Nano Banana 2 Trump: How Google’s AI Creates Presidential Images

Google’s viral AI image generator can produce stunningly realistic images of Donald Trump with no restrictions. Nano Banana 2 — officially named Gemini 3.1 Flash Image — does not block political figures the way Midjourney and OpenAI do, and that single decision turned into the biggest AI safety controversy of 2026.

The tool passed every test users threw at it: fake CNN election broadcasts, cinematic presidential portraits, Sims 4 characters, even fabricated government announcements. A NewsGuard audit found that the underlying technology generated convincing misinformation images for 30 out of 30 false claims on the first attempt. Here is what happened, how the technology works, and what it means for AI-generated political imagery.

What Is Nano Banana 2?

Google DeepMind developed three generations of its image model in under a year. The original Nano Banana launched in August 2025 and went viral within days, redefining what casual users expected from AI image editing. Nano Banana Pro followed in November 2025, built on the Gemini 3 Pro architecture, and pushed photorealism to a level that made people question whether traditional photography was finished. Then on February 26, 2026, Google released Nano Banana 2, combining Pro-level intelligence with the speed of Gemini Flash.

The official technical name is Gemini 3.1 Flash Image, though internally Google used the codename Gempix 2 during development. The model runs on a Multimodal Diffusion Transformer (MMDiT) architecture with approximately 1.8 billion parameters and uses Dynamic Quantisation-Aware Training (DQAT) to minimize memory usage without sacrificing output quality. Grouped-Query Attention (GQA) enables the model to generate images fast enough to run on a mobile device’s NPU without thermal throttling.

What sets NB2 apart from its predecessor is a multi-step generation loop: Plan, Evaluate, Improve. Instead of producing a single output, the model analyzes the prompt’s intent, generates an initial image, reviews it through built-in image analysis, identifies errors, and iterates before delivering the final result. This loop is why text rendering in Nano Banana 2 dramatically outperforms every previous version.

One of the most powerful new features is Image Grounding: the model searches the internet for specific images to understand what real-world subjects look like before generating them. This works exceptionally well for locations, monuments, and biological species. There is one critical limitation, however: Image Grounding cannot search for people. When generating faces, including Trump’s, the model relies entirely on its training data.

Nano Banana 2 and Trump: What Happened

The Trump image controversy started before Nano Banana 2 even officially launched. A preview version appeared on Media IO, a third-party platform, and testers immediately discovered that Google’s latest text-to-image generator had minimal restrictions on political figures. The results spread across X, Threads, and Instagram within hours.

One image became the defining example of the controversy: a fake CNN broadcast showing the 2027 presidential election results, with Donald Trump projected to win a third term in office. The screenshot looked indistinguishable from a real news broadcast. It included accurate CNN graphics, realistic text overlays, and the kind of production quality that would fool anyone scrolling through a social media feed quickly.

Users then pushed the boundaries further. Prompts like “Donald Trump and Xi Jinping in Sims 4” produced results that demonstrated the model’s ability to render any public figure in any context. Typographic prompts added political commentary to photorealistic scenes. Cinematic portraits with studio lighting made fictional Trump scenarios look like professional editorial photography.

The core issue was not just quality but policy. Midjourney explicitly blocked all requests for images of Joe Biden and Donald Trump ahead of the 2024 presidential election. OpenAI designed its Sora video generator to reject prompts depicting living public figures without their consent. Google took a fundamentally different approach with Nano Banana 2: no stated block on political figures at all.

Google’s generative AI use policy states that users cannot “engage in misinformation, misrepresentation, or misleading activities,” including “facilitating misleading claims related to governmental or democratic processes.” But the policy does not specifically mention political figures, does not prevent their generation, and does not state what penalties apply to users who violate the rules.

The Misinformation Controversy

The safety concerns around Nano Banana 2 and Trump images did not emerge in a vacuum. Before NB2 even launched, NewsGuard had already tested its predecessor, Nano Banana Pro, with devastating results.

NewsGuard compiled 30 provably false claims from its proprietary database and organized them into six categories: public health, U.S. politics, European politics, the Middle East, global brands, and Russian influence operations. Each category contained five claims that had recently circulated online and could plausibly be advanced through images. The researchers created short, standardized prompts for each claim and tested them using Google Gemini’s Nano Banana Pro.

The result: Nano Banana Pro produced convincing images for every single false claim on the first attempt. Not one prompt was rejected. That is a 100% failure rate on basic safety guardrails.

What made the findings more alarming was that the model sometimes added unprompted details to make fabricated images more convincing. When NewsGuard prompted an image of an MSNBC broadcast, Nano Banana Pro added a real anchor’s face without being asked. When prompted about a Cracker Barrel CEO controversy, the model looked up the CEO’s actual name, Julie Masino, and included a fabricated social media handle with her first and last name.

The SynthID Watermark Problem

Google embedded two layers of provenance into Nano Banana 2 images. Every generated image carries a visible watermark in the bottom left corner and an imperceptible digital SynthID watermark developed by DeepMind. The company also announced plans to add C2PA Content Credentials for interoperable verification across platforms.

The SynthID verification feature in the Gemini app has been used over 20 million times since its November launch. Users can upload any image and ask whether Google AI generated it. On paper, this sounds like a robust system.

In practice, the visible watermark can be cropped out in seconds using any basic photo editor. The imperceptible SynthID watermark requires access to specialized AI detection tools that ordinary social media users do not have. And a 2023 University of Waterloo study found that only 61% of people could correctly identify AI-generated images without tools, well below the anticipated 85%.

The age of photographic evidence is over. Images are no longer witnesses and now they are mere hypothesis.

