" and "Oil painting". After the first style, a page of its own with its sample image asks whether a second one should be added: "Continue with one style" or "Choose a second style". The second one is picked from the group overview again, the first is marked there, and a tap on the second continues straight away. Both styles go to the provider in a single run: it costs as much as one style and counts once towards the quota. Off by default.

Three rules apply to a combination: the person only stays recognisable if both styles keep them. The Beauty rules only apply if both styles come from the "Beauty" group. And if either of the two is photorealistic, the image is labelled like a photorealistic one.

Own styles#

Below the supplied styles there is an area of its own for styles you create yourself.

Your own style needs only a name and an instruction for the model. The instruction is followed far more reliably in English.

Keeping the person recognisable is what the "Keep the person’s appearance" box in the edit window does. It is set from the factory and appends a sentence when sending that holds face, gender, age, skin tone, hairstyle, beard and glasses — none of that has to be in the instruction. Clearing it is only worth it when the style is meant to change the person on purpose: a different age, a fantasy creature, a caricature. The box belongs to a single style and can be changed for the bundled styles as well.

"Render example picture" sends the same person through your style that is shown on the supplied tiles. That way your style fits into the row instead of standing there with a stranger's face. The run costs the same at the provider as a guest's picture, but does not count towards the event's quota. If a custom photo is stored under "Your own styles" > "Photo for example pictures", the picture is created from that photo.

No supplied style names a work, a brand or a character. What is described is the look - a caped hero without an emblem, a masked vigilante, a northern adventurer. Whoever wants to get closer to one particular original creates their own style; then it is their call.

"Show the guest the groups first" switches on the intermediate step: the guest picks a group and then a style within it. With many styles released that is much quicker than one long list. If only one group has styles, the step falls away by itself.

"Show “Random” button" places a dice button below the selection. With it the guest has one of the released styles drawn at random, sees it with its sample image and accepts it or draws again. Inside an opened group, only its styles are drawn. The button only appears with at least two released styles.

Reference images#

An own style can bring up to three images that the model receives in addition to the guest's photo: the club jersey, the company mascot, a product, a person. Each image comes with a short sentence describing what it shows; the style's instruction says what should happen with it. Both work best in English.

Reference images are only accepted by the models that also handle an own background. An own background takes priority: if the model does not accept all images, surplus reference images are left out, and the edit dialog points this out. Depending on the provider, each image sent along can add a small surcharge. "Render example picture" uses the reference images.

If the model carries the viewing direction over from one person to another - the guest suddenly looks like the guest of honour from the reference - the box "Keep viewing direction" on the reference image helps. Everyone then looks as in their own image: the guests as in their photo, the person or object from the reference as shown there. The box can be ticked on each reference image separately. If a figure from the reference should turn towards the guests, it stays off.

These and further templates are ready under "Ideas and recipes" right below your own styles. "Create as own style" fills the edit window with name, instruction and suitable boxes; the reference image is added there.

For all events: the device library#

Own styles and changed factory styles belong to the event. Whatever should be part of every party goes via the pencil and "Move to device library" into the device library. In the event its styles carry the badge "Device" and can only be switched on and off there.

Example pictures from a custom photo#

Out of the box, all tiles show the same artificially generated person. Under "Your own styles" > "Photo for example pictures" a custom photo can be stored for each event — the bridal couple, the club mascot, the face of the brand. Guests then see on the tiles what the style looks like on that person. This pays off above all at corporate events and weddings, where the tiles on a screen or in the style selection for the customer advertise the booth anyway.

"Render example pictures …" renders the pictures for many styles at once; nothing is preselected. Each picture costs as much as a guest's picture — across all styles this quickly adds up to a double-digit amount. So only select the styles that are enabled in the event. The series takes as long per picture as a capture; the admin area has to stay open meanwhile.

Style selection by the customer#

With more than 220 styles, the choice can hardly be settled over the phone. "Create style selection" under "Style selection by customers" writes the chosen styles with sample images into a single HTML file that goes to the customer by e-mail. It works without internet; the customer ticks the styles in the browser and sends back a small answer file. "Read style selection" then shows which styles change in the loaded event and applies them only after confirmation.

A preselection pays off: offering only the groups that suit the occasion brings a faster answer. Whatever was not in the file stays as it is when reading. Logo, contact note and language are taken from the customer configuration.

Background#

Most styles place the person in a scene of their own — a superhero on a rooftop above the city, a knight in front of a castle. "Background with AI styles" decides whether that stays so:

Styles in the "Beauty" group keep the background anyway; the setting changes nothing for them.

