Nano Banana 2.1 Guide: Generate, Edit, and Export Images

Nano Banana 2.1 Guide: How to Generate, Edit, and Export Images in 1K, 2K, and 4K

Nano Banana 2.1 is more than a text-to-image model. It accepts text, images, video, and PDF inputs, supports conversational image editing, combines multiple reference images, controls aspect ratios and resolutions, and offers configurable thinking levels for complex tasks.

Google positions it as its latest high-efficiency model for image generation and conversational editing. Compared with Nano Banana 2, Google says the update improves visual quality, prompt adherence, multi-turn character consistency, text rendering, and performance across panoramic formats.

In this guide, you will use the Gemini API to:

• Generate images from text
• Edit images with visual references
• Export images in 1K, 2K, or 4K
• Control the output aspect ratio
• Configure the thinking level
• Combine multiple reference images
• Continue editing through multi-turn conversations
• Ground image generation with Google Search
• Generate new images from video content

4K Image Generation	Video-to-Image Generation
An official high-resolution image-generation example.	An official example generated from the visual context of a YouTube video.

What Is Nano Banana 2.1?

The official model code for Gemini Nano Banana 2.1 is:

gemini-nano-banana-2.1

According to the official Google model page, Nano Banana 2.1 is an update to Nano Banana 2 designed for:

• Image generation
• Conversational image editing
• Multi-reference image fusion
• Multi-turn image refinement
• Visual content containing text
• Infographics and marketing assets
• Character and object consistency
• Search-grounded image generation

The model was released on October 6, 2026, and is currently listed as a stable model.

Which Inputs and Outputs Does Nano Banana 2.1 Support?

Item	Support
Text input	Supported
Image input	Supported
Video input	Supported
PDF input	Supported
Image output	Supported
Text output	Supported
Audio input	Not supported for image generation
Audio output	Not supported
Structured output	Not supported
Live API	Not supported
Batch API	Supported
Google Web Search grounding	Supported
Google Image Search grounding	Supported

The model has an input limit of 131,072 tokens and an output limit of 32,768 tokens. See the Nano Banana 2.1 model page for current capabilities and restrictions.

What You Need Before Starting

To call the Nano Banana 2.1 API, prepare:

1. A Google account
2. A Gemini API key
3. A Python or JavaScript environment
4. The Google GenAI SDK
5. A local directory for generated images
6. Reference images that you have permission to use, if you plan to edit existing assets

The examples below use Python.

Step 1: Generate an Image from Text

Here is a basic text-to-image example:

from google import genai
import base64

client = genai.Client()

prompt = """
Create a premium commercial product photograph of a transparent
glass perfume bottle on a polished dark stone surface.

Use soft directional studio lighting, realistic reflections,
subtle shadows, a clean luxury composition, and no visible logo.
"""

interaction = client.interactions.create(
    model="gemini-nano-banana-2.1",
    input=prompt,
)

with open("generated_image.png", "wb") as output_file:
    output_file.write(
        base64.b64decode(interaction.output_image.data)
    )

This code:

1. Calls gemini-nano-banana-2.1
2. Sends the text prompt as input
3. Retrieves the returned image data
4. Saves it as generated_image.png

Step 2: Set the Aspect Ratio and Resolution

Nano Banana 2.1 generates 1K images by default. Use response_format when you need a specific aspect ratio or resolution:

from google import genai
import base64

client = genai.Client()

prompt = """
Create a cinematic ecommerce campaign image featuring a luxury
perfume bottle in the center, surrounded by deep red silk,
soft golden light, realistic glass reflections, and generous
negative space for advertising copy.
"""

interaction = client.interactions.create(
    model="gemini-nano-banana-2.1",
    input=prompt,
    response_format={
        "type": "image",
        "aspect_ratio": "16:9",
        "image_size": "2K",
    },
)

with open("campaign_16x9_2k.png", "wb") as output_file:
    output_file.write(
        base64.b64decode(interaction.output_image.data)
    )

Supported Resolutions

Setting	Recommended Use
1K	Prompt testing, composition previews, and standard online content
2K	Marketing images and product visuals that require more detail
4K	High-resolution assets, large-format layouts, and further design work

Nano Banana 2.1 does not support 0.5K output.

