Batch Image Generation with Nano Banana 2 API: The Automation Guide

Google’s Nano Banana 2 — built on Gemini 3.1 Flash Image — dropped today, February 26, 2026, and the jump from the original is not subtle. Subject consistency across up to five characters, genuine 4K resolution output, real-time web grounding, and text rendering that doesn’t look like someone sneezed on a keyboard. Those aren’t incremental upgrades. They’re the features that make automation actually viable.

If you’ve been manually babysitting an image generator prompt by prompt, that ends here. The Nano Banana 2 API — available through AI Studio, the Gemini API, Vertex AI, and Antigravity — lets you queue up batches, maintain visual consistency across a full asset library, and build pipelines that don’t require a human in the loop for every single frame. This guide walks through exactly how to do that, from API setup to production-ready batch workflows.

What You’ll Actually Achieve

By the end of this tutorial, you’ll have a working batch pipeline that can generate dozens of images in a single run, maintain consistent characters or brand subjects across every output, render accurate text inside images, and tap real-time web grounding for contextually current visuals. Whether you’re building a content operation, an e-commerce asset pipeline, or just automating your social media graphics, the setup is the same.

Requirements Before You Start

You need a Google account with access to Gemini API or Vertex AI — AI Studio is the fastest entry point if you’re starting fresh. Grab your API key from aistudio.google.com. Python 3.10+ is assumed throughout this guide, with the google-generativeai SDK installed (pip install google-generativeai). If you’re on Vertex AI, you’ll need a Google Cloud project with the Generative Language API enabled and your gcloud credentials sorted. Antigravity users get a slightly different endpoint but the same prompt logic applies — check your Antigravity dashboard for the base URL swap.

Step 1 — Authenticate and Initialize

Authentication is the boring part that everyone skips and then spends two hours debugging. Don’t skip it. In AI Studio, copy your API key and set it as an environment variable: export GEMINI_API_KEY=your_key_here. Then initialize the client in Python:

import google.generativeai as genai
import os

genai.configure(api_key=os.environ["GEMINI_API_KEY"])
model = genai.ImageGenerationModel("nano-banana-2")

The model string "nano-banana-2" points specifically to the Gemini 3.1 Flash Image backend. If you’re on Vertex AI, swap the initialization for the vertexai SDK and replace the model string with your project-specific endpoint.

Pro tip ✅

Store your API key in a .env file and use python-dotenv to load it. Hardcoding API keys in scripts is how you accidentally push secrets to GitHub at 2am and have a very bad morning.

Step 2 — Build Your Prompt Queue

The core of any batch pipeline is a structured prompt list. Nano Banana 2 responds well to prompts that specify style, subject, lighting, and output resolution explicitly. Vague prompts get vague results — even with a strong model. Here’s a production-ready prompt structure for a batch of five product images maintaining subject consistency:

A sleek matte black wireless speaker on a white marble surface, 
soft diffused studio lighting, product photography style, 
ultra-sharp detail, 4K resolution, no background clutter, 
photorealistic render, brand consistent subject: Speaker Model X

The brand consistent subject: Speaker Model X tag at the end is the subject consistency anchor. Nano Banana 2 uses this across all prompts in a session to keep the same object looking like the same object — same proportions, same finish, same identity. Change it per character or product in your queue to track up to five distinct subjects simultaneously.

Editorial portrait of a mid-30s South Asian woman architect, 
natural window light, architectural office background slightly blurred, 
confident expression, professional but approachable tone, 
4K resolution, photorealistic, consistent subject: Architect Character A

This editorial portrait prompt is built for publishing workflows. The lighting specification, background treatment, and tone descriptor all do real work here — Nano Banana 2 reads the full prompt, not just the noun phrase. The consistent subject: Architect Character A anchor means every image of this character in your batch will share the same face, bone structure, and approximate look.

Social media square graphic, bold sans-serif headline text reads: 
"LAUNCH DAY", vibrant coral and navy color palette, 
geometric abstract background pattern, high contrast, 
clean modern design, 1:1 aspect ratio, 4K, text rendered sharp and legible

Text rendering is where most image generators still trip over their own feet. Nano Banana 2’s precise text rendering means the phrase “LAUNCH DAY” will actually say “LAUNCH DAY” — not “LAUHCH DAV” or some cursed approximation. The trick is to quote the exact text you want inside the prompt and add text rendered sharp and legible as an explicit instruction. It works.

Pro tip ✅

Keep rendered text short — five words or fewer per text element. Nano Banana 2 handles short phrases reliably. Long paragraphs inside images are still a gamble with any generator, and this one is no exception.

Step 3 — Write the Batch Loop

With your prompt queue as a Python list, the batch loop itself is straightforward. Add rate limiting — the API has request quotas and hammering it without delays will get your key throttled faster than you’d like:

import time

prompts = [
    "A sleek matte black wireless speaker on white marble, studio lighting, 4K, photorealistic, consistent subject: Speaker Model X",
    "Same wireless speaker from a 45-degree angle, white marble, soft shadow, 4K, photorealistic, consistent subject: Speaker Model X",
    "Speaker Model X in lifestyle context, living room shelf, warm ambient light, 4K, photorealistic, consistent subject: Speaker Model X",
]

results = []

for i, prompt in enumerate(prompts):
    print(f"Generating image {i+1} of {len(prompts)}...")
    response = model.generate_images(
        prompt=prompt,
        number_of_images=1,
        aspect_ratio="1:1",
    )
    results.append(response.images[0])
    time.sleep(2)  # rate limit buffer

print(f"Batch complete. {len(results)} images generated.")

