What’s new
- The fastest, leanest tier of the family: launched 2026-06-30, Google positions it as "the fastest, most efficient Gemini Image model"; its official developer name is Gemini 3.1 Flash-Lite Image.
- As fast as 4 seconds per image: Google says latency is dramatically reduced compared to the main Nano Banana 2 tier.
- Fewer controls than the main tier: only an aspect-ratio parameter is exposed — no resolution or output-format options — so it's simpler to use but leaves less to fine-tune.
- Up to 10 reference images per call: fewer than the main Nano Banana 2's 14, but it keeps core capabilities like character consistency and real-world knowledge.
What it’s good for
- Producing lots of variations quickly — A/B-testing ad creative or generating a batch of rough directions.
- Real-time generation features serving many users at once, like in-app generation inside a social app, where speed matters most.
- E-commerce virtual try-ons or storyboard sketches — cases that need speed more than pixel-level polish.
Prompting tips
- Cover all five elements up front: style, subject, setting, action and composition, written out clearly to cut down on back-and-forth revisions.
- Use it to explore directions fast, and save fine detail for a heavier model: Lite has no resolution or format controls, so it's a good fit for generating several directions before picking one to polish with a more capable model.
- Put on-image text in quotes: quote the words you want rendered and add a typography note, like "handwritten" or "bold sans-serif."
- Name your characters when using reference images: upload a reference and give the person a consistent name in the prompt so it recognizes them as the same character.
- Ask directly for multiple variations: e.g. "give me four different compositions," to get several options in one pass.
Product photography style, a mint-green wireless earbud charging case on a wooden desk, soft natural
light from a window on the right, shallow depth of field. Give me four different compositions: a flat
front shot, a 45-degree overhead angle, a close-up with the case open, and a shot held in a hand.
1:1, clean background.
Known limits
- Small faces, exact spelling and fine detail can still go wrong: the same issue applies as with other models in this family.
- Data-heavy visuals can get misread: when generating infographics or data visualizations, it can misinterpret or misrender the underlying information.
- Complex edits are less reliable: more involved editing operations can sometimes produce unnatural results, visible artifacts, or elements that don't quite fit together.
Using it in Nomi
On the generation canvas, add a image card, pick “Nano Banana 2 Lite” in the model picker and choose a mode (Text to image, Image editing). Add your references, write the prompt and generate. In the storyboard you can pick it per shot. The first time, connect any one of the providers above under Settings → Models.