Supported Aspect Ratios

Format	Available Ratios
Square	1:1
Standard landscape	3:2, 4:3, 5:4
Standard portrait	2:3, 3:4, 4:5
Mobile portrait	9:16
Landscape video	16:9
Ultrawide	21:9
Panoramic	1:4, 4:1, 1:8, 8:1

According to the Google image-generation documentation, Nano Banana 2.1 improves tiling artifacts that could appear in 2K and 4K panoramic formats.

This is an official update claim. It does not guarantee that every panoramic generation will be free of composition problems or repeated textures.

Step 3: Control the Thinking Level

Nano Banana 2.1 supports three thinking levels:

Level	Recommended Use
minimal	Simple generations, localized edits, or latency-sensitive tasks
medium	The default setting for most generation and editing tasks
high	Multi-element compositions, multiple references, and complex instructions

Configure the thinking level with generation_config:

from google import genai
import base64

client = genai.Client()

interaction = client.interactions.create(
    model="gemini-nano-banana-2.1",
    input="""
    Create a futuristic city built inside a giant transparent
    glass bottle floating in space. Preserve physically plausible
    reflections, atmospheric depth, internal architecture, and
    a clearly readable silhouette.
    """,
    generation_config={
        "thinking_level": "high"
    },
)

with open("high_thinking_result.png", "wb") as output_file:
    output_file.write(
        base64.b64decode(interaction.output_image.data)
    )

Google describes thinking levels as a way to balance quality and latency. Choosing high does not guarantee that every result will better match your preferences, and it cannot compensate for an unclear prompt.

For ordinary tasks, start with the default medium setting. Try high when the composition is complex, multiple references are involved, or the model overlooks important relationships.

Step 4: Edit an Image

Nano Banana 2.1 can combine an image with text instructions to add, remove, or modify visual elements.

The official Google documentation demonstrates this workflow with a photograph of an orange cat and an edited result featuring a knitted wizard hat.

Original Image	Edited Result
A photograph of an orange cat	The original style is preserved while a knitted wizard hat is added

Python structure:

from google import genai
import base64

client = genai.Client()

with open("/path/to/cat.png", "rb") as image_file:
    image_bytes = image_file.read()

interaction = client.interactions.create(
    model="gemini-nano-banana-2.1",
    input=[
        {
            "type": "text",
            "text": """
            Add a small knitted wizard hat to the cat.
            Preserve the cat's identity, fur texture, facial
            expression, original lighting, perspective, and
            background.
            """
        },
        {
            "type": "image",
            "data": base64.b64encode(image_bytes).decode("utf-8"),
            "mime_type": "image/png"
        }
    ],
)

with open("cat_with_hat.png", "wb") as output_file:
    output_file.write(
        base64.b64decode(interaction.output_image.data)
    )

Before uploading an image, confirm that you have the necessary rights to use and edit it.

Step 5: Combine Multiple Reference Images

Nano Banana 2.1 accepts up to 14 reference images in a single workflow.

Google further divides this capacity into:

• Up to four character images for character consistency
• Up to ten object images for object-detail fidelity
• Up to 14 total reference images

Multi-Reference Character Example

This example from the official documentation demonstrates a multi-character composition. A single result does not prove that every generation will consistently preserve four characters, so each output still requires inspection.

Multi-Image Ecommerce Example

Clothing Reference	Model Reference	Combined Result
A flat-lay image of a blue floral summer dress	A full-body female model reference	A complete ecommerce scene combining the model and dress

Basic code structure:

from google import genai
import base64

client = genai.Client()

def encode_image(path):
    with open(path, "rb") as image_file:
        return base64.b64encode(
            image_file.read()
        ).decode("utf-8")

interaction = client.interactions.create(
    model="gemini-nano-banana-2.1",
    input=[
        {
            "type": "text",
            "text": """
            Dress the woman from the second reference image in
            the exact blue floral dress from the first reference.

