A character portrait generator that remembers the face

Give the model a few photos of your character. Then put them in any outfit, any world, any light — and still recognise them in the result.

0/4 reference images · photos are resized on your device before upload

Public in Explore

Try one of these

Character work breaks generic image generators. Ask for the same hero twice and you get two attractive strangers who share a haircut. Anything with continuity — a comic, a game cast, a recurring ad character — needs the face to survive repetition.

This page is about building a character that holds up: what to put in the reference set, how to describe a scene without describing the face again, and how to keep a cast consistent across dozens of images.

Build the reference set like a casting folder

Three to six photographs of one person, from different angles, is the sweet spot. A straight-on portrait establishes proportions; a three-quarter view captures the bridge of the nose and the jawline; a profile pins the silhouette. Add one full-body shot if the character will appear in wide frames, because the model also learns height and build from it.

Consistency across the set matters more than quality. If half your references are shot in warm indoor light and half in cold daylight, the model has to guess which skin tone is true, and the output drifts toward a flat midpoint. Photographs taken in similar light produce the most stable character.

For drawn characters, the same logic applies to line art: supply several views of the same design rather than several styles of the same view.

  • Front, three-quarter and profile — in that order of importance
  • Similar lighting across all references
  • One character per set; never mix a cast in a single request
  • Add a full-body shot only if you need the body to stay consistent too

Describe the scene, not the person

The most common mistake in character work is re-describing the character in every prompt. "A woman with silver hair and green eyes in a market" tells the model to invent those features again, competing with the references. The reference already carries the face; the prompt should carry the situation.

Write "in a crowded spice market at noon, dust in the air, handheld documentary framing" and let the references do the identity work. When you do mention a feature, mention it only to change it — "hair now braided", "a fresh scar across the left eyebrow" — so the model treats it as an edit rather than a description.

This is also why the Character mode here defaults to a square frame: character sheets, avatars and dialogue portraits are usually square, and a square composition keeps the head large enough for facial detail to survive.

Keeping a cast consistent across dozens of images

Once a character works, freeze the inputs. Save the exact reference set, the style chips and the tier, and change only the scene description between renders. Every variable you leave alone is a variable that cannot drift.

For a cast of several characters, keep a separate reference folder per character and render them individually before attempting any group shot. Group scenes are the hardest case, because the model must hold two identities in one frame and will often let the less-referenced one blur.

Name your files by character and angle. Six months into a project, the difference between "IMG_2043.jpg" and "mara-profile-01.jpg" is the difference between a rebuildable character and a lost one.

Generic text to image vs a reference-driven character workflow

Prompt onlyCharacter mode
Same face across imagesNo — a new person each runYes, held by the reference set
Cost to fix a driftRewrite and re-roll, repeatedlyAdd a reference from the missing angle
Works for a castNot practicallyOne folder per character
Best frameWhatever the prompt suggestsSquare, head large in frame

3 example prompts are preloaded above — tap one to try it.

Questions, answered

About fidelity, credits and output sizes on this tool.

Ready to make something that looks like the person you had in mind?

Start generating