Nano Banana 2 — Instruction-Following AI Image Editing on ZenCreator

Why pick Nano Banana 2

What is Nano Banana 2?

Nano Banana 2 is Google Gemini's image-edit model on ZenCreator — exceptional at understanding reference images and following editing instructions. The defining differentiator is how references are handled: Nano Banana 2 accepts them inline as base64 in the request itself. It's the only model on the platform that does this, which means user image data never gets exposed via signed URLs to a third-party host. For client work where image privacy matters, the posture is the feature.

The model supports up to 4 reference images per call alongside the editing instruction, generates fast, and renders at 4K-class output. The strength is precision over free-form creativity — tell it exactly what to change, and it changes that without touching the rest of the scene.

The important caveat: Nano Banana 2 is the only censored model in the platform's image lineup. Google safety filters cannot be disabled at any level — any hint of nudity, suggestive content, or sensitive themes returns empty. There's also no seed control (output is non-reproducible across runs) and a hard cap of 4 reference images. For brand-safe commercial editing work this isn't a problem; for unrestricted creative work or seedable reproducibility, switch to a different model. Available in Text-to-Image and the Image Editor.

See Nano Banana 2 in action

Six prompts, six results. Copy any prompt to start from the same place.

Brand-accurate product

Hyper-detailed product photo of a vintage 1960s Leica M3 rangefinder camera with 50mm Summicron lens on a tan leather notebook. Engraved "Leitz Wetzlar" text on lens clearly legible. Slate-grey desk, brass-cornered notebook, Pelikan fountain pen at angle. Soft window light from upper-left, 100mm macro at f/4. Editorial product photography.

In-image typography

Editorial movie poster, minimalist composition. Bold uppercase sans-serif title "THE QUIET HOUR" in white on deep navy gradient. Sub-title "A FILM BY JANA MAREK". Lower two-thirds: silhouette of a woman walking down a long empty hotel corridor toward the vanishing point, warm amber light from one open doorway midway. Modernist film festival poster aesthetic.

Photoreal portrait

Editorial portrait of a young woman late twenties, freckled olive skin, dark wavy hair tucked behind one ear, soft natural makeup. Warm linen cream blouse. Seated on a Lisbon apartment windowsill holding a porcelain coffee cup, gentle half-smile. Soft golden afternoon light from camera left. 85mm at f/1.8. Warm cream highlights, soft amber mid-tones.

Web-grounded location

Architectural photograph of the Casa Batllo facade in Barcelona at blue hour. Undulating organic stone facade, mosaic dragon-scale roof tiles iridescent blue and green. Wrought-iron balconies with skull-like details. Warm amber street lamps along Passeig de Gracia. Long exposure on passing pedestrians. 24mm at f/8. Cool blue ambient, jewel-tone mosaic colors.

Editorial fashion

Fashion editorial of a tall model in a tailored cream wool double-breasted coat, beige cashmere turtleneck, wide-leg cream trousers, walking through a sunlit minimalist concrete corridor. Warm directional sunlight from overhead skylight casting long shadows. Single black leather tote. 85mm at f/2.5. Precise cream and warm taupe palette.

Editorial infographic

Illustrated infographic poster titled "HOW A DRIP COFFEE BREWS" in bold serif at the top. Central cross-section V60 dripper above a glass server. Five labeled callouts: "1. Bloom", "2. Pour", "3. Drawdown", "4. Aroma", "5. Cup" with thin connecting lines to image. Cream background, terracotta accents, dark chocolate line work. Mid-century modern editorial infographic.

Nano Banana 2 vs other ZenCreator models

When NOT to pick Nano Banana 2

Nano Banana 2 is a specialist. Three categories where another model is the cleaner choice:

• Any NSFW or edgy content — Google safety filters cannot be disabled at any level. Even mildly suggestive prompts return empty. Switch to Flux Klein NSFW for photoreal mature work, or SDXL NSFW for the unrestricted alternative.
• Reproducible output across runs — Nano Banana 2 has no seed control. If you need the same output twice, or are debugging a prompt, switch to a seedable model like WAN 2.7.
• More than 4 reference images — hard cap on reference count. WAN 2.7 Pro accepts a wider reference window for mood-board-driven scenes.

Get started in 4 steps

1. Open the Text-to-Image generator (or the Image Editor for reference-driven work).
2. Pick Nano Banana 2 in the model picker.
3. Write your prompt — be explicit. Nano Banana 2 follows instructions literally; vague briefs get vague output.
4. For Image Editor, attach up to 4 references with clear role descriptions ("subject reference", "color palette reference", "outfit style"). Hit Generate.

