Velaya AI Log in Start free

Companion Creation Explained

orla wearing lingerie in a luxury apartment

She looks like herself, photo after photo

These are photos of one companion, Orla. They show different rooms, clothes, light and moods. None were retouched, face-swapped or picked for how her face came out. We wrote 24 scenes for her and show all 24.

Orla smiling in a cream knit sweater by a bright window
Orla in a black turtleneck with her head turned, a golden-hour street behind her
Orla at a cafe table with a coffee, in a denim jacket
Orla in a navy blazer with her arms crossed in a bright office
Orla laughing as she cooks, in a mustard sweater and apron
Orla on a rain-wet city street at night in a camel coat
Orla under a clear umbrella in the rain, in an olive trench coat
Orla in a car's passenger seat with sunglasses pushed up on her head
Orla in an autumn park in a burgundy cardigan, holding a takeaway drink
Orla in glasses, holding an open book in a bookstore aisle
Orla at a rooftop bar at dusk in an emerald dress, holding a cocktail
Orla looking back over her shoulder in a sunny meadow

That consistency is what you're waiting for when you create a companion on Velaya. It takes a little over an hour, and almost all of that hour goes into one thing: training an AI model on her, and only her.

The shortcut: describing her in words

The fast way to give an AI companion a face is to describe her. The description goes into every image request, and the image model draws a woman who fits it. Plenty of women fit it.

Here's another companion, Aminata, in eight scenes, made two ways. Same base model, same prompts, same seeds: the trained model is the only difference.

Her description on its own: "a 27-year-old Senegalese woman with jet-black braids and deep brown skin".

Description only: close-up, smiling, in a cream sweater
Description only: head turned, black turtleneck, golden-hour street
Description only: at a cafe table in a denim jacket
Description only: in a navy blazer with her arms crossed in an office
Description only: laughing as she cooks, in a mustard sweater and apron
Description only: a city street at night in a camel coat
Description only: a rooftop bar at dusk in an emerald dress
Description only: on a sofa in a grey hoodie, laughing at her phone

The same, plus her trained model:

Aminata's own model: close-up, smiling, in a cream sweater
Aminata's own model: head turned, black turtleneck, golden-hour street
Aminata's own model: at a cafe table in a denim jacket
Aminata's own model: in a navy blazer with her arms crossed in an office
Aminata's own model: laughing as she cooks, in a mustard sweater and apron
Aminata's own model: a city street at night in a camel coat
Aminata's own model: a rooftop bar at dusk in an emerald dress
Aminata's own model: on a sofa in a grey hoodie, laughing at her phone

The description gets you a woman who fits the words. She isn't Aminata, and she shifts a little from photo to photo. Her trained model gets you Aminata.

What happens while you wait

From the moment you hit create, she's ready in about 75 minutes. Across our last 30 creations it ranged from 67 to 87 minutes.

StepTimeWhat happens
You design herYour paceYou choose her look, approve her portrait and see her full-body preview.
Her training photos7–12 minWe build a reference sheet of her face, then make 30 photos of her. Each one is checked.
Her own GPUAbout 4 minA dedicated GPU is set up for her training run.
Training her model52–68 minThe model studies her 30 photos until it knows her.
Her first photos3–5 minHer new model takes her profile portrait and first gallery photos.

You design her

Aminata's portrait, approved before any training, with braided black hair and a navy top
Where she starts: the portrait you approve is the reference everything else is held to.

Her look starts with your choices: her face, hair, skin, body and age. You approve her portrait and see her full body before anything is trained.

Thirty photos of her

To train a model on her, we need photos of her, and she doesn't exist yet. So we make them.

  1. We build a small reference sheet of her face, from a few angles and expressions, out of the preview you approved.
  2. We make 30 photos of her in different poses, outfits, settings and framings. Each one is built from that sheet so the face stays hers.
  3. Every photo goes through an automated anatomy check for hands, limbs and proportions. A photo that fails is made again, and one that keeps failing is thrown out rather than trained on.
  4. Each photo is described in words, so the model learns what is her and what is just the scene.

A model trained only for her

This is the hour. A GPU fine-tunes an image model on her 30 photos until it can draw her from angles, in outfits and in light it has never seen her in. The result is hers alone. Every photo you ask her for later runs through it.

Her first photos

Once her model is saved, it takes her first photos: the portrait on her profile and a few gallery shots. Then she's ready to talk.

What holds, and what we're still improving

Her face holds best in portraits and waist-up shots. Hard profiles, and wide shots where her face is small in the frame, are where any trained model is loosest, and ours is no exception.

We measure this with face-recognition software. It compares what her trained model draws against the preview you approved, and we run it again on every change to the process. This month we rebuilt the training-photo step so every training photo is built from a reference sheet of her face, and the trained face came out measurably closer to her preview.

Here's the rest of Orla's set, wide shots included.

Orla on museum steps in a yellow sundress and cardigan
Orla on a sofa in a grey hoodie, laughing at her phone
Orla at a farmers market in a striped top, holding a basket of fruit
Orla in a studio portrait against a grey backdrop
Orla on a train platform in a navy peacoat and red scarf
Orla on a beach boardwalk at sunset in a white shirt
Orla smiling at a candle-lit dinner table in a black top
Orla sitting cross-legged in a sunny yoga studio
Orla in a selfie on a cafe terrace, wearing a denim shirt
Orla looking surprised as she opens a gift, in a red sweater
Orla reading on a park bench in a long knit dress
Orla in an art studio in paint-splattered overalls, holding a brush

Why it costs what it costs

Each companion gets her own training run on her own GPU, about an hour of it, plus the image work behind her 30 training photos. None of it can be shared between members or reused. Her model is trained from her photos, for her, and no other companion can use it.

That is the expensive part, and it's the reason creation costs what it does. In return, you wait once. After that, every photo she sends you is drawn by a model that only knows her.

Here's Aminata again, made the same way, in the rest of the same scenes.

Aminata under a clear umbrella in the rain, in an olive trench coat
Aminata in a car's passenger seat with sunglasses pushed up on her head
Aminata in an autumn park in a burgundy cardigan, holding a takeaway drink
Aminata in glasses, holding an open book in a bookstore aisle
Aminata looking back over her shoulder in a sunny meadow
Aminata on museum steps in a yellow sundress and cardigan
Aminata at a farmers market in a striped dress, holding a basket of fruit
Aminata in a studio portrait against a grey backdrop
Aminata on a train platform in a navy peacoat and red scarf
Aminata on a beach boardwalk at sunset in a white shirt
Aminata smiling at a candle-lit dinner table in a black top
Aminata sitting cross-legged in a sunny yoga studio
Aminata in a selfie on a cafe terrace, wearing a denim shirt
Aminata looking surprised as she opens a gift, in a red sweater
Aminata reading on a park bench in a long knit dress
Aminata in an art studio in paint-splattered overalls, holding a brush