A portrait sheet is only useful if every panel reads as the same person. Without a locked identity anchor, diffusion models drift face-to-face inside a single grid render — different jawline in panel 2, different eye spacing in panel 5. These 12 prompts pair a fixed identity description with a neutral plate so the only thing that changes is the variable you actually want: expression, outfit, age, or lighting. For the full seed-and-reference workflow, see AI consistent character images.
TL;DR
- Write one identity anchor — five fixed descriptors (hair, eyes, key feature, costume detail, face structure) — and paste it verbatim into every prompt. That repeated string is the model’s fingerprint.
- Pin identity with a reference image, not just text. As of June 2026, use Midjourney V7 Omni Reference (
--oref), a Nano Banana Pro reference upload, or Flux Kontext edit chaining. The old Midjourney--crefno longer works on V7. - Vary exactly one axis per sheet (expression or outfit or age or lighting). Keep aspect ratio and background identical so panels stay comparable.
- Swap the
[identity anchor]placeholder below for your own descriptors. Brackets are placeholders, not literal syntax.
Best for
- Character design pipelines for games and TTRPGs
- Comic and webtoon series with a recurring lead
- Indie game NPC sheets shared with writers and animators
- Brand mascots that need expression variants for marketing
Which reference tool to use (June 2026)
Text alone gets you maybe 60% consistency; a reference image pushes a good workflow to roughly 85%. Pick the tool by how much identity control you need.
| Tool | Reference method | Identity strength | Best for |
|---|---|---|---|
| Midjourney V7 | Omni Reference --oref + --ow (0–1000, default 100) | Strong; raise --ow to ~200–400 to lock a face | Stylized concept art, hero shots |
| Nano Banana Pro (Gemini 3 Pro Image) | Reference upload, holds up to ~5 characters | Strong; top-rated on image arenas mid-2026 | Multi-character sheets, clean line art |
| Flux Kontext / FLUX.2 | Conversational edit on a base image | Strong for edits, weaker from scratch | Outfit/expression swaps off one locked base |
| SD / Flux + LoRA | Custom LoRA trained on 15–30 images | Highest at scale once trained | Long-running series, many sheets |
A practical note on Midjourney: --cref (and its --cw weight) only applies to legacy V6. On V7, use --oref with --ow — keep --ow in the 200–400 band for faces, and lower it toward 50 if you want the model to restyle. --sref still works for style lock alongside --oref. (See Midjourney’s Omni Reference docs.)
1. Four-expression base sheet
character portrait sheet, same character: [identity anchor — hair, eyes, key feature, costume], 4 expressions in a row: neutral, smile, angry, sad, consistent face structure, neutral grey background, even soft lighting, concept art style, 16:92. Eight-expression extended sheet
character expression sheet, same character: [identity anchor], 8 expressions in a 4x2 grid: neutral, smile, laugh, smirk, angry, sad, surprised, scared, identical face structure across panels, neutral grey background, flat lighting, concept art style3. Outfit variation row
character outfit variations, same character: [identity anchor], 4 outfits: casual, formal, combat, festival, identical face and hair, neutral background, full body, three-quarter view, concept art style4. Three-quarter turnaround
character turnaround sheet, same character: [identity anchor], three views in one row: front, three-quarter, side profile, identical proportions, neutral grey background, even lighting, model sheet style with clean line art5. Age progression sheet
character age progression, same character: [identity anchor — keep hair color and key feature], 4 ages: child (8), teen (15), young adult (25), elder (55), consistent facial bone structure across ages, neutral background, concept art style6. Lighting-locked variants
same character: [identity anchor], 4 lighting setups in a grid: soft daylight, dramatic rim light, golden hour, moonlight, identical pose and outfit, only lighting changes, painterly concept art style7. Emotion close-up tiles
character emotion close-up sheet, same character: [identity anchor], tight headshots only, 6 emotions: determined, exhausted, joyful, suspicious, heartbroken, defiant, identical face structure, soft studio light, neutral background8. Class or role variants (RPG)
same character: [identity anchor — face and hair locked], 4 class variants: warrior, mage, rogue, cleric, identical face across variants, only armor and props change, full body, three-quarter view, fantasy concept art9. Idle pose panel set
character pose sheet, same character: [identity anchor], 4 idle poses: standing relaxed, arms crossed, leaning, mid-walk, identical outfit and face, neutral grey background, clean line art with flat color, model sheet style10. NPC dialogue portrait set
NPC dialogue portraits, same character: [identity anchor], 4 dialogue states: greeting, explaining, surprised, farewell, shoulders-up framing, consistent face and outfit, slight 3/4 angle, JRPG visual novel style, flat lighting11. Hairstyle exploration row
hairstyle exploration sheet, same character: [identity anchor — face locked, hair variable], 4 hairstyles: short crop, shoulder length, long braid, updo, identical face, neutral background, soft front lighting, concept art style12. Cover-vs-sheet hero shot
hero cover shot of same character: [identity anchor], dynamic three-quarter pose, dramatic lighting, painterly background hinting at the world, identical face to the rest of the sheet, 9:16, key-art styleHow to refine
Pin identity with a reference image: --oref on Midjourney V7, an uploaded reference on Nano Banana Pro, or IP-Adapter on SD/Flux. Lock the seed across grids so random variation comes from your prompt, not the noise. Write the identity anchor as the same five descriptors verbatim in every prompt — hair, eyes, key feature, costume detail, face structure — so the model has a fingerprint to hit. Render expression and outfit sheets at the same aspect ratio so panels stay directly comparable.
If a single grid still drifts, split it: generate each panel as its own image off the same reference and seed, then composite the sheet in an editor. One face per render is far more reliable than four faces in one tile.
Common mistakes
- No identity anchor — the face drifts panel to panel even with the same seed
- Background varies and shifts skin tone or palette, faking inconsistency
- Pose varies wildly (frontal plus profile plus 3/4) without a base front pose to anchor on
- Outfit changes accidentally change hairstyle — fix hair as part of the identity, never the outfit
- Mixing lighting between panels of the same sheet, which reads as a different character
- Reaching for
--crefon Midjourney V7, where it is ignored — use--orefinstead
FAQ
Does Midjourney --cref still work for character sheets?
Not on V7. As of June 2026, --cref and its --cw weight only apply to legacy V6. On V7 use Omni Reference: pass your reference with --oref and tune --ow (0–1000, default 100; 200–400 for a firm face lock).
Which tool gives the most consistent faces right now? For text-plus-reference, Nano Banana Pro (Gemini 3 Pro Image) and Midjourney V7 Omni Reference both hold a face well in mid-2026. For the highest consistency across a long series, train a LoRA on 15–30 images of the character on SD or Flux — slower to set up, most reliable at scale.
Why do the prompts use brackets like [identity anchor]?
Those are placeholders. Replace the whole bracketed phrase with your own five descriptors (for example, silver bob, green eyes, scar over left brow, navy field coat, oval face). Do not keep the brackets in the final prompt.
Should I put everything in one grid or render panels separately? Start with one grid for speed. If faces drift, render each panel individually off the same reference image and seed, then assemble the sheet manually. Single-face renders are more consistent than multi-panel tiles.
Can I lock identity and style at the same time?
Yes. On Midjourney V7 you can combine --oref (identity) with --sref (style) in one prompt; the engine separates the two. On SD/Flux, pair IP-Adapter for the face with a style LoRA or style reference.
Related
- Game character class prompts — class-driven character prompts
- Game character design prompts — broader design library
- Consistent image style prompt — style lock across grids
- AI consistent character images — full consistency workflow
Tags: #Game character #Consistency