What Nano Banana could keep intact during a Gemini edit
In early September 2025, someone trying to change a person's outfit gave Nano Banana a clothing reference as a JPG. They liked the first two results. Then the tool cropped the photograph without being asked and returned the crop when prompted for more changes. Another user in the same thread reported that opening a new Gemini chat did not fix their own repeated-image problem. An edit could be convincing on its own and still leave its author unable to carry the image through the next revision. September 4 clothing-swap thread
What changed on August 26
Gemini already offered image generation with Imagen 4, which Google introduced in May with an emphasis on spelling and typography. Native editing also predated the August update. On August 26 Google introduced Gemini 2.5 Flash Image, nicknamed Nano Banana, in the Gemini app and highlighted a different task, editing an existing image while retaining a person's or pet's likeness and details the user wanted to keep. Google advertised localized changes, reference-image blending, style transfer and edits across multiple turns. The same-day developer announcement covers access via Google AI Studio, Gemini API and Vertex AI. Those routes have distinct controls and deployment conditions. Nano Banana Pro and Nano Banana 2 were later releases with their own capabilities. Gemini Apps release notes · Gemini app announcement · Developer announcement · Nano Banana Pro · Nano Banana 2
An August 26 showcase drew praise and concerns in the same thread. The poster later said they had trouble recreating a consistent character from a face photo. One participant described receiving an unrelated image when asking Gemini to edit the last result, criticized the resolution and watermark, and later said the tool seemed to be working better. Another participant said they did not encounter the same problem. On August 31, an Android user reported an update-required message upon attempting to edit an uploaded image and said they could access those image generations in a browser. The two accounts describe failures in different parts of the workflow, from retaining the source image to accessing the resulting chat. Neither account establishes how common the failure was. Launch-day thread · Android app and browser report
What happened when testers revised a picture
In a hands-on report published September 13, Mitja Rutnik described editing a portrait from Lake Bled in Gemini. Several wardrobe and accessory changes worked to his satisfaction. When he asked for greener trees around the lake, the picture lost a castle and an island and gained a mountain. He reported that asking Gemini to revert worked. He observed more grain as the edits accumulated. Uploading the original again with one comprehensive brief produced a less grainy image in his comparison, with some instructions ignored. A separate attempt to blend two personal photos into a scene of the couple holding hands yielded a composite in which they did not hold hands. The same review shows a generated cat that he successfully revised across several turns. The tests document successes and failures on different inputs without controlled repeat counts or request metadata. The precise work dates are unreported. Rutnik's illustrated hands-on report
On August 31, a Reddit contributor posted a comparison with their own prior Photoshop composites. They found a prompt for simulated hood damage broadly reproduced the intended effect and reported changes to car details. In the second edit, Gemini exposed the chair behind a seated reporter but substituted the world map, even though the brief asked to keep everything else the same. The images and author replies document what one editor accepted and flagged. The exact client and model routing are unspecified, so this case informs the editing question without identifying behavior unique to the app. The author supplied informal timings without equivalent output criteria. Photoshop comparison and prompts
Google's September 12 examples show intended uses from the provider's own selection. In one, a white dress becomes a garment made of tennis balls. In another, a house photograph becomes a tropical concept with thatch, bamboo columns and dense planting. The house example shows how an uploaded photo can become a visual concept. Google supplied paired images and prompts, with the number of attempts and any subsequent design use unreported. Google's Gemini app examples
The early reports show that a small requested change could preserve much of an existing picture, and that an apparently successful edit could still alter the crop, the surroundings, or a recognizable detail. They also make likeness and preservation separate questions for a design review. A person can remain recognizable while the map, background or framing changes. The available cases establish concrete revision behavior and leave open how often it recurred, how it varied by client or account, and whether the outputs met a downstream production brief. Clothing-swap thread · Rutnik's hands-on report · Photoshop comparison