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.

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"

Real example from the gallery: /nano-banana-pro-prompts/4309 (tagged man · introspective).

> Create an urban scene with a young man standing on dark asphalt, looking upwards into a dimly lit industrial backdrop. The subject should be in casual attire with a focused expression, set against a dark background that includes a vent and railings.

Why this lands: The first noun phrase — young man standing on dark asphalt — is the subject. The style language (urban, industrial, dimly lit) comes after. Strip the style and you still know what's being rendered; that's the test of a subject-first prompt.

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

Real example from the gallery: /nano-banana-pro-prompts/4170 (tagged photorealistic · woman).

> A poised traveler, dressed in a chic blazer and trousers, strides confidently through an airport terminal during the golden hour. The sunlight filters in from the windows, casting a warm glow on her face and creating beautiful highlights that accentuate her features.

Why this lands: Even without an explicit focal length, the language highlights that accentuate her features + warm glow on her face puts Nano Banana in portrait-lens mode — shallow depth, soft compression. The next iteration would just add `shot on 85mm` to lock it in.

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

Real example from the gallery: /nano-banana-pro-prompts/4170 (tagged golden hour · photorealistic).

> A poised traveler strides through an airport terminal during the golden hour. The sunlight filters in from the windows, casting a warm glow on her face and creating beautiful highlights that accentuate her features.

Why this lands: Lighting line is doing 80% of the aesthetic work here. Replace golden hour with fluorescent overhead lighting and the same composition reads as a stock airport shot. The lighting modifier is the highest-leverage knob.

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

Real example from the gallery: /nano-banana-pro-prompts/3970 (tagged dramatic).

> A person in a vibrant red hoodie stands confidently against a backdrop of swirling shadows, their gaze directed towards the sky as a chaotic flock of black birds takes flight in an explosion of fiery red and orange hues. The scene is set against a dramatic contrast between the dark shadowy background and the explosive bursts of color.

Why this lands: One style commitment: dramatic, swirling, chaotic, explosive. Every modifier reinforces the same energy. The prompt would collapse if you added watercolor or isometric — pick one mood and double down on its vocabulary.

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

Real example from the gallery: /nano-banana-pro-prompts/4106 (tagged dynamic · energetic).

> A chaotic cooking scene in a kitchen setting. The main focus is on a clumsy chef who has dropped ingredients onto the floor and scattered food items all around him, creating an exaggerated mess. He is holding a tray with multiple pizzas that are falling.

Why this lands: Composition is explicit: main focus on clumsy chef, food items all around him, pizzas that are falling. Three sentences each name a different spatial element. Without these, Nano Banana defaults to a centered subject with an empty background.

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."

Real example from the gallery: /nano-banana-pro-prompts/4301 (tagged pastel · calm).

> A person standing in front of a room with a desk, computer monitor, keyboard, and chair that has a blue and white color scheme. The individual is wearing a sky blue top with a white logo, denim shorts, and striped socks. There's also a small plant on the desk.

Why this lands: Color is named explicitly four times — blue and white color scheme, sky blue top, white logo, denim shorts. Nano Banana respects the palette because there's nothing to interpret. Strip the color names and you'd get a randomized palette every generation.

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."

Real example from the gallery: /nano-banana-pro-prompts/4317 (tagged winter · man).

> A young man with glasses, dressed in a sweater and coat, stands alone on a snowy pathway, admiring the winter landscape. The scene captures the tranquility of a frost-covered botanical garden under a soft, pale sky. The image is presented in a vertical format with high resolution.

Why this lands: Aspect ratio called out explicitly — vertical format. Without that line, Nano Banana defaults to a horizontal landscape composition and the snowy pathway depth-of-field disappears. One sentence; one line of control.

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"

Real example from the gallery: /nano-banana-pro-prompts/4314 (tagged photorealistic · winter).

> A contemplative samurai embodying serenity and resilience during a cold winter day. The subject is a woman dressed in traditional samurai attire, her armor lightly dusted with snow. Composed and serene posture against a snow-blanketed landscape.

The targeted negatives you'd add for this prompt: no extra fingers or distorted hands gripping the sword, no blurred or asymmetric eyes, no anachronistic modern elements in the background. Specific failure modes you've actually seen in your last two generations — not cargo-cult negatives copied from a Midjourney guide.

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.

Anchor (Turn 1) — pick a gallery prompt as the base:

Real example from the gallery: /nano-banana-pro-prompts/4170 (tagged golden hour · photorealistic).

> A poised traveler in a chic blazer and trousers strides confidently through an airport terminal during the golden hour. The sunlight filters in from the windows, casting a warm glow on her face.

Iteration patterns from this base:

- Turn 2 — "same subject, same composition, but change the time of day to late blue hour. Replace the warm window light with cool overhead fluorescents."
- Turn 3 — "same lighting as turn 2, but change her outfit to a winter coat and add a small carry-on suitcase."
- Turn 4 — "same outfit and subject, but switch the location to a snowy parking lot outside the terminal."

The subject identity (her face, hair, body) stays locked across all four turns because Gemini reuses the base seed. Re-prompting from scratch would drift the face every time.

The 4,000+ prompts in /nano-banana-pro-prompts are working examples. Each one has already been generated, rendered, and saved by a real user — meaning the prompt structure passes Nano Banana's interpretation reliably. The fastest way to ship a new image is to find the closest existing prompt and modify it, not write from scratch.

The workflow that works:

1. Search by tag for the closest aesthetic. Filter /nano-banana-pro-prompts by the dominant tag — mood (playful, serene, dramatic), subject (portrait, woman), or style (photorealistic, illustration, vintage). The Library post above lists the most-populated tag clusters.
2. Pick the example whose image is closest to what you want — not whose prompt text looks similar. The image tells you what Nano Banana actually produced; the prompt is just the input.
3. Copy the prompt, swap the subject, keep the technique stack. The lens, lighting, palette, composition language stays — only the subject noun changes.

This skips 80% of the failure modes from writing a prompt cold.

Concrete example. The gallery prompt at /nano-banana-pro-prompts/4190 (tagged photorealistic, soft-light, woman) is:

> Elegant 1950s portrait of a woman in a flowing silk dress, poised at a window with gentle, soft light creating a dreamy atmosphere. The subject is wearing jewelry and her hair is styled in a vintage manner. She is looking out of the window onto the street. The background features a vintage mirror reflecting the warm interior tones.

Look at the technique stack: subject (1950s woman portrait) → action (poised at window, looking out) → setting (flowing silk dress, vintage interior) → lighting (gentle, soft light, dreamy atmosphere) → style (vintage, warm interior tones).

To produce a 1920s flapper version, you keep the entire technique stack and swap the subject decade: "Elegant 1920s flapper portrait of a woman in a sequined fringe dress, poised at a window with gentle, soft light creating a dreamy atmosphere…" Same scaffolding, different decade. Two minutes from gallery browse to live generation.

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.

Two browsing patterns work in /nano-banana-pro-prompts:

- Subject-led: when you know what you want to render. Filter by portrait, woman, man, selfie, then pick the example whose aesthetic is closest.
- Aesthetic-led: when you have a feel but not a specific subject. Filter by playful, serene, cozy, dramatic, vintage. The mood and style tag clusters are populated in the thousands.