Nano Banana 2 Lite Prompts: Fast, Cheap Image Generation with Gemini 3.1 Flash Lite

Most people who read Simon Willison's writeup on Nano Banana 2 Lite close the tab knowing the model exists and nothing else. That's where this guide picks up. Below you'll find copy-paste prompts tested against the model's actual strengths, a working API snippet, and a clear-eyed look at where Nano Banana 2 Lite earns its place in a workflow versus where you should route the job elsewhere.

Nano Banana 2 Lite is Google's internal nickname for Gemini 3.1 Flash Lite Image, a lightweight model built for speed and low cost. It runs fast, it runs cheap, and it doesn't pretend to be Midjourney. If you go in with the right prompts and the right expectations, it's genuinely useful for ad concepting, social media mockups, and rapid visual brainstorming. If you go in demanding photorealism, you'll be disappointed every time.

What Nano Banana 2 Lite Actually Is (and What It Isn't)

Nano Banana 2 Lite is Google's lightweight Gemini 3.1 Flash Lite image model, designed for speed over fidelity. It's cheap to run, fast to return results, and best pointed at tasks that match that profile.

Where it fits well:

Rapid ad concepting before committing to a designer
Social media visual drafts for internal review
Style exploration when you're not sure what direction you want
High-volume thumbnail or placeholder generation

Where it doesn't:

Final production assets
Photorealistic portraits or product photography
Complex multi-element scenes with spatial accuracy
Anything where a client will see the raw output

The biggest mistake beginners make with this model is prompting it like they'd prompt DALL-E 3. Long, detailed descriptions of lighting, texture, camera angle, and post-processing yield mediocre results here. Short, style-forward prompts yield surprisingly good ones.

Here's what that looks like in practice:

Before (generic, over-specified):
Generate an image of a coffee shop.

A cozy, brightly lit coffee shop interior with warm wood tones, focus on atmosphere, digital art style, low detail.

Adding "low detail" isn't about lowering your standards. It's telling the model where not to spend its limited capacity. Specifying an artistic style ("digital art style") gives the model a constraint it can actually satisfy. The result lands closer to what you want because you've matched the prompt to the model's capabilities.

If you want a deeper look at how prompt framing shapes image output across different tools, the guide on cutting image prompt writing time while keeping creative control covers the same principle applied more broadly.

The Best Prompts for Ad Mockups and Visual Brainstorming

The model responds well to style keywords over descriptive keywords. "Cyberpunk" outperforms "a futuristic city at night with neon signs reflecting off wet pavement." "Minimalist" outperforms "clean, uncluttered, with lots of white space and a single focal element." Give it a genre and a subject. Let it fill the rest.

Ad Mockup Prompts

A vibrant advertisement for a new eco-friendly running shoe, minimalist design, focus on sustainability and performance, athletic lifestyle, digital illustration, --no people

Product mockup for a premium cold brew coffee brand, muted earth tones, overhead flat lay, clean background, label visible, lifestyle photography style, --no text

Social media banner for a fintech app launch, bold geometric shapes, dark background with electric blue accents, modern sans-serif aesthetic, digital art, --no photorealism

Visual Brainstorming Prompts

Futuristic cityscape with flying cars and neon lights, cyberpunk aesthetic, concept art, low poly style, --no photorealism

Brand mood board for a sustainable outdoor apparel company, earthy greens and tans, flat lay of gear and natural textures, editorial style, --no people

Logo concept exploration for a wellness brand, soft gradients, botanical motifs, minimalist circular composition, vector illustration style

A few patterns worth noting across these prompts:

Style anchors first ("digital illustration," "concept art," "editorial style")
Subject and mood second
Negative prompts at the end to cut the most common failure modes
No sentence structure, no verbs. Just descriptive phrases.

If you're running these for a client-facing marketing workflow, the Figma AI prompts guide pairs well here. You can rough out concepts with Nano Banana 2 Lite, then refine in Figma once a direction earns approval.

Integrating Nano Banana 2 Lite via the API

Running the model manually is fine for exploration. For anything at volume, you want the API.

Here's the setup process:

Step 1. Create a free account at Google AI Studio and generate an API key from the dashboard.

Step 2. Install the current Python client library for Google's Generative AI API:

pip install google-genai

Step 3. Run your first image generation call. Before you do, check the current model identifier string in the google-genai documentation or your AI Studio model list, since the exact string for Gemini 3.1 Flash Lite Image can shift between preview and stable release. Swap it in where you see "YOUR_MODEL_STRING" below:

from google import genai

client = genai.Client(api_key="YOUR_API_KEY")

response = client.models.generate_content(
    model="YOUR_MODEL_STRING",  # look up the current Gemini 3.1 Flash Lite Image identifier in the google-genai docs
    contents="A cozy coffee shop interior, digital art style, warm tones, low detail"
)

# Extract the image from the response — see the google-genai docs for current response handling

The call structure above is correct: client.models.generate_content() is the right method for image generation with this SDK. The only thing to confirm before deploying is the model identifier string and the current response-handling path, both of which the live documentation will show you.

