Why do you see 2 temporary images during Nano Banana Pro API calls? Complete analysis of the official thinking process

When calling the Nano Banana Pro API to generate images, have you noticed that two temporary images appear during the process instead of just getting the final result? This isn't an error—it's actually the "Thinking Process" feature of the Gemini 3 Pro Image model at work. In this post, we'll dive deep into the technical principles and practical value of this mechanism.

Core Value: By the end of this article, you'll understand how the Nano Banana Pro API's reasoning workflow works, learn how to view and use temporary images to optimize your prompts, and master the role of the Thought Signature in multi-turn conversations.

The Core Reason Behind Temporary Images in Nano Banana Pro API

The temporary image phenomenon in the Nano Banana Pro API stems from the reasoning mode design of Gemini 3 Pro Image. This model uses a multi-step reasoning strategy to handle complex image generation tasks rather than outputting a result all at once.

Official Documentation Clarification

According to the Google AI official documentation, this behavior of the Nano Banana Pro API is by design:

The Gemini 3 Pro Image preview model is a thinking model that uses a reasoning process ("thinking") to handle complex prompts. This feature is enabled by default and cannot be disabled in the API. The model will generate up to two temporary images to test composition and logic. The last image in the "thinking" process is also the final rendered image.

This means that when you call the Nano Banana Pro model via the APIYI (apiyi.com) platform, those two temporary images you see are evidence of the model actively performing quality validation, not a system glitch.

Technical Principles of Nano Banana Pro's Thought Process

How the Reasoning Workflow Works

The thought process of the Nano Banana Pro API follows this technical path:

1. Prompt Parsing Phase: The model first analyzes the text prompt entered by the user, identifying key elements, style requirements, and composition logic.
2. Initial Composition Test: It generates a 1st temporary image to verify the basic layout and the feasibility of the main elements.
3. Logical Optimization Iteration: Based on the results of the 1st image, it adjusts details and generates a 2nd temporary image.
4. Final Rendering Output: Using the experience gained from the first two tests, it generates a high-quality final image (which is usually the same as the 2nd temporary image or an optimized version of it).

Why Do We Need Temporary Image Testing?

The core value of the temporary image generation mechanism is reducing the failure rate of complex prompts. Traditional image generation models usually output a result in one shot; if the model misunderstands the request, the user has to manually adjust the prompt. In contrast, Nano Banana Pro uses an internal testing mechanism to self-correct before providing the final output.

💡 Technical Tip: For real-world development, we recommend using the APIYI (apiyi.com) platform for API calls. It provides a unified interface supporting Nano Banana Pro, DALL-E 3, Stable Diffusion, and other leading image generation models, helping you quickly verify technical feasibility and compare reasoning efficiency.

How to View Nano Banana Pro's Thought Process

Accessing Reasoning Details via Python API

The Nano Banana Pro API allows developers to retrieve the model's thought process and temporary images. Here’s a minimalist implementation:

import google.generativeai as genai

# Configure API key and base URL
genai.configure(
    api_key="YOUR_API_KEY",
    client_options={"api_endpoint": "https://vip.apiyi.com"}
)

# Call the Nano Banana Pro model
model = genai.GenerativeModel("gemini-3-pro-image-preview")
response = model.generate_content("A cyberpunk-style cat wearing sunglasses")

# Iterate through response parts to extract the thought process
for part in response.parts:
    if part.thought:  # Check if it contains thought content
        if part.text:
            print(f"Thought text: {part.text}")
        elif image := part.as_image():
            image.show()  # Display temporary image

import google.generativeai as genai
import json

genai.configure(
    api_key="YOUR_API_KEY",
    client_options={"api_endpoint": "https://vip.apiyi.com"}
)

model = genai.GenerativeModel("gemini-3-pro-image-preview")
response = model.generate_content("A cyberpunk-style cat wearing sunglasses")

# Store thought signatures for subsequent editing
thought_signatures = []

for part in response.parts:
    if part.thought:
        if part.text:
            print(f"Thought text: {part.text}")
        elif image := part.as_image():
            image.show()

        # Save the thought signature
        if hasattr(part, 'thought_signature'):
            thought_signatures.append(part.thought_signature)

# Save signatures to a file for multi-turn dialogue editing
with open("thought_signatures.json", "w") as f:
    json.dump(thought_signatures, f)

print(f"Captured {len(thought_signatures)} thought signatures")

🚀 Quick Start: We recommend using the APIYI (apiyi.com) platform to quickly build prototypes. It offers out-of-the-box API access with no complex configuration needed, letting you integrate and view the full thought process output in just 5 minutes.

Example of Thought Content Output

When you access response.parts, you might see a data structure similar to this:

The Role of Thought Signatures in Nano Banana Pro Multi-turn Dialogues

What is a Thought Signature?

A Thought Signature is an encrypted representation of the reasoning process that the Nano Banana Pro API has mandated starting with the Gemini 3 series. It records the model's internal logic on how it interprets the original prompt to generate an image.

Why it Matters for Multi-turn Editing

The role of the Thought Signature becomes crucial when you're editing images or engaging in multi-turn generations:

API Mandatory Validation Mechanism

According to official documentation, starting with Gemini 3 Pro Image, the API performs strict validation on all model response parts. Missing the Thought Signature will result in a 400 error:

Error 400: Missing thought signature in model parts

This means when using the Nano Banana Pro API for multi-turn dialogues or image editing, you must:

1. Save the thought_signature returned from the first generation.
2. Pass that signature back through the specific parameter in subsequent requests.
3. Ensure the signature format remains intact—don't try to modify it manually.

💰 Cost Optimization: For projects requiring frequent iterative edits, consider using the APIYI platform to call the API. They offer flexible billing and more competitive pricing, which is great for small-to-medium teams and individual developers doing multi-turn testing.

Cost Calculation for Nano Banana Pro Temporary Images

Are temporary images billed?

According to Google Cloud's official pricing docs, temporary images are not billed. You only pay for the final generated image.

Cost Comparison with Other Image Models

Even though Nano Banana Pro performs two additional internal image generation tests, the actual costs remain on par with or even lower than traditional models since those temporary images are free (and you'll have fewer failed retries):

🎯 Selection Advice: Choosing the right model mostly depends on your specific use case and quality requirements. We suggest running real-world tests via the APIYI platform to find what fits your needs best. The platform supports unified interface calls for various mainstream Large Language Models, making it easy to compare costs and results quickly.

FAQ

Summary

The 2 temporary images you see when calling the Nano Banana Pro API are the thought process feature of the Gemini 3 Pro Image model at work, not a system error. Here are the core takeaways:

1. Inference Mechanism: The model tests composition and logic by generating up to 2 temporary images; the last one is the final rendered result.
2. Cost Calculation: Temporary images aren't charged; you only pay for the final image.
3. Thought Signature: Saving and passing the thought signature in multi-round conversations can significantly improve editing consistency and efficiency.
4. Cannot be Disabled: The thought process is a built-in feature and can't be disabled via API parameters.
5. Quality Advantage: This mechanism increases the success rate for complex prompts from the 60-70% seen in traditional models to 85-90%.

We recommend using APIYI apiyi.com to quickly verify the thought process effects of Nano Banana Pro and perform real-world comparison tests with other image generation models.

Author: Tech Team
Technical Exchange: Visit APIYI apiyi.com for more technical documentation and best practice cases on AI image generation APIs.

📚 References

1. Google AI Developers – Nano Banana Image Generation: Official API documentation

   • Link: ai.google.dev/gemini-api/docs/image-generation
   • Description: Detailed technical notes on the reasoning process mechanism.

2. Google Cloud – Gemini 3 Pro Image Documentation: Vertex AI platform documentation

   • Link: docs.cloud.google.com/vertex-ai/generative-ai/docs/models/gemini/3-pro-image
   • Description: Enterprise-level deployment and configuration guide.

3. Google Developers Blog – Gemini API Updates: Official blog

   • Link: developers.googleblog.com/new-gemini-api-updates-for-gemini-3/
   • Description: New features and best practices for the Gemini 3 series.

4. Medium – Testing Gemini 3 Pro Image: Community technical review

   • Link: medium.com/google-cloud/testing-gemini-3-pro-image-f585236ae411
   • Description: Real-world use cases and performance analysis.