Nano Banana 2 on Lucidpic: Capabilities, Prompts, and Tests

Nano Banana 2 is Google's name for the Gemini 3.1 Flash Image model. It is available through Lucidpic for image generation and editing, but a new model name does not remove the need to test identity, text, hands, and product details.

What Google documents

Google's current Gemini image-generation guide describes Nano Banana 2 as the high-efficiency counterpart to Gemini 3 Pro Image, designed to balance performance, cost, and latency. The same documentation covers text-and-image input, iterative editing, supported resolutions and aspect ratios, and states that generated images include SynthID.

That first-party documentation is the right place to check current limits because model capabilities and identifiers change. It is more reliable than static comparison tables with unrepeatable speed or “accuracy” percentages.

Where it fits in a Lucidpic workflow

Nano Banana 2 is useful when you want to:

• Generate a new portrait or full-body scene from a detailed brief.
• Edit an existing image using natural-language instructions.
• Explore several compositions before selecting a campaign direction.
• Create source stills for image to video.
• Iterate on a character while retaining approved reference material.

For a reusable identity, start with the AI person generator or consistent character generator. A capable base model helps, but consistency also depends on the references, prompt structure, and how strictly you reject drift.

A prompt pattern that travels well

Use complete instructions rather than a pile of adjectives:

Create a vertical 4:5 editorial photograph of an adult man in his early thirties wearing a charcoal overshirt and cream T-shirt. He is standing beside a large north-facing studio window after rain. Use soft daylight from camera left, a relaxed expression, realistic skin texture, and a 50mm photographic perspective. Keep the background quiet and leave clear space above his right shoulder for layout. No text or logos.

The prompt specifies subject, wardrobe, scene, light, expression, camera language, composition, and exclusions. If the first result misses something important, edit one instruction at a time.

For reference edits, state both the change and the invariants:

Change the background to a restrained hotel lobby with warm evening light. Preserve the person's face, hairstyle, pose, jacket, camera angle, and crop. Do not add text, jewellery, or other people.

Test claims instead of repeating them

Use a small benchmark before adopting any image model as your default:

1. Run the same portrait prompt three times.
2. Run one full-body prompt with visible hands and footwear.
3. Edit a reference image while changing only the background.
4. Ask for a short, known phrase on a simple sign.
5. Save every output, not only the best one.

Score prompt adherence, face and skin, anatomy, text, reference preservation, and time to an acceptable result. Record the date, model version, aspect ratio, quality setting, and any automatic prompt processing. The broader photorealistic model comparison guide contains a reusable evaluation sheet.

Character consistency is an editorial process

When creating a series, lock the identity description and approved reference images. Keep temporary variables such as pose, location, and clothing in a separate shot brief. Compare outputs on a contact sheet so changes in jaw, eye spacing, age, hairline, complexion, and build are easy to spot.

Do not rescue every near miss. Regenerating from a sound reference is often faster than repairing an image whose identity has already drifted.

Text and product details need manual checks

Even when a model can render text, inspect every character at full size. For commercial images, compare labels, packaging shape, colour, materials, and included accessories against the approved product source. Treat a changed label or invented feature as a failed production image, not a harmless flourish.

Google's own responsive display ad guidance recommends high-quality, in-focus images and warns against excessive overlays and misleading button graphics. It also lists current landscape, square, and vertical asset specifications.

Provenance, consent, and disclosure

Google says its generated images include SynthID, but no watermark or provenance signal replaces honest disclosure where context requires it. The C2PA explainer makes a useful distinction: provenance can document an asset's origin and edits, but it does not prove that the content itself is true.

Use real-person references only with permission. Do not create misleading depictions of identifiable people. For sponsored creator content, the FTC's disclosure guide says the commercial relationship should be clear and hard to miss.

Try it on Lucidpic

Browse the Nano Banana prompt collection, start a new image in Studio, and keep a record of the prompt and approved result. The strongest workflow is not “generate once and trust the output.” It is brief, generate, compare, correct, and approve.