This article tests the image editing capability of the new Gemini "Nano Banana" (aka Gemini Flash 2.5 Image) model.

Our Tests

1) First, let's create a test subject using a different model (OpenAI GPT Image 1).

[GPT Image 1]
Create a half-body portrait of a young east asian lady, in her 20s.
Straight, short brown hairs. Black eyes. Fair Skin.
with a natural smile.

2) Let's make a small change. 

[Nano Banana]
wink with finger showing V sign

• The character features are almost the same, though the hair color is slightly different, which is overall acceptable.
• It is also observed that the wrinkles on the clothes near the character's left arm became more obvious. This is natural because now the character has raised her hand.

3) Try to adjust the light source and shadow.

[Nano Banana]
move the light source to slightly lower than the current one

• It seems like nothing changes except that the head angle moves slightly.
• This could be due to bad instruction given.

4) Next, let's add something that does not exist in the original photo.

[Nano Banana]
print text logo calvin.my on the left side of the T-shirt. it must looks natural, a if it were part of the T-shirt deign. use cozy font.

• The result is very decent. The logo looks like it is sewn onto the shirt and follows the texture of the cloth.

5) Let's try to test the light source one more time, with more significant changes, e.g. adding a window.

[Nano Banana]
add a window at the left side in the background. A clear glass window, showing a noon with clear sky. please consider additional light source from the window on top of the existing indoor light source.

• This is much better then the earlier test. We can see the left side of the character has been significantly impacted by the new light source.

Despite being processed multiple times, we can still recognise the same character from the original image. This seems to be a big improvement and opens up a lot of new use cases.