Nano Banana Pro: The Google Image Editor That Fixes Your Background Without Touching Your Product

Nano Banana Pro is Google's Gemini 3 Pro Image model — the image editing and revision specialist in AnyCap's catalog. At approximately 7 credits per call, it's the right choice when an agent already has a draft image and needs targeted, prompt-directed changes rather than a fresh generation from scratch.

What Is Nano Banana Pro?

Nano Banana Pro is Google's Gemini 3 Pro Image model, released November 2025. Where most image generation models work best when starting from a blank text prompt, Nano Banana Pro is purpose-built for editing existing images. Pass a reference image and an edit instruction, and it applies the described changes while preserving the overall composition, product shape, or subject consistency.

Through AnyCap, Nano Banana Pro fits into the same capability runtime as Seedream 5, FLUX.1 Kontext Max, Nano Banana 2, and other image models — one CLI, one auth, one credit system for the full catalog.

Nano Banana Pro at a Glance

Why Agents Choose Nano Banana Pro

1. The specialist for image editing and revision

Nano Banana Pro's strength is targeted editing: change the background, adjust lighting, swap an element, refine a composition — all with a text prompt applied to an existing image. This is distinct from general image generation models that tend to reinterpret the entire scene when given an edit instruction.

2. Agent revision loops without designer intervention

In automated content pipelines, feedback often arrives as text: "make the background darker," "center the product," "remove the text in the corner." Nano Banana Pro translates these instructions into precise image edits, enabling fully automated revision cycles without human design intervention.

3. Preserves subject and composition during edits

A recurring limitation with general image models used for editing is drift — the model changes things you didn't ask it to change. Nano Banana Pro is specifically engineered to apply the requested change while preserving subject identity, product shape, and structural composition.

4. Sequential refinement across multiple passes

Because each Nano Banana Pro edit returns an image that can immediately feed the next edit call, agents can chain multiple revision steps. Feedback → edit → feedback → edit — each step is precise, and the subject drifts less across passes than with general generation models.

Using Nano Banana Pro via AnyCap

Setup:

curl -fsSL https://anycap.ai/install.sh | sh
anycap auth login

Image-to-image edit (primary use case):

anycap image generate \
  --model nano-banana-pro \
  --mode image-to-image \
  --prompt "replace the cluttered desk background with a clean warm cream studio gradient, center the product, add soft rim lighting, keep the product shape and color exactly the same" \
  --param images="./product-draft.png" \
  -o product-revised.png

Inspect the schema:

anycap image models nano-banana-pro schema --operation generate

Live Demo: Nano Banana Pro Image Editing with AnyCap

Everything below was generated live using AnyCap. No Photoshop. No designer. An AI agent running CLI commands in sequence.

Step 0: Generate a Base Draft with Nano Banana 2

Start with a quick draft using Nano Banana 2 — the fast, cost-efficient model for first-pass generation:

anycap image generate \
  --model nano-banana-2 \
  --prompt "Product photo of a premium white wireless Bluetooth speaker on a cluttered home desk — coffee mug, notebook, pen, cables in background. Natural window light, realistic smartphone camera photo quality." \
  --param aspect_ratio="4:3" \
  -o product-draft.png

Generated with Nano Banana 2. Realistic draft with a distracting background — ready for editing.

Step 1: First Edit Pass — Background Swap and Studio Lighting

Pass the draft into Nano Banana Pro with a targeted edit instruction:

anycap image generate \
  --model nano-banana-pro \
  --mode image-to-image \
  --prompt "Replace the cluttered desk background with a clean warm cream studio gradient. Center the speaker perfectly. Add soft professional rim lighting from the upper left. Keep the speaker shape, color, and details exactly the same. Remove all clutter." \
  --param images=product-draft.png \
  --param aspect_ratio="4:3" \
  -o product-v1.png

Nano Banana Pro edit 1: cluttered background replaced with a studio gradient, professional rim lighting added. Speaker shape and color preserved without drift.

Step 2: Second Edit Pass — Shadow and Color Warmth Refinement

Feed the v1 output back into Nano Banana Pro for a second refinement. The v1 image becomes the new reference:

anycap image generate \
  --model nano-banana-pro \
  --mode image-to-image \
  --prompt "Add a soft drop shadow beneath the speaker. Slightly warm up the color temperature to make the product look more premium. Add a very subtle gradient from cream-white at top to light warm gray at the bottom. Keep all other elements exactly the same." \
  --param images=product-v1.png \
  --param aspect_ratio="4:3" \
  -o product-v2.png

Nano Banana Pro edit 2: drop shadow added, color temperature warmed. Each pass applies only what the prompt instructs — the subject doesn't drift across iterations.

Three CLI calls. Two targeted edits. The revision loop that took a designer 30 minutes is now fully automated.

