Nano Banana Pro: the image editing model that turned "edit it by talking to it" into a product
nano-banana-pro
Gemini's native image generation and editing tier, known in the community as Nano Banana. Its point is not peak image quality but the interaction shape: keep editing one image across a multi-turn conversation, read the text inside an image and change it, and fuse several references into one output. That made image generation feel like talking to a designer instead of pulling a slot-machine lever. Most of its ranking comes from preference arenas, which are votes rather than measurements, so we grade it as needing reproduction.
- CONFIDENCE
- Needs Reproduction
- Single source, not yet verified by us
- MATURITY
- Product
- research → demo → product → production
Our take<p>We grade this D (needs reproduction), and that is clarification rather than criticism: its reputation comes mainly from <strong>preference arenas</strong> where many users vote between two images. Arenas answer "which one is more likeable" well; they do not answer "which one is more accurate against your spec". Design production needs the second answer - is the text right, are the product proportions right, did changing one thing move something else.</p><p>On interaction shape it is the most influential entry in this tier: continuous multi-turn editing, reading text inside an image, and multi-reference fusion together turned image generation from a one-shot output into an iterable work object. That shift is closer to the AGI direction than any single quality bump, because it means the model holds <strong>persistent state</strong> over an artefact instead of redrawing from scratch each time.</p><p>It is also closed and reachable only through the vendor's API: no self-hosting, no fine-tuning, no offline reproduction. For teams whose data cannot leave their perimeter, that is a hard boundary.</p>
Why conversational editing is a capability step
Plain text-to-image is stateless: a prompt goes in, an image comes out, and dissatisfaction means starting over. To make a second change to the same picture you have to feed the previous one back as a reference and hope nothing else moves. That path hits two walls quickly - identity drift (each edit shifts the face) and global perturbation (change one thing and the whole image's colour and detail are recomputed).
Conversational editing turns both into product requirements: the model must keep an understanding of the same artefact across a session and confine changes to what was asked. Prompt tricks cannot deliver that; it requires generation and understanding to close a loop inside one model, which is why this entry sits under both image and multimodal.
The three things that actually separate it
- Reading text inside the image before changing it: copy on a poster, package or UI is understood and then re-set, not guessed as texture. That makes "change the wording" a reliable operation for the first time.
- Multi-reference fusion: hand in a product shot, a scene and a style reference separately, and get one composite satisfying all three. That is the real shape of e-commerce and advertising workflow.
- Multi-turn continuity: the fifth edit still remembers what the first one settled. The longer the session, the closer the artefact gets to shippable.
Boundaries and failure modes
- Exact dimensions and proportions: it holds no engineering semantics, so "reproduce this package at true size" still returns something visually plausible and dimensionally wrong.
- Long-form layout: short headlines and labels are dependable; body-copy layout still belongs to a proper typesetting tool.
- Drift over long sessions: accumulated small changes wander from the original intent, so you need to periodically return to a clean version.
- No self-hosting: data must leave your perimeter, which rules out compliance-sensitive work outright.
- Nature of the evidence: existing rankings are mostly preference votes, with little third-party measurement judged against a spec. That is exactly why this is graded D and what our reproduction work has to fill in.
Our reproduction protocol (the measurements we will publish)
We plan to test it on three fixed task families rather than keep quoting arena position. First, text fidelity: take a poster with mixed Chinese and English copy, change one line, and measure the glyph error rate on everything else. Second, blast radius: edit only the specified region and measure pixel change outside it. Third, identity retention: after five consecutive edits of one subject, measure recognisable consistency against the first image. Until those three run, this entry does not move up a grade.