This is the only image you generate completely from scratch. Keep it simple: front-facing, plain background, even light. Every detail you name here — hair, eyes, skin, expression — is what the AI has to remember and repeat in every image after this one.
A [gender, age range] with [hair color and style], [skin tone], [key facial features — e.g. freckles, scar, dimples], [eye color], [expression]. Simple front-facing portrait, chest-up crop, plain neutral light grey background, even soft studio light from the front, no harsh shadows. Natural skin texture with visible pores, individual eyebrow hairs, natural asymmetry in the features. Shot on an 85mm portrait lens, shallow depth of field, true-to-life skin tones. No plastic skin, no beauty filter, no AI smoothing, no unnatural symmetry.
Everything in [brackets] is what you swap for your own brand’s character — the rest of the structure is what keeps the result consistent, so leave it as is.
Feed Step 1 back in as a reference image and ask for a full character sheet in one generation — front, three-quarter and profile angles side by side. If the face holds up across every panel, you have a character you can actually reuse, not just one lucky image.
The same [gender] from the reference image, preserving their exact facial identity — [restate 2–3 defining features, e.g. hair, eye color, freckles]. Generate a character reference sheet: six portraits of the same person arranged in a grid, from different angles — front-facing, three-quarter left, full profile left, three-quarter right, full profile right, and a candid downward angle — all on the same plain neutral background with consistent studio lighting. Natural skin texture, consistent facial proportions and features, same [defining features] in every panel. No plastic skin, no identity drift, no altered facial structure between panels.
One generation, six proofs. If any panel drifts — a different nose, lighter eyes — that's the version of the prompt to fix before you move on.
This step runs on two references at once, not one. The first is your identity — and this time, use the full multi-angle sheet from Step 02, not a single photo. The second is the outfit — and this one can be anything: a product photo, a Pinterest save, a fit check from a total stranger. Feeding in every identity angle lets the model match each outfit angle to the closest matching pose instead of guessing.
Reference B doesn’t need to match your character’s identity, pose or lighting — the model only reads it for the clothes. Upload both images before you generate.
The [gender] shown across reference image A, preserving their exact facial identity and features in every angle — [restate 2–3 defining features]. Dress them in the exact outfit from reference image B — [describe the garments, fit and colors you see: e.g. light blue oversized button-up, striped tie loosened at the collar, brown leather belt, wide-leg denim] — matching the fit, styling and color of reference B exactly. Generate a reference sheet: five portraits of this person in this outfit from different angles — front-facing, three-quarter, full profile, three-quarter opposite side, and back view — matching each pose to the closest corresponding angle in reference A, on a plain neutral background with consistent studio lighting. Natural fabric texture, drape and stitching detail, realistic fabric weight. Same face, same hair, same [defining features] as reference A in every front and side panel. No plastic skin, no identity drift, no altered facial features, no invented garment details not shown in reference B.
If the face drifts once clothes are added, tighten the identity language — restate the defining features again inside this same prompt, don’t assume the model remembers them.
The final shot runs on three references together: who they are, what they're wearing, and where they are. The first two you already have. The third — the environment — can come from anywhere: a moodboard photo, a campaign still, a shot you saved months ago for a completely different reason.
Reference C sets the mood, not the person — the model reads it for the room, the light and the composition only. Upload all three before you generate.
The [gender] shown in reference image B, preserving their exact identity and outfit exactly as shown. Place them into the setting and composition from reference image C — [describe the environment: e.g. a sunlit creative studio desk cluttered with sketchbooks, watercolors, a camera and open art books, beside a tall window with sheer curtains, herringbone wood floor]. [Describe the pose — e.g. seated in a cane chair, one leg propped up on the desk, laptop resting on their lap, calm and focused expression toward camera]. Camera angled slightly from above, natural window light from the left with soft warm falloff, true-to-life color grade with a light film grain, shallow depth of field softening the desk clutter behind them. Same face, same outfit, same identity as references A and B. Realistic fabric folds, natural skin texture. No plastic skin, no identity drift, no invented facial or outfit details not shown in the references.
Swap out reference C and you get a new piece of content in the same identity and outfit — that reuse is the entire point of locking the character in Steps 01–03.