How to Write Prompts for Nano Banana: 10 Tips for Creators

Nano Banana is Google's codename for Gemini's image generation models — the same family that powers image features in Google AI Studio, the Gemini API, and consumer Gemini apps. The two checkpoints you'll typically see:

- `gemini-2.5-flash-image` — fast, lower-cost, the everyday workhorse
- `gemini-3-pro-image-preview` — slower, higher fidelity, better at fine detail and text rendering

What makes Nano Banana different from Midjourney / DALL·E:
- Native multi-turn editing. Keep the subject the same and change only the lighting / pose / background in the next turn. Other models usually force you to rewrite the prompt from scratch.
- Strong text rendering. Posters, infographics, and signage with readable text are within reach.
- Tight prompt adherence. Specific colors, named camera bodies, and aspect ratios are usually honored.

Where Curify fits:
Curify is a templating layer on top of Nano Banana. Templates pre-bake the prompt structure (composition, style tokens, negative-space rules) so non-experts get consistent results without writing each tip below from scratch. This article is for when you want to write prompts directly. If you'd rather skip ahead, browse a few of the [Nano Banana templates] in the gallery to see what a well-crafted prompt looks like in production.

Lead with the noun, not the modifier. A common rookie mistake is starting with the style: *"Cinematic photo of a..."* — the model treats the subject as a secondary detail. Instead, name the subject specifically, then describe how it's rendered.

Concrete subject signals to include:
- Species or role: "a red panda", "an elderly Tibetan monk"
- Action or pose: "standing on a moss-covered log", "mid-stride at sunrise"
- Setting: "in a Himalayan rhododendron forest"
- Eye contact / framing: "centered, eye contact with camera"

A red panda standing on a moss-covered log in a Himalayan rhododendron forest, photographed in a cinematic documentary style. Soft early-morning fog drifting between the trees. Subject is centered, fully in focus, gentle eye contact with camera.

Photographer vocabulary puts the model into "photo" mode and gives you direct control over depth, perspective, and intimacy. Useful tokens to memorize:

- Focal length: 24mm (wide environmental), 35mm (documentary), 50mm (natural), 85mm (portrait), 135mm (compressed), macro
- Depth: shallow depth of field, deep focus, bokeh background
- Angle: top-down (flat-lay), eye-level, low angle (heroic), Dutch tilt (unease)
- Format: shot on Fujifilm Pro 400H, Kodak Portra 400, 35mm film grain

Close-up portrait of an elderly Tibetan monk in red robes, shot on 85mm lens with shallow depth of field. Soft window light from the left, gentle fall-off on the right side of the face. Background fades to warm bokeh. Photographed on Kodak Portra 400, subtle film grain.

Lighting words are the highest-leverage style modifier in image generation. Two prompts identical except for the lighting line will produce wildly different outputs. Build a vocabulary:

- Time-of-day: golden hour, blue hour, harsh midday, overcast, twilight
- Quality: hard light, soft diffused light, dappled light (through leaves), high-contrast
- Direction: rim light from behind, side light from camera left, top-down (god rays)
- Studio language: softbox, key light, fill light, Rembrandt lighting, butterfly lighting

Studio portrait of a chef arranging fresh ingredients on a marble counter. Key light: one large softbox from camera left, raised to 45°. Fill: subtle white reflector on the right. Background fades to deep charcoal. Crisp catchlight in the eyes.

Layering more than one style muddies the output. "Watercolor + photorealistic + anime" produces an incoherent mess. Pick a single style and lean into its vocabulary:

- Cinematic photography: anamorphic, color graded, filmic contrast
- Watercolor: loose brush strokes, paper texture, soft edges, no outlines
- Vintage editorial: 1970s magazine, muted earth tones, grain
- Isometric illustration: clean vectors, 30° angle, soft shadows
- Anime cel-shaded: bold outlines, flat color regions, highlight on hair
- Ink wash: brushwork, negative space, monochrome plus one accent color

Watercolor illustration of a Parisian café terrace at sunset. Loose brush strokes, visible paper texture, soft warm color palette of amber, dusty rose, and warm grey. No hard outlines. Wet-on-wet bleeds in the sky.

Tell Nano Banana where the subject sits in the frame. Without a composition instruction, you'll get default-centered subjects every time. Useful phrases:

- Rule of thirds, subject on the right third
- Centered with negative space above (good for adding text)
- Leading lines from bottom-left to upper-right
- Subject in the lower-left, environment fills the rest
- Symmetrical composition, subject centered, mirrored elements

Wide landscape photograph of a single red rowboat on a still alpine lake. Rule of thirds: boat positioned on the lower-right third. Snow-capped mountains fill the upper two thirds. Mirror-perfect reflection in the water. Negative space in the upper-left for a potential title overlay.