The time.sleep(2) is a minimal buffer — for large batches, push it to 3-5 seconds depending on your API tier. The loop structure here is deliberately simple; slot in your own error handling and retry logic around the generate_images call for anything going to production.

Warning ⚠️

Every image generated by Nano Banana 2 carries an invisible SynthID watermark embedded by Google. This is non-removable and persists through screenshots and format conversions. It doesn’t affect visual quality — it’s imperceptible to the eye — but if your workflow involves claiming fully original unattributed images, you should know it’s there.

Step 4 — Save and Organize Outputs

Raw API responses return image bytes. Save them with structured filenames so your batch doesn’t become an unmanageable pile of image_001.png files with no context:

import os
from datetime import datetime

output_dir = f"batch_{datetime.now().strftime('%Y%m%d_%H%M%S')}"
os.makedirs(output_dir, exist_ok=True)

for i, image in enumerate(results):
    filename = f"{output_dir}/img_{i+1:03d}.png"
    image.save(filename)
    print(f"Saved: {filename}")

Timestamped output directories keep batches separated when you’re running multiple jobs in a day. The zero-padded index (:03d) ensures files sort correctly in any file browser — a small thing that matters a lot when you’re looking at 50+ images.

Step 5 — Real-Time Web Grounding in Batch Context

Nano Banana 2’s real-time web grounding means the model can pull current visual context when generating — useful for anything tied to real-world aesthetics, trends, or specific environments. In a batch workflow, you can use this for prompts that need to reflect current events, seasonal contexts, or trending visual styles without manually updating every prompt:

Street photography style, busy Tokyo Shibuya crossing, 
current February 2026 winter fashion visible on pedestrians, 
natural ambient light, overcast sky, 4K, photorealistic, 
real-world grounded aesthetic, no staged look

The grounding kicks in automatically when the prompt references real-world specifics. You don’t pass a separate flag — the model recognizes contextual anchors and pulls current reference data. For evergreen product shots, grounding doesn’t add much. For anything trend-sensitive, it’s the difference between visuals that look current and visuals that look like they were generated eight months ago.

Pro tip ✅

When using web grounding for trend-sensitive batches, run a small test batch of three images first. Grounded outputs can vary more than fully synthetic ones — verify the aesthetic is consistent before committing to a full run of 50.

Step 6 — Multi-Character Consistency Across a Batch

The five-character consistency feature is Nano Banana 2’s headline capability and the one most likely to change how editorial and storytelling workflows operate. Here’s how to run a batch that maintains two distinct characters across multiple scenes:

Two characters in a modern café: 
Character A: tall Black man, late 20s, orange puffer jacket, natural hair, 
Character B: petite East Asian woman, early 30s, green oversized blazer, reading glasses, 
Scene: both seated at a wooden table, morning light through large windows, 
candid conversation, 4K, photorealistic, 
consistent subjects: Character A, Character B

Same two characters from previous scene now walking outside on a city street, 
Character A: orange puffer jacket, natural hair, 
Character B: green oversized blazer, reading glasses, 
midday natural light, urban background, 4K, photorealistic, 
consistent subjects: Character A, Character B

The key is restating the visual identifiers for each character in every prompt — don’t assume the model will remember the full description from a previous call. Within a session, the consistency anchors do the heavy lifting, but reinforcing the distinctive visual traits (the orange jacket, the reading glasses) keeps outputs tighter across a long batch.

Note 💡

Subject consistency works within a single API session. If you close the session and start a new one, the consistency context resets. For long-running projects, complete your full batch in one session rather than splitting across multiple runs.

Avoid 🚫

Don’t overload one prompt with more than five consistent subjects. Nano Banana 2 caps at five for a reason — beyond that, consistency degrades noticeably and you’ll spend more time on manual corrections than the automation saved you.

Nano Banana vs. Nano Banana Pro in Batch Workflows

Nano Banana Pro offers higher request rate limits, priority processing during peak hours, and extended session length — which matters for large batches. The free and standard tiers throttle more aggressively and sessions time out faster. For a 10-image test batch, standard is fine. For a 200-image e-commerce run that needs to complete in under an hour, Pro is the tier that makes it practical. The prompt logic and subject consistency behavior are identical across both — it’s infrastructure, not capability, that separates them.

What This Means for Your Workflow

Batch automation with Nano Banana 2 is genuinely useful in a way that’s hard to oversell without sounding like a press release — so here’s the concrete version: a content team that was hand-generating 20 product images a day can run 200 before lunch. An editorial operation can maintain visual consistency across a character-driven series without a dedicated illustrator babysitting every frame. A solo creator can build a week of social content in a single automated run. The SynthID watermark is always there, the five-character cap is a real limit, and you still need to write good prompts — but the infrastructure to scale image generation into a real production workflow is now sitting in a free API call away.

promptyze

See Full Bio