Replace background. In the same sub-section, own images can be uploaded that the model places the person into — the venue, a beach, a castle hall. The chosen style applies to the whole image, including the new background. It works like the Background Studio: switch images active, set one as the main image, add labels. If several are active, the guest picks one after the style; "Random" and "Continue without background" can be switched on. Without a selection, the main image applies.

Images are reduced to at most 2048 pixels on the long edge and travel with the event when it is duplicated, exported or used in presets. Some models charge extra for every image sent along; a portrait with an own background can therefore cost a little more.

Questions for the guest#

An own style can ask the guest up to three questions, for example "What colour should the jersey be?". The answer goes into a sentence that belongs to the style - "The football jersey is {answer}." - and so stands where the model understands it. The free wish from the next section, by contrast, hangs loosely behind the whole instruction.

This is useful wherever a style leaves a choice open: jersey colour, shirt number, favourite club, decade, hair colour of the comic character. It is set up in the edit window of the style under "Questions for the guest".

The guest sees the questions after choosing the style, one per page, and can skip each of them; its sentence is then left out and the model decides for itself. With two combined styles the questions of both follow one after the other. Questions are available for own styles of the event and of the device library; they travel along when exporting and in the event package.

Style choice by the guest#

With "Allow the guest to add a wish" switched on, the guest can append a short sentence to the chosen style - for instance "astronaut on the left, knight on the right". It is meant for group pictures where each person is to become something different. The wish adds to the style, it does not replace it, and the sentence that keeps the person stays in place.

The input is off out of the box. What the guest types goes to the provider as an instruction: if they name protected characters, its content filter decides what comes of it. And typing costs time at the booth - with a queue that is the real price.

"Always mark when a wish was added" is on out of the box: where the guest writes along, the picture is more likely to show people doing something that never happened. The rule then applies even when labelling is otherwise set to "Never"; it can still be switched off.

Quota#

The count is listed under Counters and limits and can be reset there on its own. Only what was actually computed is counted — a failure at the provider consumes nothing.

Limit and message are also listed under Counters and limits - the same field, reachable in two places. Once the quota is used up, the guest gets a notice card with the stored text instead of the style selection and can go back or continue without a style from there.

If the rest of the quota is no longer enough for all photos of a layout, the whole session stays without an AI style, and a notice tells the guest so. For layouts with several photo slots a limit that is a multiple of the number of slots therefore fits best - otherwise a remainder is left unused at the end.

Below "Runs per event" there is the cost frame: the price per run of the selected model times the limit. Without a limit it only names the price per run - the total stays open. The badge in the header of "Quota" shows "Active" as soon as a limit is set.

Labelling#

The AI Act requires visible disclosure where a picture could falsely be taken for real. A cartoon style is not covered, a photorealistic portrait is. Out of the box exactly the photorealistic styles are labelled; the default can be switched to "Always" or "Never" and the position of the label chosen.

The operator of the booth makes that call — under the regulation they are the one deploying the system.

What the guest sees#

The style is chosen in a step of its own right after the mode selection, before the filter and layout selection: it determines which layouts fit and whether a digital background is still asked for. Filter and AI style can be combined: first the model redraws the picture, then the filter is laid over it. If "Filters for AI portraits too" is switched off in the filter settings, the filter selection is skipped after a style. If "Offer for AI portraits too" is switched off under "Decorate image", an AI portrait goes straight to the end screen without decorating. All pictures of a session get the same style.

The step does not appear at all when no key is stored, when the booth runs in offline mode or when the data-minimising default is switched on. In none of these cases does the guest see an error message.

While the picture is being computed the waiting screen says that the portrait is being drawn and shows the remaining time. The duration is measured on every run, so the display appears from the second shot with AI onwards.

Layouts and framing#

In the layout editor you can set per layout whether it comes into question for AI portraits. Out of the box all of them do. Once the guest has chosen a style, deselected layouts are no longer shown - and if multi-picture layouts are not released for AI portraits, only those with a single placeholder remain. Whoever chose "no AI" still sees them all.

The capture goes to the provider in the crop of its layout slot. The model therefore computes for exactly the format that will be seen; otherwise fitting it in would cut away the very part it worked on.

Three reasons can only be fixed by the operator: missing credit, an invalid key or none at all. They go out as errors and trigger the immediate status notification, if set up - otherwise the event would run on without anyone learning of the missing portraits.