Nano Banana 2 Negative Prompting: How to Tell the AI What NOT to Draw

Every AI image generator has a dirty secret: it’s much better at adding things than removing them. Ask for “a clean studio portrait” and you’ll get studio lighting, a subject, and also — surprise — a watermark logo, a blurry background smear, or three extra fingers. That’s the default state of generative AI in 2026, and Nano Banana 2 is no exception to the underlying problem. What it is an exception to is the workflow for handling it. Nano Banana 2 ships with a structured negative prompt field that sits alongside your main prompt, and it does exactly one job: it tells the model what to leave out.

This tutorial walks you through how that system actually works — not in theory, but in practice. You’ll see the exact phrasing that gets results, the combinations that protect brand safety, the resolution-tier tricks that matter, and the common negative prompt mistakes that actually make images worse. Copy the prompts, run them, iterate. That’s the whole game.

What You’ll Actually Achieve Here

By the end of this guide you’ll be able to write negative prompts that reliably strip watermarks, suppressed UI artifacts, stray text, unwanted faces, motion blur, and over-processed skin from your outputs. You’ll also understand why negative prompts interact differently with portrait versus product versus editorial styles — and how to adjust accordingly. The result is tighter output, fewer regeneration cycles, and images you can actually use without a cleanup pass in Photoshop.

Before You Start: Requirements

You need access to Nano Banana 2, Google’s image generation layer built on Gemini 3.1 Flash Image. You can reach it through the Gemini app (gemini.google.com), through AI Studio (aistudio.google.com), through the Gemini API directly, or through Vertex AI if you’re on a Google Cloud setup. The negative prompt field appears in all of these interfaces — though the exact UI label varies. In the Gemini app it’s labeled “Exclusions”; in AI Studio it surfaces as a second text box under the main prompt field; in the API it maps to the negative_prompt parameter in your generation config. Antigravity users get it as a dedicated input row in the visual editor.

Note 💡

The negative prompt field in Nano Banana 2 accepts plain English — no special syntax, no parentheses weighting like Stable Diffusion, no colons. Just comma-separated descriptors. “watermark, blurry, text overlay” works better than “(watermark:1.4)” — the model doesn’t parse weighted syntax the same way open-source diffusion models do.

Step 1: Understand What Negative Prompts Actually Do Here

Negative prompts in Nano Banana 2 work as exclusion signals during the generation process. The model reads both your positive prompt and your negative prompt simultaneously and adjusts its sampling away from the concepts flagged as unwanted. This is not post-processing — it happens at inference time, which means it genuinely affects composition, not just surface appearance.

The practical implication: negative prompts work best on things that are visually definable. “Blur” works. “Bad vibes” does not. “Text, watermark, logo” works. “Anything unprofessional” does not. Stay concrete, stay visual, and keep each entry short — two or three words maximum per exclusion item. Long-form descriptions in the negative field confuse the model more than they guide it.

Step 2: The Core Negative Prompt Stack

Most professional users in early 2026 are running some version of a “base stack” — a standard negative prompt block they paste into every generation before customizing. Here’s a solid starting point that covers the most common artifacts across all resolution tiers:

watermark, text overlay, logo, signature, username, blurry, out of focus, motion blur, low resolution, pixelated, jpeg artifacts, overexposed, underexposed, grainy, noise, oversaturated

This string handles the technical failures (blur, noise, compression artifacts) and the IP/brand safety flags (watermarks, logos, signatures) in one shot. Run this as your baseline for any commercial or editorial output. The model will occasionally still sneak in a subtle watermark-style texture — that’s when you add “watermark pattern, repeating texture” to the negative field and regenerate.

Pro tip ✅

Save your base negative stack as a text snippet in your clipboard manager or a notes app. Regenerating with the same negative block ensures consistency across a batch. Nano Banana 2’s batch mode reads the same negative prompt across all images in the queue, so your entire set stays clean without re-entering it each time.

