Nano Banana·de GoogleTexto a imagenNuevo
Nano Banana 2.1 Text-to-Image
Text-to-image with Google Nano Banana 2.1 (Gemini 3.6 Flash): Pro-level quality at Flash speed. 1K, 2K and 4K output, shapes as wide as 8:1, sharp multilingual text rendering, adjustable thinking and optional Google Search grounding.
Parámetros
| Parámetro | Tipo | Por defecto | Rango u opciones |
|---|---|---|---|
aspect ratio aspect_ratioShape of the image. auto keeps the input image's shape when editing and makes a square otherwise. Ultra-wide and ultra-tall shapes (4:1, 1:4, 8:1, 1:8) are supported. | select | auto | auto, 1:1, 3:2, 2:3, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9, 21:9, 4:1, 1:4, 8:1, 1:8 |
resolution resolutionOutput size: 1k ≈ 1 MP, 2k ≈ 4 MP, 4k ≈ 16 MP. 2k costs 1.5× and 4k 2.8× the 1k price. | select | 1k | 1K · ~1 MP (1k), 2K · ~4 MP (2k), 4K · ~16 MP (4k) |
enable image search enable_image_searchUse images found through Google Image Search as visual context. Photos of real people from the search are not used. Adds a small fee per request. | select | false | true, false |
enable web search enable_web_searchGround the image in live Google Search results (current events, weather, charts). Adds a small fee per request. | select | false | true, false |
thinking level thinking_levelHow much the model reasons before drawing. high helps with infographics, dense text and multi-character scenes but takes longer; minimal is fastest; default means medium. Does not change the price. | select | default | default, minimal, medium, high |
Entradas
Sin archivos de referencia: este modelo trabaja solo con el prompt.
Reglas de LUVI Prompt Engine para Nano Banana
TARGET MODEL: Nano Banana 2.1 (Google, gemini-nano-banana-2.1). It reads a brief the way a language model reads one: positively, in prose. There is no negative_prompt, no weighting and no style dial.
Reglas que deben cumplirse
- Narrative prose, not tag stacks. Describe a scene a person could photograph or design: register → subject and action → setting → light → camera → in-image text → finish.
- Exclusions become presences that do NOT name the excluded thing: "no people, no cars" → "a quiet, deserted street", never "no cars" and never "empty of people and cars". Naming the unwanted thing can summon it.
- Aspect ratio and resolution are parameters. Never write "16:9", "widescreen", "4K" or "8k" into the prose — on a conflict the parameter wins and the words are wasted.
- In-image text: put every visible string in double quotes, exactly as it must read, and describe the type (weight, style, colour) and its placement in plain words. Never write the words "in quotes" or "quoted" — the model draws the quotation marks. Every other visible surface (signs, apparel, packaging, screens) is either given quoted copy or described as plain/blank; unquoted incidental lettering comes back as gibberish.
- Keep quoted copy verbatim. If it contains an em dash, the model draws it as a separate rule that splits the line: add OUTSIDE the quotes that the text stays on a single line with a literal em dash between the phrases. Example: a sign reading "Taze simit — İçeride çay var" on a single line, with a literal em dash between the two phrases.
- Non-English copy (e.g. Turkish İ ı ğ ş ç) is kept character-exact and LARGE in the frame; small lettering drops diacritics at any resolution.
- People: keep every physical detail the user gave (age, build, hair, wardrobe, expression) — never flatten a described person to "a man"; vague unnamed people drift toward real public-figure lookalikes. Keep distinct on-screen identities few; give extras varied faces or defocus them.
- Never ask for a transparent background — the model has no alpha. Ask for a plain seamless white background instead.
Lo que mejor funciona
- Open with the operation and the deliverable: "Create an image: a photorealistic product photograph of …", "… an infographic poster …".
- State purpose and audience when known ("for an e-commerce product page", "suitable for a 4th grader").
- Many elements → order them in words: "First … Then … Finally …".
- When the user names light, lens or mood, phrase it concretely ("soft window light from the left" beats "cinematic lighting"). Do not add light directions, lenses, camera specs, scene elements (buildings, props, background figures) or wardrobe the user did not ask for — expand only what the user wrote.
- Prose over JSON: when the user supplies JSON, rewrite its content as prose — JSON loses counted details and soft details.
De las reglas de LUVI Prompt Engine para esta familia: las mismas que aplican el reescritor del espacio de trabajo y el conector MCP de LUVI.