Nano Banana 2 Character Consistency with Reference Images (No LoRA)

Every AI image tool promises “consistent characters.” Most of them mean: sometimes the same vibe. I needed something stricter — the same face across outfit changes, new locations, and multi-shot lookbooks — without training a LoRA.

I spent a week stress-testing Nano Banana 2 character consistency with reference images on NanoBananaTool: portrait sheets, outfit swaps, multi-pose sets, and “same person, new scene” prompts. This is the workflow that held identity best, plus the failure modes that still break it.

TL;DR workflow

Step	Do this	Don’t do this
1	Upload 1 clear face reference (or a small sheet)	Describe hair/eyes differently every prompt
2	Start prompts with identity lock language	Bury identity at the end of a long scene
3	Change only wardrobe / location / camera	Redescribe the whole person each time
4	Regenerate with the same refs if drift appears	Add more random style adjectives

Keep the same character in your browser

Upload a reference, lock identity in the prompt, then generate pose and outfit variants on NanoBananaTool.

Try Nano Banana 2 on NanoBananaTool

Why text-only prompts fail

Text is great for mood. It is bad at biometric continuity.

If prompt A says “woman with auburn hair and green eyes” and prompt B says “young female redhead, emerald gaze,” the model treats them as cousins, not clones. Nano Banana 2 gets much more reliable when a reference image carries identity and the prompt only directs the new scene.

That is the no-LoRA point: you are not training a character model. You are re-anchoring the same visual identity every generation.

Build a reference sheet (3–6 images)

Best results I got used a small set, not a dump of 20 similar selfies.

Clean face reference portrait for identity lock

Minimum viable sheet

Front face, neutral light
3/4 view
Full body or waist-up in neutral clothes

Stronger sheet (add if needed)

Profile
Expression variation (smile / serious)
One “hero” look you want preserved (signature jacket, glasses, hair length)

Reference quality rules

Sharp eyes and mouth; avoid heavy filters
One identity only per sheet
Similar lighting across refs beats “artistic chaos”
Crop tight enough that the face is readable

Prompt formula: lock identity, change only the scene

Use this skeleton every time:

Same person as the reference image(s). Preserve exact face, age, skin tone, hairstyle, and body proportions.
[NEW SCENE ONLY]: {pose}, {wardrobe}, {location}, {camera}, {lighting}.
Do not change identity. High detail, natural skin.

Example A — new location

Same person as the reference image. Preserve exact face and hairstyle.
Walking through a rainy Tokyo street at night, cinematic wet reflections, medium shot, 50mm, natural skin texture.
Do not change identity.

Same face, new location — rainy Tokyo night example

Example B — new outfit

Same person as reference image 1. Preserve face and body identity.
Wear the outfit from image 2 with realistic fabric drape and correct scale.
Studio softbox lighting, lookbook pose, ecommerce fashion photo.
Do not alter facial features.

Same face, new outfit — studio lookbook example

Example C — multi-pose pack

Same character as the reference sheet. Preserve identity across the frame.
Standing pose, three-quarter angle, hands relaxed, clean studio background, full body.
Consistent face and hair with references. Fashion lookbook.

Outfit, pose, and multi-shot recipes

1) Lookbook from one face

Upload face reference.
Generate 4 prompts that only change pose words: front / three-quarter / walking / seated.
Keep wardrobe text identical for the first pass.
Then swap wardrobe in a second pass.

2) Virtual try-on without losing the face

image 1 = person
image 2 = garment
Prompt must say preserve face and body identity from image 1 before describing the clothes.

3) Character across story beats

For storyboards or ad sequences, reuse the same reference set for every shot. Change only:

camera distance
action verb
environment

If you change hair length “for drama,” you are starting a new character.

Common drift fixes

Symptom	Fix
Face slowly morphs over a series	Re-attach original refs; shorten scene text
Hair length / color shifts	Name the hair once, identical wording every prompt
Age jumps	Add “same age as reference”
Body type changes in full-body shots	Include a waist-up or full-body reference
Style overpowers identity	Remove artist-name spam; keep “natural skin”
Group scenes merge faces	Limit to 1–2 characters; stronger single-face refs

My take: most “Banana is inconsistent” complaints are actually prompt drift — people rewrite the person every time.

When to use Nano Banana Pro instead

Use Nano Banana 2 for fast lookbooks, ecommerce try-on, and social variants.

Reach for Nano Banana Pro (or your highest-fidelity tier on NanoBananaTool) when:

you need tighter multi-reference locking
labels/logos in-frame must stay readable beside the character
you are producing hero campaign stills where small facial drift is unacceptable

If you are unsure, generate the series on 2 first. Promote only the failing shots to Pro.

Try it in NanoBananaTool

Create a 3-image reference sheet.
Run Example A and Example B above on Generate.
Save the best identity lock as your default template on Banana Prompts.
Optional: turn a winning still into motion with Seedance 2.0.

Character consistency without LoRA training

Reference in → scene out. Same face across outfits and locations.

Start on NanoBananaTool

Final verdict

Nano Banana 2 character consistency is real when you treat references as the source of truth and prompts as stage directions. No LoRA, no fine-tune — just disciplined inputs.

If you only remember one line, remember this: describe what changes, not who they are.

FAQ

How do I keep the same face with Nano Banana 2?

Upload a clear face reference (or a small sheet) and start every prompt with identity-lock language. Change only pose, wardrobe, and location.

Do I need LoRA for character consistency?

Not for most lookbooks and ecommerce try-on. Reference images on Nano Banana 2 are enough when you stop rewriting facial descriptions each time.

How many reference images work best?

Three strong views beat ten weak selfies. Add more only for missing angles.

Why does my character still drift?

Usually mixed wording, weak refs, or oversized scene prompts. Re-anchor with the original sheet and shorten the text.

Can I do this on NanoBananaTool?

Yes — use Generate with references, and keep reusable prompt text in Banana Prompts.