Nano Banana Pro in an Agentic Workflow

An agent processing stakeholder feedback and applying it to product images without human designer involvement:

import subprocess

def apply_edit(source_image: str, edit_instruction: str, output_path: str) -> str:
    """Apply a targeted edit to an existing image using Nano Banana Pro."""
    subprocess.run([
        "anycap", "image", "generate",
        "--model", "nano-banana-pro",
        "--mode", "image-to-image",
        "--prompt", edit_instruction,
        "--param", f'images=./{source_image}',
        "-o", output_path
    ], check=True)
    return output_path

# Feedback from review: three edit requests, applied sequentially
feedback_queue = [
    ("v1.png", "make the background lighter, shift to off-white", "v2.png"),
    ("v2.png", "add a soft drop shadow beneath the product", "v3.png"),
    ("v3.png", "slightly warm the color temperature, make the product look more premium", "v4.png"),
]

for source, instruction, output in feedback_queue:
    result = apply_edit(source, instruction, output)
    print(f"Edit applied: {output}")

print("All feedback processed — v4.png ready for final review")

This pattern replaces the design review and revision cycle with a fully automated agent loop. Feedback is text-in, revised image is out.

Nano Banana Pro vs Other Image Models in AnyCap

Nano Banana Pro vs FLUX.1 Kontext Max: Both are editing-oriented models. Nano Banana Pro is faster and works well for clear, single-instruction edits to commercial images. FLUX.1 Kontext Max produces richer visual output for design-demanding tasks but costs more. Use Nano Banana Pro for practical revision loops; use FLUX.1 Kontext Max when design quality at each edit step is the priority.

Nano Banana Pro vs Seedream 5: These models serve sequential steps, not competing ones. Use Seedream 5 to generate the draft, then use Nano Banana Pro to refine it. They're typically used in the same workflow — generation step followed by editing step.

What Nano Banana Pro Is Not Ideal For

• Starting from a text prompt with no reference image: Seedream 5 or GPT Image 2 produce stronger first-pass output from blank-slate prompts.
• High-volume batch generation: At ~7 credits per call, Nano Banana Pro is more expensive. For volume generation without targeted editing, Nano Banana 2 or Seedream 5 are more cost-efficient.
• Creative concept exploration without a source image: Nano Banana Pro is optimized for editing, not discovery. Use it after a direction is established.

Getting Started

# Install and authenticate
curl -fsSL https://anycap.ai/install.sh | sh
anycap auth login

# Apply your first edit with Nano Banana Pro
anycap image generate \
  --model nano-banana-pro \
  --mode image-to-image \
  --prompt "clean up the background, make it a smooth white gradient, keep the product" \
  --param images="./your-image.png" \
  -o refined.png

→ Nano Banana Pro model page → All image generation models → Image generation capability guide

FAQ

What is Nano Banana Pro best for?

Nano Banana Pro is best for targeted image editing and revision loops when an agent already has a draft image. It applies prompt-directed changes — background swaps, lighting adjustments, composition refinements — while preserving the subject and overall structure of the source image.

What is Nano Banana Pro based on?

Nano Banana Pro is Google's Gemini 3 Pro Image model, released November 2025. It's the Pro tier of the Nano Banana family, designed specifically for editing precision over generation speed. Through AnyCap, it's accessible alongside Nano Banana 2 (the speed/volume model) and the rest of the image catalog.

Should I use Nano Banana Pro or FLUX.1 Kontext Max for image editing?

Both models excel at editing existing images, but with different tradeoffs. Nano Banana Pro is faster and well-suited for practical single-instruction revisions. FLUX.1 Kontext Max produces richer design quality for complex, multi-element edits. Use Nano Banana Pro when speed and cost efficiency per edit matter; use FLUX.1 Kontext Max when each edit needs to meet a higher design bar.

Can Nano Banana Pro generate images from text without a reference image?

Yes. Nano Banana Pro supports text-to-image mode, but it's not its primary strength. For polished first-pass generation from a text prompt, Seedream 5 or GPT Image 2 are stronger. Nano Banana Pro is purpose-built for the editing and revision step that comes after.

How does the Nano Banana Pro revision loop work?

An agent generates a base image (with Seedream 5 or another model), then calls Nano Banana Pro with the draft as the reference image and an edit instruction as the prompt. The model returns the revised image. This revised image can be fed back into Nano Banana Pro for the next edit. Each pass applies only what the prompt instructs, preserving the rest.

Does Nano Banana Pro integrate with Claude Code?

Yes. The anycap image generate --model nano-banana-pro command is available in any shell-capable environment, including Claude Code. Pass the reference image path via --param images and the edit instruction as the --prompt. No additional Google API credentials required.