Nano Banana: Gemini 2.5 Flash Image Profile

Nano Banana: Gemini 2.5 Flash Image

Nano Banana originally named Google's Gemini 2.5 Flash Image model, launched publicly in August 2025 for image generation and conversational image editing. Google later expanded Nano Banana into a family brand and now describes the original model as the legacy pioneer of that series.

The name alone no longer identifies one current model. A reproducible evaluation should record gemini-2.5-flash-image when it means the original.

Original product identity

Google's August 2025 announcement emphasized maintaining a person or pet's likeness, combining photos, transferring styles, changing parts of an image, and continuing edits across conversational turns.

Those are launch claims. They establish what Google positioned the model to do, not independent performance on every subject, language, region, interface, or use.

Name origin

Google's later first-person account says the team needed a public codename for an anonymous LMArena submission after the Gemini 2.5 Flash Image technical name was already set.

Product manager Naina Raisinghani proposed Nano Banana by combining two personal nicknames, "Naina Banana" and "Nano." The model appeared under that name on LMArena in early August 2025. Google's Nano Banana origin account

The name was therefore public while the model's owner remained concealed. It was not a secret internal codename that later leaked.

Editing workflow

The original model accepted text, images, or combinations of both. The distinctive product idea was that a user could continue revising an image in a conversation rather than treat every output as a disconnected generation.

flowchart LR
    A["Source image and prompt"] --> B["Gemini 2.5 Flash Image"]
    B --> C["First edited output"]
    C --> D["Follow-up instruction"]
    D --> E["Revised output with prior context"]
    E --> F["Human visual, factual, and rights review"]

Multi-turn editing should be tested as a state-preservation problem. A successful instruction can still damage a face, product detail, background, label, pose, or earlier decision.

[[How to Test Multi-Turn AI Image Editing]] provides a controlled sequence. [[How to Evaluate Identity and Detail Preservation in AI Image Editing]] separates task success from preservation and critical errors.

Current family position

Google's current Gemini image-generation guide, reviewed July 27, 2026, uses Nano Banana for four API models.

The guide calls the original a legacy model and recommends transition to a newer option for many new API workloads. Current pricing, availability, performance, limits, and migration advice require a fresh check before implementation.

[[Nano Banana Pro Product Profile]] covers Gemini 3 Pro Image separately. The two profiles should not be collapsed because they refer to different models and different launch periods.

Watermark and provenance boundary

Google said launch-era images created or edited in the Gemini app included a visible watermark and an invisible SynthID watermark. Its current developer guide says generated images include SynthID.

[[SynthID Product Profile]] explains the tool-specific signal and its limits. A watermark does not establish whether an image is true, consensual, correctly labeled, unmodified, or safe to publish. It also does not replace the source, prompt, model identifier, rights record, disclosure, and human review.

Evaluation boundaries

An image editor should be evaluated with owned or authorized sources and a defined use. The record should preserve the model identifier, date, interface or API, account conditions, region, exact prompt, source file, every output, latency, retries, safety response, and scoring method.

Product claims about likeness and local editing should be tested separately. A person can remain recognizable while critical details change. A requested transformation can succeed while unrelated regions drift.

Current Google documentation tells users to hold the necessary rights to images they upload and not create material that deceives, harasses, or harms. Terms, data handling, retention, review, and regional availability can change. Check the product and policy in use, not only this profile.

E082 evidence boundary

Venture Step E082 records Dalton Anderson testing the original launch-era experience. He thought a running figurine remained recognizable despite odd eyes, thought a felt version of the family Shar Pei preserved the dog well, saw several style conversions fail, and was impressed by a multi-turn clothing sequence.

The source images, outputs, interface captures, exact model display, audio, video, and public episode record are missing. The transcript supports Dalton's observations. It does not support retrospective visual scoring, a current model comparison, or republication of the images.

The episode also described the name as a leaked codename and referred to Nano Banana as having dethroned ChatGPT. Google's later origin account corrects the name history. No ranking record recovered with E082 supports the broader competitive claim.

This profile reflects official Google sources reviewed on July 27, 2026. AI assistance was used for research organization, drafting, and validation. Publication remains unauthorized.