            Preserve the model's facial identity, hairstyle,
            body proportions, and pose. Preserve the dress pattern,
            colors, neckline, sleeves, length, and fabric appearance.

            Present the result as a realistic outdoor fashion
            ecommerce photograph.
            """
        },
        {
            "type": "image",
            "data": encode_image("dress.png"),
            "mime_type": "image/png"
        },
        {
            "type": "image",
            "data": encode_image("model.png"),
            "mime_type": "image/png"
        }
    ],
    generation_config={
        "thinking_level": "high"
    },
    response_format={
        "type": "image",
        "aspect_ratio": "4:5",
        "image_size": "2K",
    },
)

with open("fashion_result.png", "wb") as output_file:
    output_file.write(
        base64.b64decode(interaction.output_image.data)
    )

Do not upload more references merely to reach the 14-image limit. Every image should have a defined role, such as character identity, clothing, product angle, material, or composition.

Step 6: Preserve Character Consistency Across Different Angles

Nano Banana 2.1 can use one character reference to generate additional viewing angles.

Google recommends including previously generated images in later prompts to improve consistency. For complex poses, you can also provide a separate pose reference.

Original Character	Right-Side Profile	Front View
Original identity reference	A profile preserving the face, glasses, and clothing	A front view preserving the same identity

A prompt for changing the viewing angle can follow this structure:

Create a right-side profile view of the same exact adult man.

Preserve his facial identity, facial structure, white glasses,
hairstyle, skin tone, clothing, body proportions, photographic
style, background, and lighting.

Only change the camera angle. Do not redesign or beautify the person.

The key is not merely asking for a side view. The prompt must also lock every identity attribute that should remain unchanged.

Step 7: Continue Editing Through Multi-Turn Conversations

Nano Banana 2.1 can continue editing a previous result without requiring you to describe the entire image again.

In Google’s official example, the model first generates an English infographic about photosynthesis. A subsequent instruction changes its text to Spanish.

First Generation: English	Second Edit: Spanish
The initial infographic structure and English content	The same layout with the text translated into Spanish

The key parameter is:

previous_interaction_id

Create the first image:

from google import genai
import base64

client = genai.Client()

first_result = client.interactions.create(
    model="gemini-nano-banana-2.1",
    input="""
    Create a vibrant infographic that explains photosynthesis
    as if it were a recipe for a plant's favorite food.

    Show the ingredients: sunlight, water, and CO2.
    Show the finished dish as sugar and energy.
    Use a colorful children's cookbook style.
    """
)

with open("photosynthesis_en.png", "wb") as output_file:
    output_file.write(
        base64.b64decode(first_result.output_image.data)
    )

Continue the interaction:

second_result = client.interactions.create(
    model="gemini-nano-banana-2.1",
    input="""
    Update this infographic to be in Spanish.
    Do not change any other elements of the image.
    """,
    previous_interaction_id=first_result.id,
    response_format={
        "type": "image",
        "aspect_ratio": "16:9",
        "image_size": "2K",
    },
)

with open("photosynthesis_es.png", "wb") as output_file:
    output_file.write(
        base64.b64decode(second_result.output_image.data)
    )

A multi-turn edit should explain:

1. What needs to change
2. What must remain unchanged

This gives the model a clearer editing boundary than describing only the desired change.

Step 8: Use Google Search Grounding

Nano Banana 2.1 can use Google Web Search and Google Image Search results as grounding for image generation.

The official documentation demonstrates this workflow by generating a weather chart from time-sensitive information.

