Character drift remains a common frustration in generative image production. You generate an initial portrait, but the moment you change the camera angle, lighting, or setting, the subject morphs into an entirely different person.
The NanoBananas editor resolves this by integrating multi-image reference slots directly into the studio panel. By uploading up to 4 reference photos, creators can anchor facial bone structure, signature outfits, and product geometry across multi-scene storyboards. This guide breaks down the four-slot staging framework.
Why AI Characters Drift: The Latent Space Shift
Standard text-to-image diffusion starts from random Gaussian noise. When you prompt "Sarah walking on a beach," the model picks random facial coordinates that loosely match the text tokens.
Nano Banana 2.1 changes this by injecting spatial image tokens directly into the cross-attention layers. Rather than generating facial proportions from text prompts alone, the model anchors eye spacing, cheekbone structure, and nose bridges to your reference image.
Allocating Your 4 Reference Image Slots
The NanoBananas upload panel accepts up to 4 images per generation task. For maximum visual continuity, assign each slot a distinct role:
Slot 1: Frontal Facial Anchor
A clean frontal portrait with neutral lighting and expression. This locks eye color, nose shape, and jawline proportions.
Slot 2: Profile / 3D Turn Reference
A 45-degree angle photo showing the cheekbone slope and ear position. This provides depth information when your subject turns in dynamic scenes.
Slot 3: Wardrobe or Signature Prop
A clean shot of the jacket, uniform, or accessory. Isolating clothing here prevents garment textures from muddying facial features.
Slot 4: Environment or Secondary Object
An optional reference image for background architectural style, e-commerce product placement, or specific ambient color palettes.
Structuring Multi-Turn Consistency Prompts
When prompting new scenes with reference images uploaded, follow a 4-part structure:
1. Shot Type & Setting: "Medium shot of [Subject] standing inside a subway car at rush hour..."
2. Explicit Reference Tie: "Retain exact facial features, eye shape, and cheekbone structure from reference 1..."
3. Contextual Changes: "Subject is looking down at a phone with tired expression, wearing the dark wool coat from reference 2..."
4. Camera & Lighting: "Fluorescent overhead subway lights, reflections in glass windows, 35mm lens, f/2.8..."
Preserving Product Geometry and Logos for E-Commerce
Commercial staging requires keeping hardware details intact while changing environments:
- Isolate the Product First: Clean product photos with plain white or transparent backgrounds yield much cleaner reference transfers than lifestyle photos cluttered with props.
- Name Key Elements: Describe the product's distinctive parts in text ("brass zipper", "matte black ceramic finish", "stitched leather handle").
- Generate at 2K or 4K: High resolutions preserve metallic logos and fine seam stitching without artifacting.
Production Prompt Templates
Template 1: Character in a New Location
Locks facial bone structure and hair color while moving the subject into a distinct setting.
Cinematic medium close-up of [Character Name] walking through a bustling outdoor spice market in Marrakech. Maintain exact facial bone structure, amber eye color, and wavy shoulder-length auburn hair from reference image 1. Expression: focused curiosity, looking toward a vendor stall on the right. Clothing: textured charcoal linen duster coat with a rolled collar. Lighting: late afternoon dappled sunlight filtering through overhead canvas awnings, warm golden rim light. Shot on 35mm lens, f/2.8, natural film depth of field.
Template 2: Action Pose with Wardrobe Continuity
Preserves face identity and specific clothing items during dynamic physical movement.
Action shot of [Character Name] vaulting over a concrete barrier on an urban rooftop at night. Retain identical facial features, jawline, and short tapered black fade haircut from reference image 1. Clothing: navy blue technical windbreaker and matte black combat boots shown in reference image 2. Dynamic forward-leaning motion, athletic posture, motion blur on background neon billboards. Rain-slicked rooftop ground with neon reflections, atmospheric rain mist in air.
Template 3: E-Commerce Product Angle Variation
Transfers an exact physical product from a studio shot into lifestyle staging.
Commercial lifestyle photograph of the wireless ceramic desktop speaker from reference image 1. The speaker sits on an oiled walnut Scandinavian desk next to an open hardcover design notebook and a brass desk lamp. Maintain exact speaker proportions: rounded cylindrical shape, textured cream ceramic surface, and brushed copper control dial. Natural morning daylight pouring through an adjacent loft window, casting soft directional shadows. High dynamic range, commercial catalog composition, 50mm macro lens.
Template 4: Multi-Turn Storyboard Sequence
Maintains consistent character features across sequential narrative storyboards.
Panel 2 of 4 narrative sequence: [Character Name] seated inside a vintage train compartment looking out the rain-streaked window. Maintain facial identity, thin metal-frame wire glasses, and navy pea coat from reference 1. Introspective expression, chin resting lightly on knuckles. Reflections of passing countryside trees visible in the wet window pane. Cool blue overcast daylight balanced by a warm overhead brass reading lamp.
Common Mistakes That Break Continuity
If reference 1 shows a big smile and reference 2 shows a serious frown, the model synthesizes an unnatural hybrid facial tension. Keep baseline references relaxed and neutral.
Altering the camera lens, lighting mood, wardrobe, and action pose in a single generation step increases drift risk. Change environment and action first, then refine wardrobe details.
Frequently Asked Questions
Put Consistency Workflows to the Test
Upload up to 4 reference photos and run multi-scene consistency workflows directly in NanoBananas Studio.
Open AI Studio