Step 3: Portrait-Specific Negative Prompts

Faces are where Nano Banana 2 earns its reputation for clean output — and where a bad negative prompt can actually make things worse. If you over-suppress face-adjacent concepts, you start getting strange deformations. The goal is surgical exclusion, not carpet bombing.

extra limbs, extra fingers, deformed hands, distorted face, asymmetrical eyes, skin texture, overly smooth skin, plastic skin, airbrushed, unnatural lighting, harsh shadows, red eye

This stack targets the classic portrait failure modes without touching the underlying facial structure. “Skin texture” sounds counterintuitive to exclude — but in practice it pushes the model away from the uncanny pore-level detail that reads as “AI face” immediately. If you’re generating executive headshots or editorial portraits, add “background clutter, distracting background, busy background” to force cleaner separation between subject and environment.

extra limbs, deformed hands, distorted face, plastic skin, harsh shadows, background clutter, multiple people, crowd, strangers

The addition of “multiple people, crowd, strangers” is critical for subject consistency across a series. If you’re generating five portrait variations of the same character, suppressing extra faces prevents the model from hallucinating additional subjects into the frame — a problem that compounds badly in wider shots.

Step 4: Product Photography Negative Prompts

Product shots have their own pathology. The model loves to add props, backgrounds, context, and people that you never asked for. It also over-renders reflections and shadows in ways that look physically wrong. Negative prompts fix most of this.

person, hands, background objects, props, clutter, drop shadow, unnatural reflection, warped surface, distorted product shape, text label, barcode, price tag, watermark, studio equipment, visible lighting rig

The “visible lighting rig” exclusion is one that catches people off guard — at higher resolution tiers Nano Banana 2 sometimes adds realistic studio equipment into the scene as contextual fill. It looks authentic, which makes it worse. Flag it explicitly and the model keeps the frame clean.

Pro tip ✅

For product work at 4K resolution tiers, add “depth of field blur, bokeh” to your negative prompt if you need the entire product in sharp focus. The model defaults toward cinematic depth of field at higher resolutions — beautiful for portraits, wrong for a flat-lay product where every label needs to be readable.

Step 5: Editorial and Social Media Negative Prompts

Editorial images need to dodge anything that looks like ad content, and social media cuts require suppressing the UI decorations the model occasionally hallucinates — notification bubbles, like buttons, interface chrome. Yes, really.

advertisement, commercial look, product placement, watermark, interface elements, UI overlay, notification icons, social media buttons, text caption, hashtag, emoji overlay, over-edited, HDR, oversaturated, vignette

The “HDR, oversaturated, vignette” cluster is doing a lot of work here. Social media has trained these models on billions of heavily processed images, and left to its own devices Nano Banana 2 will cheerfully crank the contrast and slap a vignette on everything. Suppress it explicitly and you get images that actually look like they were made by someone with taste.

watermark, text, caption, advertisement, UI elements, interface chrome, oversaturated, HDR look, vignette, artificial lens flare, heavy grain filter, Instagram filter effect

Warning ⚠️

Don’t add “filter” to your negative prompt without a qualifier — it’s too broad and will suppress legitimate photographic effects you might want. “Instagram filter effect” or “heavy grain filter” gives the model enough context to understand what you’re actually excluding.

Step 6: Text and Watermark Suppression in Detail

Text rendering is one of the places where Nano Banana 2 has improved significantly over the first version, but it’s still not perfect. When you don’t want text in an image — for a clean background, a stock-style shot, or anything destined for overlay in a design tool — you need to be specific about what kind of text you’re excluding.

text, written text, words, letters, numbers, watermark, copyright symbol, signature, handwriting, typography, font, label, caption, subtitle, headline, printed text, digital text

That’s the aggressive stack. For most use cases you don’t need all of it — “text, watermark, signature” handles 90% of situations. But when you’re generating backgrounds for design work where even a single character would ruin the asset, run the full list.

