AI Consistent Character Prompts
Identity is built from one canonical reference, not five "ideal" portraits. These recipes assume you start with a single canonical character image (Nano Banana 2 is the canonical face-locker) and then spin out turnaround views, expression sheets, outfit variants, and scene swaps — all anchored to that one image. Flux Kontext handles outfit and scene swaps with face preservation; Runway Gen-4 Image holds character-aware compositions; Midjourney develops the look. The downstream payoff is a video pipeline (Vidu, Kling O3) that inherits the same locked identity.
When to use this prompt
- Building a 12-week AI influencer pipeline around one recurring face ("Mia").
- Producing recurring-protagonist artwork for a Substack serial or web novel.
- Designing a brand mascot bible — multi-pose, multi-expression, multi-wardrobe.
- Generating a character bible for an AI short film before the video phase.
- Refreshing a legacy character across new outfits, ages, and scenes without re-casting.
Required inputs
- A single canonical reference image of the character — front-facing three-quarter is the most reusable angle.
- A short identity descriptor (age range, build, distinguishing features) — used sparingly to support the reference.
- A wardrobe / scene reference image when swapping outfit or environment.
- Aspect ratio target for the deliverable (1:1 for portraits, 3:4 for editorial, 16:9 for cinematic).
- Optional: a style anchor image if you want a specific rendering aesthetic (photoreal vs editorial vs anime).
Prompt recipes
Master portrait (front three-quarter)
The canonical anchor. Generate this once on Nano Banana 2 and reuse as the reference for every other recipe on this page. Do not regenerate — refine.
Front three-quarter portrait of [identity descriptor: e.g., woman late 20s, mid-length brown hair, soft features], soft front-key plus rim light, neutral studio backdrop, sharp eyes in focus, natural skin texture, photoreal rendering, 3:4 framing.
This is the single source of truth. All other recipes reference this image.
Three-quarter turnaround
Builds the back-three-quarter angle from the master. Useful for character bible turnarounds and for video shots that pan around the subject.
Same character as referenced canonical portrait, three-quarter back view, identical wardrobe, identical hair, identical lighting, neutral studio backdrop, locked face structure, locked features.
Profile silhouette
The pure-profile angle. Best when you need a silhouette for graphic design or a side-on close-up.
Same character as referenced canonical portrait, full profile, identical wardrobe and hair, soft side rim light, neutral backdrop, natural skin texture, locked nose and chin profile.
Outfit swap (same identity)
Wardrobe variation while preserving identity. Flux Kontext is edit-aware and handles outfit swaps without face drift.
Same character as referenced canonical portrait, now wearing [referenced wardrobe image: e.g., a navy blazer over white shirt], identical face, identical hair, same studio lighting, no change to body proportions.
Variations
- Substitute the referenced wardrobe with seasonal variants (summer, winter, formal, casual).
Expression sheet
Five-expression set in one generation. Use the result as a reference panel for downstream video reaction shots.
Same character as referenced canonical portrait, expression sheet — five identical compositions side by side: (1) neutral, (2) warm smile, (3) surprised, (4) determined, (5) thoughtful. Locked face, locked hair, identical wardrobe, identical lighting.
Model output is one wide image with five panels — crop downstream as needed.
Scene swap (locked identity)
Drops the locked character into a new environment. Flux Kontext preserves identity through the scene swap better than re-prompting from scratch.
Same character as referenced canonical portrait, now placed in [referenced scene image: e.g., a sunlit Brooklyn rooftop at golden hour], same face, same wardrobe, same hair, lighting matched to scene, natural integration with environment.
Age progression
Single-character age progression for narrative arcs. Stay close to the canonical features — large jumps lose identity.
Same character as referenced canonical portrait, aged to [target age, e.g., late 40s], identical bone structure, identical eye color, hair shows subtle gray, fine lines around eyes, same wardrobe style adapted, neutral studio backdrop.
20-year jumps work; 50-year jumps drift.
Lighting variations
Cinematic lighting variant for hero placements. Useful as a video shot anchor when the character moves into a dramatic scene.
Same character as referenced canonical portrait, dramatic backlit lighting from camera right, deep shadow on left side of face, atmosphere haze, locked features, locked wardrobe, editorial mood.
Variations
- Swap backlit for golden-hour side light.
- Swap dramatic for soft beauty light.