Enable web and image search:

from google import genai
import base64

client = genai.Client()

interaction = client.interactions.create(
    model="gemini-nano-banana-2.1",
    input="""
    Create a clear visual weather forecast for San Francisco
    covering the next five days. Use an easy-to-read horizontal
    infographic layout with weather icons and temperatures.
    """,
    tools=[
        {
            "type": "google_search",
            "search_types": [
                "web_search",
                "image_search"
            ]
        }
    ],
)

with open("grounded_weather.png", "wb") as output_file:
    output_file.write(
        base64.b64decode(interaction.output_image.data)
    )

Keep these limitations in mind:

• One request may trigger one or more search queries
• Search grounding may result in additional charges
• Real-world images of people returned by search cannot currently be used as image-generation references
• Image Search grounding requires compliance with Google’s citation and search-suggestion display requirements
• Grounding does not guarantee that every detail in the final image is completely accurate

Step 9: Inspect the Result

Review the following before approving an output:

Check	What to Review
Prompt adherence	Whether the subject, action, environment, and composition appear
Reference preservation	Whether people, products, colors, and structures remain intact
Anatomy	Whether hands, feet, faces, and body proportions look natural
Text rendering	Spelling, hierarchy, line breaks, and information relationships
Composition	Whether subjects are cropped or overlap
Aspect ratio	Whether the output matches the requested ratio
Resolution	Whether the target image size was used
Multi-turn editing	Whether unchanged areas were modified
Object count	Whether objects are missing or duplicated
Search-grounded information	Whether time-sensitive data needs manual verification
Watermark	Whether the presence of SynthID is understood

Google states that the model may not always return the exact number of requested images or objects. Any task involving precise quantities still requires manual inspection.

How to Write More Reliable Prompts

Be Specific

Instead of:

Create fantasy armor.

Write:

Create ornate silver elven plate armor engraved with leaf motifs,
featuring a high collar and shoulder pieces shaped like falcon wings.

Explain the Intended Use

Instead of simply asking the model to design a poster, specify its purpose:

Create a vertical product launch poster for a premium minimalist
skincare brand, designed for an ecommerce homepage.

Break Complex Tasks into Steps

For a complex scene, describe the sequence:

1. Create the background
2. Position the primary subject
3. Add supporting elements
4. Define the lighting and camera
5. Identify the details that must remain unchanged

Use Photography and Cinematography Language

For example:

• wide-angle shot
• macro shot
• eye-level camera
• low-angle perspective
• shallow depth of field
• soft studio lighting
• cinematic rim lighting

Use Positive Semantic Descriptions

Instead of:

No cars.

Write:

An empty and quiet street with no signs of traffic.

Frequently Asked Questions

Why Does Text Still Contain Spelling Errors?

For text-heavy images:

1. Generate and confirm the copy first
2. Ask the model to place the approved text into the image
3. Check spelling, line breaks, and hierarchy
4. Correct localized errors through multi-turn editing

Which Thinking Level Should I Use?

• Simple edits: start with minimal
• Standard generation: use the default medium
• Multi-reference or complex compositions: try high

These are configuration guidelines, not fixed quality guarantees.

When Should I Use 4K?

Consider 4K when you need large-format output, detailed inspection, or further layout and design work.

For prompt and composition testing, begin with 1K or 2K. Generate the 4K version after confirming the direction to avoid unnecessary testing costs.

Current Nano Banana 2.1 Limitations

• Image generation does not support audio input
• 0.5K image output is not supported
• The model may not return the exact requested number of images
• Search grounding cannot use real-world images of people returned by search
• URL context is not supported
• Live API is not supported
• Function calling is not supported
• Structured output is not supported
• All generated images include SynthID
• Multi-character scenes, text, and complex compositions still require manual review
• API pricing and service limits may change and should be rechecked before publication

Its Core Value Goes Beyond One-Off Text-to-Image Generation

Nano Banana 2.1 combines text, image, video, and PDF inputs with multi-turn editing, multi-reference composition, character consistency, and high-resolution output in one workflow.

For quick previews, start with 1K and the default medium thinking level. For complex multi-image composition, character consistency, or commercial assets, gradually add references, increase the thinking level where necessary, and move to 2K or 4K only after confirming the composition.