AI girlfriend photos and videos: how they work

How AI companion photos are made

A photo of an AI girlfriend is made by an image model: software trained on a very large collection of pictures and their descriptions, so that it can produce a new image from a written one. Most current models work by starting from random visual static and refining it, step by step, until it matches the description it was given. Nothing is photographed, and no existing photo is edited.

For a companion, the description has two layers. The first is fixed: the character's own look, from hair and eye color to build and style, which is what keeps her recognizable from one picture to the next. The second is what you ask for in the moment: a pose, an outfit, a setting, a mood. Apps usually offer those in one of three ways:

• Presets, where you pick a pose, an outfit and a scene from a list. Quick, predictable and a good place to start.
• Your own prompt, where you describe the picture in words. More control, more room for surprises.
• Edits, where you start from a picture you already have and change part of it, such as the outfit or the background.

Writing a prompt that works

If you write your own description, a little structure goes a long way. Name the setting, the outfit and the mood in plain words (“on a rooftop at sunset, in a denim jacket, laughing”) rather than stacking adjectives. Ask for one idea at a time; prompts that pull in two directions tend to produce muddled pictures. Leave her looks out, because the app already knows them, and adding your own can fight with her profile. And if a prompt doesn't work, change one thing at a time so you can see what made the difference.

In a chat, the same thing happens behind the scenes. When a companion “sends” you a photo, the app has generated one to match her look and the moment in the conversation.

How AI companion videos are made

Video works a little differently. Companion apps typically make video from a still image: you choose a photo, and a video model animates it into a short clip, predicting how the scene would move over a few seconds. Some also generate sound to go with it.

Three things follow from how this works:

• Clips are short. Seconds rather than minutes, because keeping every frame consistent gets harder the longer a clip runs.
• They take a while. A video is many images that have to agree with each other, so a render usually takes minutes, not the few seconds a photo takes.
• They cost more. More computing means a higher price than a photo, and longer or higher-resolution clips usually cost more again.

Always AI-generated, always fictional

A reputable companion app makes pictures of its own fictional characters and nothing else. It won't let you upload photos of yourself or other people to build a companion, and it won't make images of anyone who exists. That isn't a technical limit; it's a rule, and an important one. Pictures of a person made without their agreement can be used to embarrass, harass or mislead, and the person in them has no say in it.

Lerk AI follows this rule: it doesn't accept uploads, every companion is a fictional adult never based on anyone who exists, and every image and video is generated by AI.

How pricing usually works

Generating pictures and video costs real computing power, so apps usually charge for them in some way, even when chatting is cheap or included. The common models are:

• Credits per image or video. You hold a balance of credits (stars, gems, tokens) and each picture or clip uses some. Videos use more than photos, and higher resolutions use more than lower ones.
• A monthly allowance in a subscription. A plan adds a set number of credits each month, sometimes with a discount on extra purchases.
• A mix of both, which is the most common: a plan with an allowance, plus the option to buy more.

Whichever model an app uses, four questions save a lot of frustration:

1. Do you see the price before you ask for a picture?
2. If a generation fails or is blocked, do you get your credits back?
3. Do bought credits expire, and does a monthly allowance roll over?
4. Are videos priced by length, by resolution, or both?

Our checklist for choosing an AI companion app puts these alongside the other billing questions worth asking.

Quality: what to expect

Today's image models are very good, but they aren't flawless, and knowing their weak spots makes the results easier to enjoy:

• Consistency can drift. A companion should look like herself in every picture, but small details such as a hairstyle or an accessory may vary from one image to the next.
• Fine details are the hardest part. Hands, fingers, jewelry, writing on signs or clothing and busy backgrounds are where odd results most often appear.
• Motion can warp. In video, fast movement or complicated scenes may bend or blur for a moment.
• Style matters. Illustrated, anime-style pictures often hide small flaws better than photoreal ones. Our guide to anime vs realistic companions covers how the two styles compare.

If a result isn't right, a simpler prompt or a preset usually does better than a longer, more detailed description.

Privacy of your gallery

The pictures you make can be personal, so it's worth knowing what happens to them. Before you start, check three things in the app's privacy policy and settings: whether your images are private by default or shared to a public feed, whether you can delete them yourself, and whether the app says anything about how it uses them.

On Lerk AI, your gallery is private. Chats, characters and media are private to your account and never published, and you can delete them whenever you like. Our page on whether Lerk AI is safe covers privacy in full.

How it works on Lerk AI

Here is how all of the above looks on Lerk AI, the app we make, in brief:

• Photos in chat. Ask for a photo or tap the photo button; you see the price in stars before you send, and it arrives in your chat once it's ready.
• Surprise photos. Now and then a companion sends a locked photo with a blurred preview, which is yours to keep once you unlock it.
• The studio. Photos from presets (pose, outfit and scene) or your own prompt, in portrait, square or landscape, plus edits of photos you've saved.
• Videos. Turn a photo into a short clip with generated sound, in HD or Full HD. A clip takes a few minutes to render.
• Refunds on failures. If a generation fails or is blocked by a safety check, its stars come back automatically.

For the details, including video lengths and what each photo and video costs today, see photos and videos on Lerk AI. How stars and the plans fit together is in Lerk AI pricing.

Frequently asked questions