Nano Banana 2 Guide: Reasoning-Guided Images for Creators

Traditional diffusion models guess patterns. Nano Banana 2 adds a planning step: identify required text, validate spatial relationships, check diagram logic, then refine until the brief passes. That is why "OPEN" renders as four letters—not blob shapes—and why mirror scenes can keep reflection logic straight.

Prompt like you are instructing a careful illustrator: state explicit requirements ("text must read exactly 'SALE'"), limit simultaneous text elements (roughly three to five per frame), and put literal copy in quotes with placement cues (centered, above door, on towel embroidery).

Pair this guide with Nano Banana 2.1, the deeper briefs in Nano Banana Pro, and Nano Banana Pro prompting. Compare tradeoffs in GPT Image 2 vs Nano Banana 2.

Use quotes for every string that must appear on-image. Add typography family or use case (medical textbook illustration, professional diagram). For educational content, name each label and what it points to—leader lines should terminate on the correct structure.

Keep handwriting prompts short and medium-specific ("clean handwriting, blue marker") for whiteboard math. For flowcharts, enumerate boxes and arrows in order so the planner can validate sequence.

Nano Banana models excel at typography in layouts, product realism, and multi-step edits when you describe composition ratios (for example, two-thirds copy, one-third portrait). They also respond well to explicit lens language—35mm, 85mm product, isometric illustration.

Use these for text-to-image. For image-to-image, upload a reference still and append only the lines that should change (lighting, background, medium). Catalog shots get channel-specific wording in our product photography prompts and deeper Lock/Change syntax in the Nano Banana Pro prompting guide.

Complex arrangements—foreground signage, reflected poster text, props on counters—need explicit spatial roles (actual scene vs mirror reflection). Nano Banana 2 validates those roles before final render, which reduces swapped labels and duplicated text.

The model composes natively across aspect ratios instead of cropping a landscape frame. Treat 9:16 as vertical storytelling (posters, stories), 16:9 as hero banners, 1:1 as product tiles, and 21:9 as cinematic width. In UltraPix, pick ratio before generating so subject placement matches the destination feed.

Resolution strategy: draft at standard size, select the best composition, then regenerate at 2K/4K for print or marketplace zoom. PNG preserves sharp type; JPEG is fine for photo-only scenes without lettering.

Paste the prompts above into text-to-image, then vary one constraint at a time. When typography almost works, switch to image-to-image with a Lock clause instead of rewriting from scratch.

For photoreal habits (lens, materials, grade), read photorealistic AI image generation. For FLUX-style dense scenes, see the FLUX 2 MAX prompt guide. Run them in UltraPix.

Reasoning-guided models still fail when you overload one frame: ten labels, three paragraphs of poster copy, and a reflective floor will compete for planner attention. Split infographics into panels or generate a clean background plate first, then add type in a second image-to-image pass with a Lock on the plate.

Mirror and reflection prompts need explicit roles—what exists in the room vs what appears only in glass. If labels swap between reflection and reality, rewrite with "visible in mirror only" and "on actual wall" clauses like the bathroom example above.

For model strengths and more sample prompts, see Nano Banana 2.1.