How to write prompts that land on Nano Banana 2

Nano Banana 2 rewards specificity. Five tactics built around its differentiators:

1. Be explicit about every visual decision. Nano Banana 2 doesn't infer — it executes. Spell out the subject, the action, the camera angle, the light direction, the colour palette. Words you skip get filled with average values from training. Words you write get rendered exactly.

2. Name real-world subjects accurately. Web grounding pulls from Google's index — naming "Casa Batlló in Barcelona" or "Patagonia retro-pile jacket" or "1960s Leica M3" produces factually accurate renderings. Generic descriptors ("a fancy building", "a vintage camera") trigger generative invention instead of grounded retrieval.

3. For references, give each one a role. In the Image Editor, attach up to 4 references and label them in prose: Reference 1 = subject's face, Reference 2 = outfit style, Reference 3 = colour palette and mood, Reference 4 = setting and architecture. Nano Banana 2 routes signal from the labelled reference into the corresponding output slot.

4. Quote in-image text and name the language. Wrap text content in quotes (title reading "THE QUIET HOUR" in white sans-serif). For non-Latin scripts, write the actual characters. Nano Banana 2's typography rendering is the strongest on the platform — exploit it with explicit spec.

5. Stay clearly on the SFW side. Safety filters fire on borderline language as well as explicit prompts. Words like "intimate", "sensual", "barely-dressed" can trigger empty returns. For brand-safe commercial work this isn't a problem; if your prompt feels even mildly edgy, switch to a different model rather than fight the filter.

What to avoid: vague mood-led prompts ("cinematic, beautiful, atmospheric"), implicit instructions ("make it more dramatic"), seed-dependent workflows, anything beyond 4 references.

Bottom line

Nano Banana 2 is the answer when you need precision: literal instructions, real-world subjects with factual accuracy, in-image text at design quality, and brand-safe output ready for client delivery. The trade-off is rigid safety filters and no seed control — pay attention to those before committing to a Nano Banana workflow for sensitive content or repeat-output work. For editing-heavy or reference-driven briefs that stay safely in SFW territory, it's the strongest model on the platform.

Available in

Nano Banana 2 powers two image tools on ZenCreator. Pick the entry point that fits your input.

Questions

Why is Nano Banana 2 the only censored model on ZenCreator?

It's the only model on the platform built directly on Google's Gemini API. Google enforces safety filters at the API level, and there is no off switch — not for trusted users, not via flags. The other models on ZenCreator either run on private deployments or use providers that allow filter disabling.

Can I disable the safety filters?

No. Any prompt with a hint of nudity, suggestive content, or sensitive themes returns empty. If you need unrestricted output, switch to Flux Klein NSFW, SDXL NSFW, or the photoreal NSFW-capable WAN 2.7 family.

What is "web grounding"?

When Nano Banana 2 sees a real-world subject in your prompt — a specific place, product, brand, or person — it pulls factual references from Google's knowledge base instead of inventing them. "Casa Batlló in Barcelona" produces an accurate Casa Batlló; "a famous Barcelona building" produces a generic Gaudí-influenced fabrication.

How many reference images can I attach?

Up to 4. They're sent inline as base64 inside the request, so your images never get exposed via signed URLs to a third-party host. For mood-board-driven scenes that need more than 4 references, WAN 2.7 Pro accepts a wider reference window.

What's SynthID and why does it matter?

SynthID is Google's invisible watermark embedded into every Nano Banana 2 output. It survives compression and minor edits. For brand-safety teams that need a chain of custody on AI-generated assets, it's a meaningful artefact — you can verify which images came from Nano Banana 2 even after they've been processed downstream.

Can I use Nano Banana 2 for fast batch generation?

Yes — speed is one of its strengths. The model is optimised for low-latency, high-volume generation. For 50-image bulk runs of brand-safe content, it's one of the fastest options on the platform.

When should I pick Nano Banana 2 over Seedream 5?

Pick Nano Banana 2 when your work needs literal instruction-following, in-image text, real-world brand accuracy, or reference editing. Pick Seedream 5 when you need fast cinematic generation, uncensored output, or rich photoreal color. The two models solve different problems.

Sources

1. Google DeepMind — Gemini Flash Image (Nano Banana 2): deepmind.google/models/gemini-image/flash
2. Google AI — SynthID watermark documentation
3. ZenCreator AI Models Review (internal) — Nano Banana 2 strengths and weaknesses
4. Internal benchmark comparisons across Nano Banana 2, Seedream 5, and WAN 2.7 — ZenCreator testing, May 2026