Testing Nano Banana: Gemini 2.5 Flash Image Editing

Gemini 2.5 Flash Image, nicknamed Nano Banana, is an image generation and editing model that I find particularly strong at preserving a character across edits. I test it on viewpoint changes, composition, image restoration, text, object counting and image sequences, while pointing out places where it fails.

I ask it to flip over two phones in an MKBHD thumbnail. The generated screens resemble iPhone and Android interfaces, although some icons and letters are wrong. This is a convincing imagined reverse view, not a way to recover details actually hidden in the original image.

In another test I ask it to rotate the two rightmost KAWS figures. It rotates three and changes one position, but a follow-up instruction improves the result. With a Coca-Cola can, it produces three viewpoints while preserving the logo and condensation convincingly.

I add the can to a photograph of myself, then add reflective glasses. The glasses reflect yellow flowers, matching the surrounding scene. A further request generates a plausible rear view of my head. These edits show how the model can invent coherent details, even when those details were not visible in the source.

I ask for several poses and angles of two characters. The model maintains their appearance across the images, though it produces fewer images than requested and has a little clipping. A group image includes two front views and a rear view.

For a composition test, I edit a moon-landing photograph to add camera operators, then expand it into an imagined soundstage. The generated additions match the original’s grain and lighting while preserving the astronaut. This is a fictional edit, not evidence about the actual moon landing.

At launch, Google’s announcement and the displayed image-edit leaderboard rank the preview highly. I then transform my portrait into a pilot in front of an SR-71 and ask to zoom out. The model retains facial details across edits and adds a photographer-like reflection to the helmet.

In a group photograph I remove two people and insert myself. The generated pose, expression and shadows fit the scene surprisingly well, although the result is a synthetic alteration. I also try adding a beard, changing hair and editing another person’s expression; generated tears look artificial, so I remove them.

A four-panel comic mostly keeps the same character across breakfast, a meeting and home, but omits her from the commute panel. A cat appears first in the background and later on her lap. The example shows both continuity and a missed instruction.

Background removal works well on a portrait, but a subsequent muscular edit leaves an inconsistent sleeve on a hand. I also turn a portrait into anime, then a 3D-style character, a full-body image and a sheet of different poses. The sequence suggests uses in character design and product imagery.

For generation from scratch, I ask for an arbitrary everyday moment. One output has effective shadows but a stylized lens flare; another includes a floating leash and a very large pigeon. I like the depth-of-field effect, even as these small inconsistencies remain visible.

A car prompt tests reflections: the generated grille and wet ground show reflected elements, including a photographer. In thumbnail experiments, I remove the background, change facial expressions and request a challenge-style composition. One attempted face replacement does not follow my request well.

Changing thumbnail text from “big news” to “huge news” preserves the original style reasonably well. I also repair damaged black-and-white photographs and colorize them. The results look convincing, but a plausible generated color or reconstructed detail is not necessarily historically accurate.

At the time of the video, I demonstrate access through Google AI Studio and the Gemini interface. I discuss settings such as temperature and show where the image generation option appears. Availability, names and limits can change after the demonstration.

A sequence prompt shows a candle unlit, burning and melted. The three panels mostly follow the timeline, though the burning candle appears taller than the unlit one. A second prompt requests a burger decaying over years and generates several stages of mold and decay.

For material changes, I ask the model to replace an ice teapot with a metal one while keeping the rest of the scene. Much of the composition, including the steam, remains similar. I also create a meme by adding a banana and text to a whiteboard.

A counting test asks for seven identical apples. There are seven, but two lack leaves that appear on the others, so they are not truly identical. A final close-up of interlocking hands has five fingers and convincing skin texture in this example. The successes are impressive, but individual outputs still need inspection.