Maalvika Bhat, podcaster, in an Instagram reel viewed over 800,000 times

A viral X post comparing the original Nano Banana output with Nano Banana Pro output was viewed 56 million times. The old version had the classic AI look: sheeny, airbrushed, and immediately identifiable as synthetic. The Pro version looked like a real photograph pulled from someone’s Instagram feed. That 56-million-view post crystallized what researchers had been warning about: AI-generated photorealism had crossed a threshold where human detection was no longer reliable.

How to Create Trump Images with Nano Banana 2

Despite the controversy, Google has not removed the ability to generate political figures with its AI image generator. Here is how the process works.

1. Open the Gemini app or visit gemini.google.com in a browser
2. Select any model mode (Fast, Thinking, or Pro) — Nano Banana 2 is the default image generator across all three
3. Type a descriptive prompt that includes “Donald Trump” along with the desired scenario, style, and composition details
4. Wait a few seconds for the image to generate, then download or iterate with follow-up prompts
5. For maximum quality, switch to Nano Banana Pro via the three-dot regeneration menu (requires Google AI Pro or Ultra subscription)

The key to getting better results lies in prompt specificity. Generic prompts like “a photo of Trump” produce generic results. The images that went viral on social media used detailed, cinematic prompts that specified lighting conditions, camera angles, artistic styles, and environmental context.

Thinking mode matters for complex scenes. By default, Thinking mode is off, and Google’s own developer guide recommends keeping it off for standard image generation to save time and processing. But for multi-element political scenes that require spatial reasoning, coherent text overlays, or complex compositions, toggling Thinking mode on produces dramatically better results.

Resolution strategy affects both cost and quality. Nano Banana 2 supports resolutions from 512px to 4K. A cost-effective workflow is to generate multiple variations at 512px (which costs roughly the same as the original Nano Banana), select the best composition, and then upscale that specific image to 4K. The Batch API offers an additional 50% discount for bulk generation.

Image Grounding cannot help with faces. While NB2 can search the internet for visual references of locations, buildings, and animal species, it explicitly cannot search for people. All face generation comes from the model’s training data, which means some prompts may produce less accurate facial likenesses than users expect.

Nano Banana 2 vs. Competitors: Who Blocks Political Figures?

The AI image generation space fractured in 2024-2025 over the question of political content. Each major platform made a different decision, and those decisions defined their relationship with users, regulators, and the media.

Midjourney took the most aggressive stance. Ahead of the 2024 U.S. presidential election, the company blocked all requests for images of Joe Biden and Donald Trump. The block was categorical: no political figure generation, no exceptions. This approach eliminated the misinformation risk entirely but frustrated users who wanted to create political satire, editorial illustrations, or artistic commentary.

OpenAI positioned itself between restriction and permissiveness. Its Sora video generator is designed to block the generation of images depicting living public figures without their consent. DALL-E applies similar restrictions, though enforcement varies. The policy acknowledges the creative value of political imagery while trying to prevent malicious use.

Adobe Firefly focused on a different dimension of the problem entirely. Rather than policing individual prompts, Adobe built its model on licensed and copyright-safe training data. This approach limits what the model can produce, including political figures, but frames the restriction as a copyright issue rather than a safety one.

ByteDance’s Seedance entered the market amid Hollywood copyright lawsuits from Disney and Paramount over intellectual property violations. The tool’s guardrails around political figures remain unclear, and the company has not published explicit policies.

Google, with Nano Banana 2, chose permissiveness. No stated block on political figures. No prompt rejection for public figure requests. The only guardrail is the general AI use policy prohibiting “misleading activities,” which places the burden of responsible use entirely on the user.

Ethical and Legal Implications

U.S. law does not specifically prohibit creating AI images of public figures for non-commercial, non-defamatory purposes. The First Amendment protections around political speech are broad, and courts have historically treated satire and commentary involving public figures as protected expression.

The legal landscape shifts when intent crosses into deception. Using AI-generated political imagery to deceive voters, fabricate government announcements, or spread false claims about elections may violate state election laws and federal regulations around election interference. At least 10 states have introduced or passed legislation addressing AI-generated deepfakes in political campaigns since 2024, though enforcement mechanisms remain weak.

Google’s AI use policy adds another layer. Users who “facilitate misleading claims related to governmental or democratic processes” violate the terms of service. But Google has not stated what penalties apply, has not described any enforcement mechanism, and did not provide comment when NewsGuard presented its 30/30 failure rate findings.

I find it irresponsible to release this to the market without sufficient guardrails and then to sell it with problematic stereotypes and outdated views about women.

Sandra Wachter, Professor of Technology and Regulation, Oxford Internet Institute

The tension between creative freedom and safety will only intensify. AI image quality improves with each model generation. Human detection rates remain low at 61%. Watermarking solutions exist but are easily circumvented. And the demand for political imagery, both legitimate and malicious, is not decreasing.

Where to Access Nano Banana 2

Google made NB2 available across its entire product ecosystem on launch day, rolling out to 141 new countries and territories with support for 8 additional languages.

• Gemini app — the primary consumer interface. A small number of free image generations are available without a subscription. Google AI Pro and Ultra subscribers get full access and can switch to Nano Banana Pro for specialized tasks through the regeneration menu
• Google Search — integrated into AI Mode and Google Lens, accessible through the Google app and both mobile and desktop browsers
• AI Studio + Gemini API — available in preview for developers at aistudio.google.com
• Vertex AI — Google Cloud’s enterprise platform, available in preview through the Gemini API
• Flow — Google’s creative tool where Nano Banana 2 is the default image model, available for zero credits
• Google Ads — powers AI-generated image suggestions while creating ad campaigns

For developers optimizing costs, 512px image generation with Nano Banana 2 costs roughly the same as the original Nano Banana. The Batch API provides a 50% discount for processing multiple prompts simultaneously, making large-scale generation projects feasible at around $45 for 100+ images.

FAQ