Creating Fantasy Portraits With AI: A How-To Guide

Fantasy portraits fail more often than other restyles, and the reason is not that the style is harder to render. It is that the instruction is longer, and every extra demand competes with your face.

The competition problem

A watercolour instruction says one thing. A fantasy portrait instruction says four: wear this, stand here, be lit like this, feel like that.

The model weighs all of them, and your likeness has no clause arguing for it unless you write one. So it loses. That is why people get back a beautifully rendered stranger in excellent armour.

Two fixes, and they are the whole article:

1. Write a preservation clause. Explicitly: "keeping the face, hair and skin tone unchanged". It competes.
2. Keep the face large in the frame. A head-and-shoulders crop keeps far more structure than a full-body heroic pose. If you want the full pose, accept that the face becomes a stylised suggestion.

Choosing a source photo

The requirements are the same as for any heavy restyle, with one addition.

• Front-facing or three-quarter. Dramatic angles are outside what the style handles well.
• Even, directional light. One side brighter than the other gives the model form to work with; flat light produces flat results.
• Hair fully visible. Along with face shape, it is doing most of the identification.
• Neutral expression or a slight one. Extreme expressions stylise badly.
• And: simple clothing, or shoulders only. Whatever you are wearing has to be replaced. A plain crew neck is easier to overwrite than a patterned jacket, which tends to leave ghosts of its pattern in the new costume.

Describing the picture, not the character

This is where most fantasy prompts go wrong. They read like character sheets.

The model has no access to backstory. It responds to visual description: materials, colours, light, texture, shape.

Not this:

A brave ranger from the northern kingdoms, weathered by years of hardship, noble and determined.

This:

Weathered green wool cloak with a bronze clasp, leather shoulder guard, damp hair, lit from below by a low fire, cold blue ambient light behind, shallow depth of field, keeping the face and hair unchanged.

Every clause in the second version is something you could point at. That is the test.

Structure the instruction in four parts

Writing it in a fixed order makes it easier to change one thing later:

1. Costume — materials first, then colours, then details. "Dark leather, brass buckles, deep red under-tunic."
2. Setting — a few words, suggested not described. "Stone hall, distant torchlight."
3. Light — direction, colour, quality. This does more for the mood than anything else. "Warm low key light from the left, cool fill behind."
4. Preservation — "keeping the face, hair and skin tone unchanged."

Keeping costume and light as separate clauses matters, because "epic fantasy lighting" bundled into the costume description is a judgement rather than a visual instruction, and judgements are what let the model drift.

What always breaks

Plan around these rather than fighting them:

• Hands. Especially holding a weapon, a staff, a book. The most reliable failure in the whole field. Compose them out of frame.
• Armour and jewellery. Rigid repeating structure; the model regularises it and the regularity looks wrong. Keep it simple and expect to check.
• Text. Banners, runes, book spines. Always nonsense. Ask for no text.
• Symmetry. Paired items — pauldrons, earrings — frequently do not match.
• Fine chain and lacing. Dissolves into texture.

A portrait crop from the chest up avoids most of this list, which is another reason it is the safer framing.

Building a consistent character

If you want several images of the same character:

• One anchor photo, used every time
• One costume description, reused verbatim
• Change one clause per image — the setting, or the light, not both

Generation varies between runs even with identical inputs, so consistency is something you impose upstream. The full method is in creating consistent avatars.

Judging it

Two checks:

Does it still read as you? Show someone who knows you. Hair, face shape and skin tone are carrying this, as they do in any stylised avatar.

Does it hold up at 100 percent? Hands, paired items, any lacing or chain, and the boundary between hair and background.

Three generations before judging the source photo, not one — heavy styles vary enormously between runs.

Restyling runs in Kitana's browser studio, and the thirteen photo tools, including the dedicated avatar tool, are in the apps.