Nano Banana runs on Gemini 3.1 Flash Image. It's fast, it handles edits, and it scales from 512px drafts to 4K finals without switching tools. The catch: vague prompts get vague results. This article gives you the prompt structure that actually works, plus a few variations to keep in your back pocket.
The Prompt
Use this template every time you want a new image. Fill in each slot. Leave nothing blank — the model fills blanks with defaults you probably don't want.
Create an image of: [subject]. Style: [style]. Composition: [camera angle or shot type]. Lighting: [lighting description]. Background: [background description]. Color palette: [color palette]. Avoid: [what to exclude].Example filled in for a coach's social content:
Create an image of: a confident woman in business casual attire sitting at a minimalist desk with a laptop. Style: editorial photography, clean and modern. Composition: three-quarter shot, slight low angle. Lighting: soft natural window light from the left, warm shadows. Background: blurred white office with subtle plants. Color palette: warm whites, sage green, muted gold. Avoid: cluttered desk, visible brand logos, harsh shadows, stock photo feel.Why Each Field Matters
- Subject: This is the only non-optional field. Everything else shapes it. Be specific — "woman at desk" is different from "confident coach at standing desk mid-presentation."
- Style: Tells the model the visual language. Options like "editorial photography," "flat design illustration," "cinematic still," or "watercolor" dramatically change the output.
- Composition: Controls where the camera sits. Use terms like "bird's eye view," "close-up portrait," "wide environmental shot," or "Dutch angle" — the model understands them.
- Lighting: This one carries more weight than most people expect. "Soft natural light" and "dramatic side lighting" are completely different images even if everything else is identical.
- Background: Vague backgrounds become noise. Say exactly what you want — or say "solid white" or "blurred bokeh" to keep focus on the subject.
- Color palette: 3-4 named colors or tones is enough. This locks the mood and keeps images consistent across a content series.
- Avoid: Don't skip this. Calling out what you don't want — watermarks, cluttered backgrounds, multiple people, specific colors — saves you iterations.
The Editing Prompt
When you're editing an existing image, the goal is surgical precision. The model needs to know what to change and what to leave alone. If you don't specify, it will reinterpret the whole image.
Change ONLY: [single specific change]. Keep identical: subject, composition/crop, pose, lighting, color palette, background, text, and overall style. Do not add new objects. If text exists, keep it unchanged.Example for swapping a background:
Change ONLY: replace the white office background with a blurred outdoor cafe scene, warm afternoon light. Keep identical: subject, composition/crop, pose, lighting on the subject, color palette of clothing, and overall editorial photography style. Do not add new objects or people. If text exists, keep it unchanged.Draft First, Then Go 4K
Do not generate at 4K until your prompt is locked. Run at 0.5K or 1K first. The composition, lighting, and style will all read clearly at low resolution — and you'll iterate 3-5x faster. Only go 4K on the version you're keeping.
Workflow: start at 0.5K, iterate until the image looks right, then rerun the same prompt at 4K. One resolution change at the end, not multiple expensive runs while you're still adjusting.
Prompt Variations
Here are three pre-filled variations for common content types. Adjust the subject and palette to match your brand.
Social post graphic (square, brand-forward):
Create an image of: bold typographic layout with the text "The Client Machine" centered. Style: modern graphic design, high contrast. Composition: centered symmetry, square format. Lighting: flat studio light, no shadows. Background: deep plum-black solid. Color palette: plum-black, violet accent, white text. Avoid: gradients, drop shadows, decorative borders, multiple font weights.Instagram Reel cover (vertical, 9:16):
Create an image of: a coach speaking on stage to a small audience, captured mid-gesture. Style: documentary photography, candid and warm. Composition: vertical portrait, subject fills upper two-thirds of frame. Lighting: warm stage spotlighting, slight rim light from behind. Background: blurred audience silhouettes, warm amber tones. Color palette: warm amber, soft ivory, deep shadow. Avoid: posed look, microphone in shot, identifiable faces in audience.Email header or course thumbnail (wide, 16:9):
Create an image of: a clean minimal workspace with an open notebook, coffee cup, and small succulent plant. Style: lifestyle product photography, aspirational and calm. Composition: top-down flat lay, centered. Lighting: soft diffused daylight, no harsh shadows. Background: warm white linen surface. Color palette: warm white, terracotta, sage green, matte gold. Avoid: phones, screens, cluttered items, visible brand labels.Tips for Consistent Results
- Save your working prompt in a text file. When you find a version that works, that file is your baseline for the whole series.
- Use the grounding flag (--grounding) only for prompts that reference real-world subjects that need visual accuracy — a specific landmark, animal species, or product. For abstract or generic content, skip it.
- Change one variable at a time when iterating. If the lighting is right but the composition is off, fix only the composition line. Changing multiple fields at once makes it impossible to know what fixed the problem.
- The Avoid field is not optional. Model defaults lean toward stock photography aesthetics — busy backgrounds, generic expressions, heavy shadows. Explicitly blocking those defaults gets you to a usable image faster.
- For a consistent brand look across multiple images, copy the Style, Lighting, and Color palette fields exactly from image to image. Only change Subject and Composition.