AI character creator

One face.
Every scene after.

Start with a photo or a line. Pick the face, build its model sheet, and bring the same character into images, stories and video.

New character

Characters is on Hobby, with every image model. Video is on Starter. See plans

Or start from a ready-made character →
Mira · The face
Mira The face

Same face, every time

One sentence in. A person who stays put.

Sam started as one line of text. The anchor is the first render; the five angles that follow are separate generations conditioned on it — not crops, not filters, not one image rotated. Compare the freckles and the silver hoop.

Sam — the anchor image, generated from a single line of text
Anchor · rendered from the typed line
Sam — Headshot reference angle
Headshot
Sam — Front reference angle
Front
Sam — 3/4 Left reference angle
3/4 Left
Sam — Side reference angle
Side
Sam — Full body reference angle
Full body

These are separate renders from an earlier character session. The current Face preset is Front, three-quarter, Side, Back and Full body. Reference support varies by model: some use the anchor, others use multiple views to help preserve identity.

How it works

From a sentence to a cast member. No training, no LoRA.

  1. 01

    Describe or upload

    One sentence is enough — "a Korean dog-treat founder, long blue-black hair, a silver hoop". Or start from a photo you already have. No dataset, no training run, no LoRA to wait on.

  2. 02

    Get an anchor

    That line gives you faces to choose from. Keep one to draft its name, appearance and notes. The chosen face is the anchor: everything downstream conditions on it rather than on your original wording.

  3. 03

    Fill the sheet

    Five Face tiles — front (the anchor), three-quarter, side, back and full body — make up the standard model sheet. Build the remaining views from your chosen face. This is the file the models read when they need to know who this person is.

  4. 04

    Make things with it

    Point the sheet at a new scene, a video model, or a storyboard. The face carries. You are no longer prompting for a person, you are casting one you already have.

One character, every medium

Lock the face once. Use it everywhere.

A saved character is an input, not an output. Open it and the next step is a menu rather than a blank prompt box — every one of these starts from the same reference sheet.

New photo — made with the Sam reference sheet
New photoHobby · Characters

Put the character in a new scene or outfit — one image, conditioned on the whole sheet.

Reel — made with the Sam reference sheet
ReelHobby to plan

Pre-cast into a storyboard: shots written in order, keyframes rendered with the same face.

Short clip — made with the Sam reference sheet
Short clipStarter

A ~5s clip built from the full reference set, so it reads as the same person and not one animated frame.

Talking video — made with the Sam reference sheet
Talking videoStarter

The character speaks to camera, driven by the voice already bound to them.

Not just a face

Your character has a voice.

Choose the voice that fits the character from a library of 45 English voices. It is what talking video speaks with and what storyboard narration reads in, so the character stays the same character once it opens its mouth.

  • Expressive Narratorexpressive, warm, storytelling
  • Confident Womanconfident, poised, clear
  • Trustworthy Manwarm, trustworthy, steady
  • Anime Heroineexpressive, animated, bright
  • Magnetic-voiced Mansmooth, magnetic, confident
  • Graceful Ladygraceful, elegant, refined
  • Wise Scholarthoughtful, scholarly, measured
  • Playful Girlplayful, cheeky, light
  • Deep-voiced Gentlemandeep, refined, dignified
  • Wise Ladywise, measured, grandmotherly

A sample of the catalog. Matching is free — you are not buying a voice per character.

Or start from ours

Six characters, already built.

A photoreal woman, an anime heroine, an elven ranger, a cyberpunk operative, a robot mascot and a forest creature — each with a reference sheet you can generate against straight away.

No surprises

Character work, minus the fine print.

Characters start on Hobby

Browse our characters and write your idea before subscribing. Hobby unlocks character creation and every image model inside Characters, including Nano Banana 2 and Pro. Drafting and the first five Face tiles are included.

One character, many scenes

Keep the face, notes and reference sheet together. Reuse them in Studio, Apps and Storyboard without training a separate model.

Failed renders are refunded

A generation that fails is refunded automatically. Completed renders are charged, even if you choose to try another take.

Monthly, cancel anytime

Putting a character in video — short clips and talking video — unlocks on Starter, $9 for your first month, then $29/mo (or $19/mo billed annually). No forced annual contract, downgrade whenever.

What is an AI character generator?

An AI character generator makes a person — or a mascot, or a creature — that you can generate again. That second half is the whole job. Any text-to-image model will hand you a striking portrait; almost none of them will hand you the same portrait subject twice, because the only thing carrying identity between runs is your prompt, and a prompt describes a category rather than a face. “A woman with dark hair and freckles” has a few billion valid answers, and the model picks a different one every time.

The fix is to stop describing and start referencing. A character here is a small set of images of one subject — five Face tiles, including the anchor — that later generations condition on directly. The models are no longer guessing what you meant; they are matching a face they can see. That is why the same character survives a change of scene, outfit, lighting, or medium.

The older way to get this was to train a LoRA: gather twenty or thirty images of a subject, pay for a fine-tuning run, wait, then do it all again for the next character. It works, and it is a substantial amount of work for one cast member. Reference conditioning skips the training step entirely — the sheet is generated in a single pass and is usable immediately — which is what makes keeping a whole cast around practical rather than a project.

What separates tools in this category now is how far the character travels after it is made. A face that holds across still images is table stakes. The useful question is whether the same character can walk into a video model, speak with a voice that belongs to it, and hold across the six shots of a storyboard — because that is where drift is most visible and most expensive to fix by hand.

FAQ

Straight answers.

What is character consistency, exactly?+
It is the ability to generate the same person again — same bone structure, same freckles, same silver hoop — in a scene the model has never rendered before. Ordinary text-to-image gives you a new stranger every run, because the only thing carrying identity is your wording, and wording is not precise enough to specify a face.
Do I need a photo to start?+
No. A sentence is enough — you choose a face from the first results, and later views use that anchor. You can also upload a photo you already have and use it as the anchor instead. Both paths end at the same place: a reference sheet the models can read.
Why do faces drift in normal AI generators?+
Because nothing is holding the identity between runs. Each generation samples fresh from the model, and a prompt like "a woman with dark hair" describes a category, not a person. Locking a character replaces that description with actual reference images, so the model is matching a face rather than re-inventing one.
Is this the same as training a LoRA?+
No, and that is the point. A LoRA means collecting a dataset, paying for a training run, and waiting — then repeating it whenever you want a new character. Here you choose an anchor and build its remaining views when you are ready. Nothing is fine-tuned, so there is nothing to retrain.
What angles do I get?+
Five Face tiles: Front, three-quarter, Side, Back and Full body. The anchor is Front; Build the rest fills the other views. Outfits and Props are named slots you add and render when needed. The number of references used depends on the model.
What is the voice for?+
Choose a voice for the character or use the voice-matching tool. It is what talking video speaks with and what storyboard narration reads in, so a character stays the same character when it opens its mouth.
Can I put my character in video?+
Yes — short clips and talking video both take the character. Video models are on the Starter plan and up; Hobby includes building the character with every image model inside Characters. Short clips, talking videos and Storyboard Animate require Starter.
Does it work for non-human characters?+
Yes. Mascots, creatures, and stylized characters all work the same way — the anchor just happens not to be a person. Two of our own official characters are a robot and a forest creature for exactly this reason.
Can I use the characters commercially?+
Yes. What you generate is yours to use, including commercially. Characters you create stay private to your account unless you publish them.
How many characters can I have?+
As many as you want — characters themselves are not metered. Credits are spent on renders, not on the act of keeping a cast around.

Who are you casting?

One line of text is enough to start. Characters is included with Hobby, with every image model. Video of a character is on Starter.