How much better is Google's image generation AI 'Gemini 3 Pro Image (Nano Banana Pro)' than previous models and other models?
Google released its high-quality image generation AI,
GenAI Image Showdown
https://genai-showdown.specr.net/image-editing
Nano Banana Pro is an image generation AI built on Gemini 3 Pro , which can generate new images or edit existing images according to user instructions. Examples of Nano Banana Pro are summarized in the following article.
Google's image generation AI 'Gemini 3 Pro Image (Nano Banana Pro)' has been released, applying Gemini's inference function to enhance language support and information organization capabilities - GIGAZINE
The top image of this article was also edited using Nano Banana Pro. Japanese text, which contains a mixture of hiragana, katakana, and the alphabet, was rendered correctly. The line breaks were also positioned appropriately.
The GenAI Image Editing Showdown is a website that tests various image-generating AIs by inputting the same image and editing instructions, and the results of Nano Banana Pro have been added.
First, input an image of stacked blocks of different colors and instruct the user to 'swap the blue and yellow blocks.'
Below are the results for 'Gemini 2.5 Flash (Nano Banana)', 'Gemini 3.0 Pro Image (Nano Banana Pro)', 'FLUX.1 Kontext [dev]', 'OpenAI gpt-omage-1', 'Qwen-Image-Edit', and 'Seedream 4'. Only Nano Banana Pro was able to edit correctly.
'Remove only the brown candy from the bowl,' he instructed.
The Nano Banana Pro and Qwen-Image-Edit were successful in this test, with the older Nano Banana changing the color of the candy entirely.
The image below will be edited to 'Remove all trash from the road. Replace people sleeping on the ground with green benches. Replace parking meters with plants.'
With the Nano Banana, there was a lot of trash on the road, but with the Nano Banana Pro, it was clean.
Input a photo of the Leaning Tower of Pisa and instruct it to 'make the tower vertical because it is leaning to the right.'
Nano Banana Pro produced images that were nearly identical to the original. FLUX.1 Kontext [dev] and Seedream 4 were also successful.
While inputting a picture of a giraffe, instruct the user to 'make the giraffe's neck extremely short.'
Although Nano Banana Pro shortened the giraffe's neck, it looked like the upper part of the neck had simply been removed and the head moved, which GenAI Image Editing Showdown deemed a failure.
Of the 12 tests, GenAI Image Editing Showdown passed 10 of them, making it the AI model with the highest success rate.
Related Posts:
