How to prompt Nano Banana 2 Lite for fast, sharp images

Google announced Nano Banana 2 Lite on June 30, 2026 — gemini-3.1-flash-lite-image, the fastest and cheapest model in the Gemini image family. It generates a 1K image in about 4 seconds for $0.034, roughly 2.7x faster than Nano Banana 2, and still lands at #5 overall on the Arena.ai Text-to-Image Arena with an Elo near 1,251. Prompting it well is less about compensating for a weaker model and more about knowing exactly which dials moved.

Speed and price aren't the only things that shift at the Lite tier. Resolution drops, Google Search grounding disappears, and — counterintuitively — the reference-image ceiling actually goes up. Getting fast, sharp results means understanding which trade-offs matter for your job and which prompting habits compensate for the ones that do.

What actually changes at the Lite tier

Two numbers are worth sitting with. Lite's $0.034 for a 1K image matches Nano Banana 2's batch-tier price and undercuts its standard tier by half — at real-time generation speed, not batch turnaround. And despite being the budget tier, Lite retains the things that actually decide whether an image is usable: character consistency, prompt adherence, and legible in-image text. Google's own positioning is "no compromise on quality" at this speed and price, with resolution as the one real ceiling.

Thinking mode is also on by default at this tier, controlled through thinking_level: minimal or high. A higher thinking level generates interim composition tests before producing the final image — worth the extra fraction of a second on prompts with several elements that need to agree with each other. Aspect ratio support is unchanged from the rest of the family: 1:1, 3:2, 2:3, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9, and 21:9 are all available, so framing belongs in the prompt itself rather than an afterthought.

The counterintuitive part: more reference images, not fewer

The instinct with a "lite" model is to assume every dial gets turned down. Reference images break that pattern. Nano Banana 2 Lite accepts up to 14 reference or object images per generation — more than Nano Banana 2's cap of 10 object images plus 4 character images.

That makes Lite a genuinely strong choice for jobs built around a large reference set: a product line with a dozen SKUs, a character sheet with many outfit or prop variants, a brand kit with several logo lockups that all need to appear together and stay consistent. You give up top-end resolution, but you gain headroom on the one input type that determines whether a multi-reference generation holds together at all. If a job needs more than ten references and doesn't need 4K output, Lite is the better-suited tool, not the compromise.

A prompt template that respects the tier

Google's guidance for Nano Banana 2 Lite is consistent with how the whole Gemini image family responds: rich, specific detail increases control, and a template-based structure outperforms a loose keyword list. Five slots cover almost every brief:

[Shot type] of [specific subject with descriptive detail], in [setting],
lit by [lighting direction and quality], shot from [camera angle / lens
characteristic]. [Style descriptor: medium + visual qualities].

Work through it as a checklist before sending anything:

• Shot type — close-up, three-quarter product shot, wide establishing shot. Don't leave the model to guess framing.
• Subject — specific, not generic. "A 34-year-old ceramicist with flour-dusted forearms" beats "a person" every time.
• Setting — where and when, including anything in the background that matters.
• Lighting — direction, quality, and color temperature. "Soft window light from the left" produces a different image than "hard overhead fluorescent."
• Camera angle / lens — eye-level versus low angle, 35mm versus macro. This does more compositional work than most prompts give it credit for.
• Style descriptors — name the medium (photograph, watercolor, 3D render) and the visual qualities (film grain, glossy studio finish) you want; descriptors are what keep a batch of variants consistent with each other.

This is the same structure that works across any Gemini image tier — it's just that at roughly 4 seconds a generation, you can afford to run the template five or six times with one slot changed and pick the best result, instead of hand-tuning a single prompt.

Five prompts to try

1. Product hero, single reference. "Three-quarter product shot of a matte-ceramic white sneaker on a slate-gray pedestal, softbox key light from top-left with a subtle rim light from behind, 50mm macro lens, shallow depth of field, editorial product-catalog style, no props." At $0.034 and ~4 seconds a frame, ten angle variants cost less than one revision cycle on a slower model.

2. Multi-reference consistency, testing the 14-image ceiling. "Using the attached reference set, generate a three-quarter studio portrait of the same character in a pose and lighting that match the references. Keep facial features, outfit, and color palette identical to the reference images; change only the background to a warm gradient studio backdrop." This is the job Lite's extra reference headroom is built for.

3. Editorial portrait, thinking_level high. "Candid editorial portrait of an elderly luthier in his workshop, golden-hour light through a dusty window, 85mm lens, shallow depth of field, natural film grain, documentary photography style." Complex scenes benefit from raising thinking_level to high — the interim composition pass earns its keep when a brief has this many parts to reconcile.