Nano Banana follows color guidance more precisely than most image models. Three approaches:

1. Name a palette: "muted earth tones", "high-contrast neon", "pastel ice cream", "corporate blue + grey"
2. Provide hex values: "dominant color #c47957, accents #a3b29b and #f4e9d6"
3. Reference an aesthetic: "Wes Anderson palette", "early Pixar 2D color script"

Add a negative for what to suppress: *"no oversaturated blues, no neon greens."*

Isometric illustration of a tiny office floating on a small island. Color palette: dominant terracotta (#c47957), soft sage accent (#a3b29b), cream backgrounds (#f4e9d6). Avoid: oversaturated primaries, neon greens, pure black. 30° isometric angle.

Don't rely on defaults. Spell out the ratio explicitly so the composition is built for the platform you'll publish on:

- 1:1 — Instagram square, LinkedIn feed
- 4:5 — Instagram portrait (the highest-engagement IG ratio)
- 9:16 — TikTok, Reels, Shorts, Stories
- 16:9 — YouTube thumbnail, blog hero
- 3:2 / 2:3 — DSLR landscape / Pinterest portrait
- A4 / 210x297 — printable poster, infographic

When in doubt, also tell the model *what fits where* in the frame — "subject occupies the lower two-thirds, with sky negative space above for caption text."

Vertical 9:16 portrait of a barista pulling an espresso shot, rosetta foam swirl visible in the cup. Phone-style aspect ratio intended for a TikTok cover. Subject occupies the lower two-thirds; upper third is darker, leaving room for a video title.

Nano Banana handles "no X" / "avoid X" naturally. But stuffing generic negatives (*"low quality, bad anatomy, blurred"*) is cargo-cult Midjourney habit and tends to underperform. Target the specific failure modes you keep seeing:

- Hands: *"no extra fingers, no fused fingers, hands fully visible"*
- Text: *"no text or watermarks, no captions overlaid on the image"*
- Style drift: *"no cartoon style, photorealistic only"*
- Distortion: *"no fisheye distortion, no warped faces near the frame edge"*
- Composition: *"no subject cropped at the eyes or chin"*

Photorealistic portrait of a marathon runner mid-stride at sunrise, city skyline blurred behind. Avoid: motion blur on the face, distorted hands, blurred or asymmetric eyes, supplemental text overlays, watermarks, fisheye distortion at the frame edges.

This is the single biggest advantage of Nano Banana over Midjourney. Other models force you to re-prompt from scratch for every variation. Gemini lets you say *"same image, but..."* and keep the subject locked while changing one dimension.

Typical iteration patterns:

- Lock subject, vary lighting: "same character, change the lighting to overcast morning instead of golden hour"
- Lock style, vary scene: "keep the watercolor style and the character, move them to a snow-covered cabin"
- Lock everything, add one element: "same image, add a black cat sitting on the windowsill"
- Generate variants: "three variations of this image, each with a different background color"

Don't try to land everything in turn 1. Sketch broadly, then refine.

Turn 1 — A samurai standing in a bamboo grove, traditional ink-wash style, monochrome with a single red accent.
Turn 2 — Same image, but change the time of day to dusk with strong rim light from behind the subject.
Turn 3 — Same image again, but add a crow perched on a bamboo branch in the upper-right corner. Keep the ink-wash style and dusk lighting.
Turn 4 — Generate three variations of the final image, each with the samurai's stance slightly altered.

Tips 2–8 are exactly what a well-designed template hardcodes. If you're generating one-off art and want full creative control, write the prompt by hand using everything above. But if you're producing a *series* — character cards across MBTI types, recipe cards across cuisines, vocabulary posters across topics — the lens, lighting, palette, composition, and aspect ratio should be fixed across the set. That's what templates do.

Curify's [Nano Banana templates] take a tested prompt skeleton and expose only the content-specific parameters (subject, theme, language pair, etc). The result is consistent across hundreds of generations without you re-typing the lighting line every time.

Rule of thumb: first time = write the prompt yourself. Second time you reach for the same prompt structure = build or find a template.

If you find yourself re-typing prompts like *"watercolor infographic, soft palette, 4:5 aspect ratio, no outlines, {topic_name}"* — that's a template. The {topic_name} is the only parameter that should change between generations. Curify's template gallery has dozens of these for common use cases.

Direct access:
- Google AI Studio (aistudio.google.com) — free interactive playground; best for one-off exploration
- Gemini API — programmatic access, pay-per-image (`gemini-2.5-flash-image` and `gemini-3-pro-image-preview`)
- Vertex AI — enterprise tier with regional deployment and compliance controls

Wrapped access (UX optimized for non-engineers):
- Curify — template-based UX on top of the Gemini image API; you fill in subject parameters, Curify constructs the full prompt
- Consumer Gemini app — image generation as part of the general Gemini chat experience

If you're writing prompts directly (this article's audience), Google AI Studio is the cheapest place to iterate. If you're producing a series of consistent images, Curify is the lower-effort path.

Every Curify Nano Banana template encodes lens, lighting, palette, composition, and aspect-ratio decisions made by someone who already followed this guide. You bring the subject, the template brings everything else. Useful when you want consistency across a series — character cards, recipe sets, language posters, MBTI illustrations.

Browse a few examples in the gallery to see what production-grade prompts look like when they're baked into a template.