Once you have the basic call working, you can drop it into a loop to generate multiple concept variations from a single prompt, or parameterize the prompt itself to test style variations at scale:

from google import genai

client = genai.Client(api_key="YOUR_API_KEY")

styles = ["digital illustration", "watercolor", "low poly", "pixel art"]
base_prompt = "Brand concept for an eco-friendly running shoe, minimalist, --no people"

for style in styles:
    prompt = f"{base_prompt}, {style}"
    response = client.models.generate_content(
        model="YOUR_MODEL_STRING",  # confirm current Gemini 3.1 Flash Lite Image identifier in the google-genai docs
        contents=prompt
    )
    # process and save response

This kind of loop is what makes Nano Banana 2 Lite genuinely useful for concepting work. You're not generating one image. You're generating eight in the time it would take a designer to open a new file. You bring the taste and the final call; the model handles the volume.

If API-based workflows are new to you, the post on running AI locally covers the infrastructure basics before you get into model-specific integration.

Cost: What You Actually Save with Nano Banana 2 Lite

This is where the model earns its name in a workflow.

Because Nano Banana 2 Lite is a lightweight model, it sits at the cheaper end of Google's image generation pricing compared to heavier, higher-fidelity alternatives. The per-call cost is lower, and that gap compounds fast once you start iterating at volume. Check the current pricing in your Google AI account, since Google updates it as models move from preview to stable release.

What the raw per-call number doesn't capture is the iteration cost. When each call is cheap, you run more experiments. More variation means better concepting. Better concepting means fewer rounds of revision with a designer. That's where the real savings compound: not in any single image, but in the 50-image exploration session you'd never greenlight if you were paying full price for a production model.

One honest caveat: if you're generating only a few images per month, the cost difference between model tiers may not matter much in absolute terms. Nano Banana 2 Lite's advantage grows with volume. For occasional use, just pick whichever model gives you results you can work with.

Limitations: What Will Go Wrong and How to Work Around It

The model struggles with three categories of requests. Knowing them in advance saves frustration.

Photorealism fails fast

Prompt that will underdeliver:
A hyperrealistic portrait of a woman with intricate skin texture, detailed eyes, studio lighting.

The output will be blurry, distorted, or uncanny. Don't fight the model's weight class.

An oil painting style portrait of a woman, impressionistic brushstrokes, soft warm lighting, painterly aesthetic.

Framing the same subject as an artistic style rather than a realistic photograph gives the model a target it can actually hit. The result won't be photorealistic, but it can be genuinely compelling.

Complex spatial scenes lose coherence

Multiple characters interacting, architectural interiors with accurate perspective, crowd scenes: all of these push the model past its reliable range. Stick to single subjects or simple compositions.

Series consistency is unreliable

If you need five images that look like they belong to the same visual identity, Nano Banana 2 Lite will struggle to maintain it without significant prompt engineering. For that use case, a larger model or a diffusion approach with seeded generation will serve you better.

When you hit these limits, the decision framework is simple: if the output is for internal concepting, Nano Banana 2 Lite is still faster and cheaper even with imperfect results. If the output is client-facing or production-ready, route it to a heavier model. The local AI vs. cloud AI breakdown covers exactly this kind of routing decision across model types.

FAQ

How do I use Gemini 3.1 Flash Lite Image for ad mockups without spending much?

Keep prompts short and style-forward: name an aesthetic ("digital illustration," "flat design") before you describe the subject. Use negative prompts to block the most common failure modes like photorealism and excess detail. Because Nano Banana 2 Lite sits at the cheaper end of Google's image model lineup, you can run more variations per dollar than you would with a heavier model, which means better concepting material without a big API bill.

What prompts work best with Nano Banana 2 Lite for quick visual concepts?

Short descriptive phrases anchored by a style keyword. "Cyberpunk cityscape, concept art, low poly, neon lighting" outperforms a three-sentence description every time. Artistic styles ("watercolor," "pixel art," "isometric illustration") play to the model's strengths. Photorealistic descriptions do not.

Can I integrate Nano Banana 2 Lite into my existing workflow via the API?

Yes. Install google-genai via pip, configure a client with your API key, and call client.models.generate_content() with the Gemini 3.1 Flash Lite Image model string. Look up the exact current model identifier in the live google-genai documentation or AI Studio before deploying, as it can change between preview and stable release.

How much cheaper is Nano Banana 2 Lite compared to standard Gemini image models?

It's priced below heavier Gemini image models, so the per-call cost is lower. The exact figures are worth checking directly in your Google AI account since pricing updates as models move through release stages. The bigger advantage isn't any single call — it's that cheaper per-call costs change how many variations you're willing to run, and more variation produces better concepting outcomes.

What are the limitations of Nano Banana 2 Lite for professional design work?

Poor performance on photorealistic requests, limited spatial coherence in complex scenes, and unreliable consistency across a series of images. It's a concepting and brainstorming tool, not a production design tool. Treat the outputs as a starting point that earns a direction, not a finished asset.

The One Rule for This Model

Match the prompt to the model, not the model to the prompt. Nano Banana 2 Lite is fast, cheap, and surprisingly capable when you give it stylistic targets and simple compositions. It breaks when you push it toward realism and detail density it wasn't designed for.

Use it to get eight rough concepts on the table before your next creative brief. Pick the direction that works, then decide whether the final execution needs a heavier model or a human designer. That's the workflow it was built for.

If you want tested prompt templates for image generation already formatted for this kind of concepting work, Ultra Prompt's Image Generation category has them ready to drop into your workflow.

Ultra Prompt