AI Avatar Generator From a Photo: How It Works

In AIGE an avatar is a saved character: a name, an avatar picture and a gallery of reference photos. Nano 2.1, Nano Pro and Nano Lite draw on those photos to keep the face across new pictures. Saving costs no credits of its own; each picture costs what its model costs.

What AIGE saves when you save a character

A character is a record in your account with three parts. The name is what you call the person. The avatar is the one picture that stands for them. The gallery is the pile of reference photos the models look at when you ask for a new picture of that person. The saved photos are used as references: each request sends them to the image model, and the model reads them again every time.

One naming quirk. In the classic grid studio, the popup that creates a character is titled "Create project", because the same record also holds that project's history.

The guide to consistent AI characters walks through a first session. This is the route map under it: where the photos go in, which of them get sent, and what that costs.

Where the photos go in

On a laptop there are two ways in, and both write to the same record.

1. Through the agent in the prompt island. Put the dial on Fast, Auto or Deep, so an agent runs, because Manual sends your words straight to a model and has no agent. Use the plus button, labeled "Add a picture, a file or a skill", to attach a photo, then name the person: "name her Mara", "this is Sam". The agent saves a character with that photo as its avatar. It can seed the gallery with up to 8 more pictures when it creates the character, and later add up to 20 per request. A picture can be up to 10 MB, and AIGE converts it to a JPEG when it saves it. Saving a character is an action the agent can ask your approval for first.

2. In the classic grid studio at aige.ws/grid. The project switcher opens a "Create project" popup with one field, a name. Then you upload pictures into that character's gallery, and the upload dialog's button reads "Save to Gallery". The grid takes image files such as PNG, JPG and WEBP and turns away HEIC files until you convert them. In the current code the canvas's tool menu has the link to the grid switched off, so you open that address yourself.

Assistants connected over MCP, such as Claude, ChatGPT or Cursor, can also create a character and add photos by link, and they land in the same gallery.

How many photos? AIGE enforces no minimum: a name alone creates a character in the grid, and the agent route needs one photo for the avatar. AIGE's own character page suggests starting with a sharp front portrait, a three-quarter view, a profile and a full-length shot. The ceiling is on the sending side. On the agent route, one request carries at most ten reference pictures, the ten newest uploads, so a bigger gallery is a pool to draw from, not a bigger set sent each time.

What the character does when you generate

Ask the agent for a picture "of Mara" and it passes Mara's id with the request. The server then adds her saved photos as references. A few rules sit behind that one step:

• Face photos first. The character's uploaded photos lead the list, ahead of any picture you attach. AIGE's code does this because it treats the first reference as the one that anchors the face.

• Your own picture rides along. Say "in the style of this picture, with Mara" and both go to the model, while there is room: one request carries ten pictures at most and the character's photos go first, so with ten or more saved photos your attached picture is left out. It never replaces the character.

• The avatar is the fallback. It goes alone only when the character has no uploaded photos.

• Face or body. The agent can ask for the face photos, for the character's generated gallery shots when build and wardrobe matter, or for both.

Not every model can use them. The table follows AIGE's own routing guide and its code, and the picker names are the ones on the Make sheet. The model picker article covers the rest of that sheet.

The grid has a second route: typing @ and a character's name. The phone board has its own character picker, and in Manual it resolves the same way. Both send the character's avatar, or a registered face set when the character has one, and do not read the gallery. On a Fast, Auto or Deep turn the phone board hands the character to the agent instead, which sends the saved photos. Pictures you pick yourself still go along. That is why the same character can behave differently on the two routes.

Does the character carry into video?

On some routes it does, and on one it does not.

• Through the agent. Ask for a clip with a saved character and no start picture, and the clip starts from one of that character's saved pictures, or from the avatar if there are no photos. A start picture of your own is used instead of the character's.

• In the grid, with @ and a name. On Kling V3 the character becomes an element: its first picture is the front view and the rest are extra references. Video models that take a single picture use the character as the start frame only if you gave none. If you did, the grid keeps your start frame and says so.

• In the Canvas video timeline. It makes clips under the canvas's own placeholder, not under a saved character, so it does not take one. Make a picture of the character first, then start the clip from that picture. The photo-to-video steps cover that route.

In every case the character sets the first frame and the video model does the moving. AIGE does not lip-sync, so a character does not speak on cue.

What it costs

Saving a character spends no credits of its own. Through the agent, each turn still carries the agent's fee, which starts at 2 credits a turn. Making pictures is where the credits go, and the price follows the model, the size, the quality and the count. The saved photos add nothing: the charge has no input for them. Here is one picture and a set of six, at the settings named:

So six Nano 2.1 pictures of one character at 1K come to 48 credits, before any agent fee. Nano 2.1 takes 0.5K to 4K and Nano Pro 1K to 4K, and each step up costs more, so name the setting before you trust a total. GPT Image 2 is priced by quality and shape, so it is not in the table; the model picker article has its tile. In Manual, Send shows the exact price. On an agent turn it shows the turn fee's floor, and the Make sheet shows the price of a model you pin. Video prices live in the AI video for YouTube article, which owns that table.

Whose face you can use

AIGE's Acceptable Use Policy names this directly. It rules out realistic synthetic media of a real, identifiable person without that person's documented consent. It also rules out generating content with another person's face or likeness unless the face is your own, you have their permission, or you hold a valid AIGE license for it, such as a FaceShield license or a character rental. That is a paraphrase, and this article is not legal advice, so read the policy for the exact wording.

The practical reading for a photo-based avatar: start from your own photos, or photos of someone who agreed. On the agent route AIGE also screens each reference picture, and a refused one comes back as "That reference image is outside what the platform allows. Attach a different one."

Honest note: a saved character is a strong anchor, not a guarantee. AIGE's own character page says drift grows with distance from the references, with extreme angles, heavy stylization, fast head turns and long clips, so look at every picture and regenerate the ones that wandered. The routes send different sets of photos (the agent up to ten, an @ mention the avatar), so do not expect one route to match another. This article follows the laptop. A phone shows the classic grid or, on accounts that have it, the phone board, and the phone-versus-laptop article sorts out which. Nothing here was run with a live account: it is read from AIGE's code and pages as of October 2026, so trust the screen where it differs.

Quick answers