Avoid 🚫

Don’t use “no text” or “without text” in your negative prompt — the model reads the negative field as a list of concepts to avoid, not as natural language instructions. “No text” might register “text” as a positive concept with a weak negative modifier, which can paradoxically make text more likely to appear. Just write “text” — the field itself is the negation.

Step 7: Brand Safety and Legal Exclusions

Creators generating content for clients or publication have a separate set of concerns beyond aesthetics. You want to make sure outputs don’t accidentally contain recognizable trademarks, faces that resemble specific real people, or brand imagery that could create legal exposure.

celebrity face, recognizable person, famous person, public figure, brand logo, trademark symbol, corporate logo, recognizable brand, copyrighted character, mascot, fictional character, real person likeness

This stack won’t make Nano Banana 2 legally bulletproof — nothing does — but it significantly reduces the probability of the model hallucinating a face that happens to look like someone famous, or generating a swoosh-shaped logo element that lands you in trademark trouble. Pair this with SynthID watermarks, which Nano Banana 2 embeds automatically in all outputs, giving you a provenance trail that shows the image was AI-generated.

Note 💡

SynthID watermarks in Nano Banana 2 outputs are invisible to the human eye but detectable by Google’s verification tools. They survive basic edits like cropping and color grading. You can’t disable them — they’re baked in at the infrastructure level. This is actually useful for brand safety documentation: you have a built-in record that the image is AI-generated, which matters for disclosure compliance.

Step 8: Combining Negative Prompts with Resolution Tiers

Nano Banana 2 offers multiple resolution outputs, and negative prompt behavior shifts slightly depending on where you’re operating. At standard resolution, the model is more forgiving of broad exclusions — you can throw a long negative list at it and it handles the contradictions gracefully. At 4K, precision matters more. The model has more capacity to interpret nuance, which means overly aggressive negative prompts can create visible holes in image quality — areas where the model has suppressed so many concepts that it doesn’t know what to put in their place.

The practical rule: at 4K, keep your negative prompt list under 15 items and make every item specific. At standard resolution, the broad base stack works fine. For batch operations, test your negative prompt on a single image at the target resolution before running the full batch — an over-suppressed 4K batch wastes generation credits and time in ways that a quick single test would catch immediately.

Pro tip ✅

When running batch operations through the API or Vertex AI, structure your negative prompt as a constant across the batch and vary only your positive prompt. This gives you a consistent quality floor across all outputs — useful when generating product variations or character consistency sequences where visual coherence matters more than individual image optimization.

The Prompts, All in One Place

Here’s every negative prompt from this guide consolidated for quick reference. Start with the base stack, then layer in the category-specific additions that match your use case.

watermark, text overlay, logo, signature, blurry, motion blur, low resolution, jpeg artifacts, overexposed, grainy, noise, oversaturated

extra fingers, deformed hands, distorted face, plastic skin, harsh shadows, background clutter, multiple people

person, hands, props, clutter, unnatural reflection, warped product shape, text label, watermark, visible lighting rig

advertisement, watermark, UI overlay, notification icons, text caption, HDR look, oversaturated, vignette, Instagram filter effect

text, watermark, copyright symbol, signature, handwriting, typography, label, caption, printed text, digital text

celebrity face, brand logo, trademark symbol, corporate logo, copyrighted character, real person likeness

What This Means for Your Workflow

Negative prompting in Nano Banana 2 isn’t magic, and anyone who tells you a perfect negative prompt eliminates all regeneration cycles is selling something. What it actually does is shift your ratio dramatically — instead of regenerating five times to get one clean image, you regenerate twice. That compounds fast at scale. A 50-image batch that used to require 80 manual review-and-regenerate cycles drops to 20. The time math is obvious.

The deeper value is consistency. Once you have a negative prompt stack that works for your use case — portrait, product, editorial, whatever — you lock it in and every image in your project starts from the same quality baseline. SynthID watermarks give you the provenance documentation. The structured exclusion field gives you the control. Put them together and Nano Banana 2 becomes a tool you can actually build a repeatable creative process around, not just a toy you poke at hoping for good results.

promptyze

See Full Bio