4. Quick social variant, thinking_level minimal. "Flat-lay of a matcha latte and a linen napkin on a marble countertop, soft overhead daylight, top-down angle, minimal aesthetic, muted pastel palette." For high-volume, low-complexity content, minimal thinking keeps throughput high without shortchanging the composition.

5. Style-descriptor stress test. "Watercolor illustration of a lighthouse on a rocky coast at dusk, visible paper texture, loose wet-on-wet washes, muted indigo and rust palette, hand-lettered caption area left blank." Naming the medium and the specific technique — not just "watercolor style" — keeps a run of these visually consistent.

Multi-turn editing without restarting the prompt

Nano Banana 2 Lite supports iterative refinement through previous_interaction_id — you reference the prior generation instead of re-describing the whole scene. A typical thread:

• Turn 1: Generate the base image with a full prompt.
• Turn 2: "Using the previous generation, change the jacket color to burgundy and add rain-soaked reflections on the pavement."
• Turn 3: "Zoom to a tighter three-quarter crop and warm the color grade slightly."

Each turn only needs to describe the change, not the whole scene — turning Lite's speed into an actual iteration loop instead of three separate cold-start prompts.

Preservation instructions protect what you don't want touched

The most common failure mode in iterative editing is scope creep: you ask for one change and get three. Explicit preservation instructions fix this. Add a clause like "keep the background unchanged" or "keep the subject's pose and expression identical" to any edit turn, and the model treats it as a hard constraint rather than a suggestion.

Where Nano Banana 2 Lite still trails

Two limits are worth planning around. Resolution tops out at 1K — 0.5K and 1K are the only outputs, with no 2K or 4K path — so large-format print or billboard work should go to a higher tier. And Lite doesn't support Google Search grounding, so prompts leaning on current events or live data won't be grounded in anything real; send those to a different model or pair them with a manual fact check.

Which techniques carry over to Nano Banana 2 on OmniArt today

This is the part that matters if you don't have Lite access yet. Nano Banana 2 (gemini-3.1-flash) is live on OmniArt's image workspace, tagged new and hot, and it shares enough prompt grammar with Lite that most of the above transfers directly.

In other words, the prompting discipline is the transferable skill, not the tier's specific limits. Build the shot-type-through-camera-angle habit on Nano Banana 2 today, and it carries over the moment Lite — or any future tier — shows up in the workspace.

For the full three-way spec breakdown, see Nano Banana 2 Lite vs 2 vs Pro: which Gemini model to use. And for the video-side counterpart to this release, Gemini Omni Flash's developer API: what's new since I/O covers what Google shipped alongside it.

FAQ

Is Nano Banana 2 Lite available on OmniArt?

Not yet. Google announced Nano Banana 2 Lite (gemini-3.1-flash-lite-image) on June 30, 2026, and it's currently available through Google AI Studio and the Gemini API. Nano Banana 2, the tier above it, is live on OmniArt today.

What's the actual difference between Nano Banana 2 Lite and Nano Banana 2?

Speed and resolution mainly. Lite generates in about 4 seconds versus roughly 2.7x longer for Nano Banana 2, at $0.034 per 1K image. In exchange, Lite caps out at 1K resolution — no 2K or 4K — and doesn't support Google Search grounding.

Why does the "lite" tier support more reference images than the flagship?

It's a genuine trade-off, not an oversight. Lite accepts up to 14 reference or object images per generation, versus Nano Banana 2's 10 object plus 4 character images. If a job depends on a large reference set more than it depends on 4K output, Lite is the better-suited tool.

Can I get 4K output from Nano Banana 2 Lite?

No. Lite's maximum resolution is 1K, with 0.5K and 1K as the only supported outputs. For 2K or 4K, use Nano Banana 2 or Nano Banana 2 Pro.

Does Nano Banana 2 Lite support Google Search grounding?

No. Nano Banana 2 does; Lite doesn't. Prompts that depend on current events or live data should go to a model that supports grounding, or be paired with a manual fact check.

How do I do multi-turn edits with Nano Banana 2 Lite?

Reference the prior generation using previous_interaction_id and describe only the change you want — a color swap, a crop, a lighting adjustment. Pair it with an explicit preservation instruction ("keep the background unchanged") to stop the model from altering parts of the image you didn't ask about.

Getting started on OmniArt

Nano Banana 2 Lite isn't in the workspace yet, but the prompting discipline it rewards — specific subjects, named lighting, explicit preservation instructions — is exactly what gets better results out of Nano Banana 2 on OmniArt. Open the image workspace, run the shot-type-through-camera-angle template on Nano Banana 2, and once you have a hero image you like, carry it into the video workspace with the photo-to-product-video workflow. For more on prompt structure across models, see how to write better prompts, and for the full flagship comparison, GPT Image 2 vs Nano Banana 2.