Cinematic hero frame
Cinema-style hero frame — useful as the opening or closing image in a multi-shot sequence.
Same character as referenced canonical portrait, full-body cinematic hero frame, 16:9 framing, atmospheric environment matched to mood, soft volumetric light, locked face and identity, sharp focus on character, depth of field falls off in background.
Editorial portrait development
Look development pass on Midjourney for high-fidelity finished hero portraits. Use after the canonical is locked.
Character look development based on referenced canonical portrait, editorial fashion photography aesthetic, soft window light, neutral wardrobe, magazine-quality skin retouch level, locked identity, 3:4 portrait framing.
Martini canvas workflow
Generate the canonical character portrait once on Nano Banana 2 and label it "canonical-ref" on the canvas. Refine it (small re-runs, never wholesale regenerations) until it is exactly the face you want. This single image is the source of truth.
Pin the canonical-ref as an upstream image-anchor node. Every other character node — turnaround, outfit, expression, scene — wires this anchor in as its reference. Do not feed multiple "ideal" portraits; the model averages them and the face drifts.
Use Flux Kontext for edit-aware operations (outfit swap, scene swap, expression change) that need to preserve identity through the change. Use Nano Banana 2 for face-locked re-poses (turnaround, profile). Use Midjourney for editorial look development on the master.
Curate a character bible on the canvas — master portrait + 3–5 turnaround views + 5-expression sheet + 2–3 outfit variants. This is your reusable bible. Future shots reference this bible, not just the original portrait.
Chain into video by wiring the canonical-ref (and the strongest turnaround / expression frame for each shot) into Vidu or Kling O3 video nodes. The downstream video phase inherits the locked identity rather than re-prompting from scratch — see /prompts/video/multi-shot-video-prompts for the sequence vocabulary.
Variations
Photoreal
Default rendering style — natural skin, natural light, photographic detail. Best for influencer / brand pipelines.
Editorial
Magazine-quality retouching, dramatic light, fashion-photography aesthetic. For premium brand work.
Anime
Anime / manga rendering. Note: do not stylize against a known IP (Ghibli, Pixar) for commercial release.
3D / stylized
Pixar-style or 3D-render aesthetic. Same identity caveats apply.
Young (20s)
Default character age. Most reference models are tuned around this range.
Mature (40s–60s)
Aged variant. Stay within ~20 years of the canonical for clean identity transfer.
Related features
Related how-tos
Related models
Related blog posts
Related docs
Frequently asked questions
- Why does my character's face drift across generations?
- Most often because you fed multiple "ideal" portraits as references — the model averages them and the face shifts. Use one canonical reference image, lock it as a single anchor, and reference it everywhere. Do not regenerate the canonical every session; refine it once and reuse.
- Nano Banana 2 vs Flux Kontext — which do I use when?
- Nano Banana 2 for face-locked re-poses (turnarounds, profiles, lighting variations) where identity is the priority. Flux Kontext for edit-aware operations (outfit swap, scene swap, expression change) where you need to alter context while preserving the face. Use both — Nano Banana for the bible, Flux Kontext for the variations.
- Can I style the character after a known IP (Ghibli, Pixar)?
- Technically the models will attempt it, but for commercial release it is legally risky — those are protected IPs. For commercial work, generate an original aesthetic (your brand style guide) rather than mimicking a studio. For personal / fan work, treat outputs as non-commercial.
- How do I avoid "more handsome" prompt regression?
- Subjective adjective edits ("more handsome", "younger looking", "prettier") cause the model to regress toward an average idealized face and drop your canonical features. Stay literal — describe wardrobe, lighting, pose, expression, scene. Save the subjective adjustments for actual look development on the canonical, not for downstream variations.
- How do I move from image to video with the same character?
- Wire the canonical-ref image (plus the strongest pose / expression for the shot) into a video node — Vidu (1–7 character refs), Kling O3 (character-aware motion), or Kling 3. The video phase inherits the locked identity rather than re-prompting. See /prompts/video/multi-shot-video-prompts for the sequence vocabulary.
- How many bible images do I need before starting the video phase?
- Master + 3 turnaround angles + 5-expression sheet + 2 outfit variants is the practical floor. That gives you the right per-shot anchor for most scenes. For action-heavy projects, add 2–3 dynamic body-pose references.
Try this prompt on Martini
Copy a recipe above, drop it into a node, and run it inside a full canvas workflow.