AI profile pictures and avatars — one face across a whole set

An avatar is a portrait with two unusual requirements. It gets displayed at about the size of a fingernail, and it has to be the same person tomorrow. Neither of those is a prompt-writing problem exactly — they are settings problems, and both have a clean answer in the Image studio.

Square, every time

Set the format to 1:1 before anything else. It renders at 1024 × 1024, and it is the only format that nobody downstream will crop, because every platform that shows a profile picture shows it in a square or a circle inscribed in one.

Render a 3:4 portrait for an avatar and something else decides which part of it you get — usually by taking the center, which is somewhere around the collarbone. Render the square yourself and the composition is yours. If you also want a banner or a cover from the same session, render that separately at 16:9 rather than crop it out of the portrait, because a native frame is composed for its shape and a trimmed one never is.

Frame for the circle, not for the wall

Most platforms mask the square into a circle, which eats the corners. So:

• Crop in tight. Head and shoulders, the chin around a third of the way up. A three-quarter body shot becomes an unrecognizable smudge at display size.
• Keep the head off dead center. Slightly high in the frame reads as confident; dead center reads as a passport photo.
• Leave the corners quiet. Whatever lives in them is going to be cut, so put nothing there you would miss.
• One value break. A light face on a dark ground or the reverse. Detail is lost at small sizes; contrast is not.

Backgrounds should be plain on purpose. A seamless studio wall, a soft gradient, an out-of-focus interior. Anything busier competes with the face and loses at 32 pixels.

Which model

Three of eroq's own engines cover most avatar work, and two of them cost 10 credits an image, so the choice is mostly about finish.

• Krea2: photoreal, and it reads up to four reference pictures. Use it when the avatar should read as a photograph of someone: professional profiles, a host face for a channel, a realistic cast for a product.
• Anima: anime, manga and illustration, with a negative prompt. The right pick for community avatars, stream identities and any cast where a consistent line style is the brand. It tolerates a compact comma-separated prompt well. It reads no reference pictures, so for one face across a set lean on a Character.
• Eroq One: reads up to ten reference pictures, at 10 credits an image. Use it when a set has to hold one face from several photos.

If you are genuinely unsure, run the same subject on two of them and stop guessing. The third-party image engines on the model list are there too when you want another look.

Holding one face across a set

Here is the part that matters if you need more than a single picture — a cast for an app, a team page, a set of variants for different platforms.

A reference photo is the direct route. Attach one and the render starts from that picture instead of inventing a new person. Use a clean, front-lit, sharp source with the face filling the frame. Note that with a reference attached, the format defaults to Retain, so the source image decides the frame: start from a square reference, or pick 1:1 yourself, when you want a square result.

A Character is the durable route. Save a name, a persona and reference photos once in Characters, then @mention that character in any prompt afterwards. The same face is then available across every image you render, and across clips too, since a Character can be animated through image-to-video and can carry a voice. For a cast of twelve, this is the difference between a project and a filing problem.

What breaks a face: a different model mid-set, a heavily stylized prompt fighting a photoreal reference, and describing the person again in words when the reference is already doing that job. There is more on that failure mode in best AI image generator for consistent characters.

A prompt to start from

Write it as a photographer's note: framing, then lens, then light, then a two-word mood. Batch four takes at the same settings and pick, rather than re-rolling one at a time.

Head-and-shoulders portrait of a woman in her thirties with short dark curls and a small scar through one eyebrow, square crop, head slightly high in frame, 85mm lens at f/2, soft window light from camera left with a gentle fill, plain warm gray background, direct unsmiling gaze, calm and unbothered.

For an illustrated avatar on Anima, the same subject compresses into tags: portrait, woman, short dark curls, scar through eyebrow, three-quarter view, clean line art, soft cel shading, muted palette, plain background, calm expression. Both are 10 credits on Krea2 and Anima. More formulas for portraits live in character portrait prompt formulas.

On Anima, keep a short negative prompt and reuse it across the set — extra fingers, deformed hands, text, watermark, logo, cluttered background — rather than growing a new one per render.

What you may not render

One rule, stated plainly, because avatars are exactly where people push on it: no real identifiable people without consent. No likenesses of celebrities, no impersonation, no photo of somebody you know turned into a profile picture for an account that is not theirs. This is enforced at the account level, and it is in the acceptable use policy in writing.

Everything else in the policy still applies: fiction only, every character 18 or older, nothing illegal. Invented people are unrestricted in this sense — build a whole cast, give each one a face, and keep them consistent.

Worth knowing how avatars talk here, too: there is no dedicated talking-head tool. An image can be animated into a short clip through image-to-video, and a Character can carry a voice. For a character who speaks on screen, write the line into a scene in Cinema on an engine that renders sound: the character says it with the picture, in their own voice where the engine supports it. Plan the deliverable around that rather than around a demo you saw somewhere else.

FAQ

What size are AI profile pictures on eroq?

Square renders come out at 1024 × 1024 on eroq's own image engines, which is comfortably above what any platform shows a profile picture at. There is no image upscaler, so pick 1:1 at generation time instead of cropping a larger portrait afterwards.

How do I keep the same avatar face across several pictures?

Attach the same reference photo to every render, or better, save the person as a Character with reference photos and @mention them in each prompt. The Character route also carries the face into clips through image-to-video, so one definition covers a whole set.

Can I make an avatar of myself?

Yes, with your own photos as references, because consent is yours to give. What is prohibited is generating real identifiable people who have not consented, which rules out celebrities and anyone else's likeness.

Can the avatar talk?

Not as a talking-head tool. Write the character's lines into a Cinema scene rendered on an engine with sound and they are spoken with the picture, in the character's own voice where the engine supports it. You can also animate a portrait into a short clip with image-to-video and generate narration separately in the voice studio.

Set the format to square, attach one clean reference, batch four — open the Image studio.