Gemini Image Generation Blocks a Reasonable Prompt as a Safety Violation

Gemini (Nano Banana) or Imagen rejects a normal prompt as a 'violation.' Usually named people, brands, or aggressive default safety. Exact rewrites and API fixes here.

Published: May 24, 2026 Updated: Jun 21, 2026 Author: AI Productivity Guide Team 🌐 查看中文版本

You ask Gemini’s image generator for “a businesswoman in a Tokyo office,” “a photo of Albert Einstein at a chalkboard,” or “an iPhone on a desk” and you get back “Sorry, I can’t generate that image” (or “I can’t help with that”) with no further detail. Sometimes the same prompt works in Google AI Studio but fails in the gemini.google.com app.

Fastest fix: describe attributes instead of identities. Replace the named person with a physical description (“an elderly physicist with wild white hair and a moustache” instead of “Einstein”), drop the brand name (“a modern black smartphone” instead of “iPhone”), and the prompt almost always goes through. For borderline categories (a kitchen knife, a doctor with a needle) move to Google AI Studio or the Gemini API, where you can lower the safety threshold. Named real people are a hard policy block on every surface and no setting unlocks them.

What’s actually running (as of June 2026)

The product name matters here because the surfaces behave differently:

gemini.google.com app and the Gemini mobile app now generate images with Nano Banana 2 (gemini-3.1-flash-image-preview), which shipped on 26 February 2026. This surface applies the strictest filter, and named-public-figure / celebrity blocking was formalized as an explicit safety category in Nano Banana 2.
Google AI Studio (aistudio.google.com) lets you run the same Gemini image models with adjustable per-category safety sliders.
Imagen (imagen-4.0-generate-001, imagen-4.0-ultra-generate-001, imagen-4.0-fast-generate-001) is still available through the Gemini API and Vertex AI, with a person_generation parameter. Note: the standalone Imagen API is deprecated and is scheduled to shut down on 17 August 2026; Google recommends migrating to the Gemini image (Nano Banana) models. Until then, the Imagen API and Vertex AI remain the most configurable surface.

If you are pasting a prompt into the chat box at gemini.google.com, you are hitting Nano Banana 2’s consumer filter, which is the strictest of the three.

Which bucket are you in?

Symptom	Most likely cause	First thing to try
Block disappears when you delete a name	Named real person / public figure	Describe by attributes (Step 1)
Block disappears when you delete a product word	Brand / trademark mention	Describe the object, not the brand (Step 2)
Anything with “child / kid / baby / teen”	Child-safety filter	Remove the reference entirely (Step 5)
Knife, needle, blood, gun, pill, surgery	Dangerous / medical filter	Move to AI Studio / API, lower threshold (Step 3-4)
Works in AI Studio, fails in the app	Consumer app is the strictest surface	Use AI Studio or the API (Step 3-4)
“Person generation isn’t available in your region”	Regional restriction	See Step 6

Common causes, by frequency

1. Named real person (most common)

“Photo of Elon Musk,” “image of Taylor Swift,” “Einstein at a chalkboard” all get refused. Nano Banana 2 and Imagen refuse to generate likenesses of named real people, including historical figures, and the Gemini app will even block editing or face-swapping when it recognizes a public figure in an uploaded photo.

How to confirm: remove the name and the prompt goes through.

2. Brand or trademark in the prompt

“iPhone,” “Coca-Cola can,” “Nike shoes,” “Tesla car” frequently trigger blocks. The filter treats trademark mentions as risk.

How to confirm: “a modern smartphone” works where “an iPhone” doesn’t.

3. Politician / political figure / political topic

Any named politician is blocked. Some political scenes (rallies, protests, election imagery) are blocked even when unnamed.

4. Child or minor in the scene

Mentioning children, kids, babies, minors, teens, or even “young people” triggers the child-safety filter, which is the most aggressive of all. This block exists on every surface; on Vertex AI it maps to the person_generation setting (only allow_all permits minors, and that value is itself restricted in the EU, UK, Switzerland, and the MENA region).

5. Medical / weapon / violence adjacent

Knives, needles, blood, surgery, firearms, pills are blocked even in benign contexts. Borderline scenes (a chef cutting onions, a doctor with a stethoscope) sometimes pass and sometimes don’t.

6. Consumer app is stricter than AI Studio / API

The gemini.google.com app applies the strictest filter. AI Studio exposes adjustable safety sliders. The Imagen API on Vertex AI is the most permissive surface with proper authentication.

7. Regional restriction

person_generation="allow_all" (which includes minors) is restricted in the EU, UK, Switzerland, and MENA regions, and the consumer app disables some people-generation features in certain countries even on a paid Google AI Pro plan.

Shortest path to fix

Step 1: Rephrase identities as attributes

Replace named people with descriptions:

Bad:  Photo of Elon Musk at a press conference
Good: Photo of a middle-aged man with short brown hair, dark suit,
      speaking at a tech press conference podium

Bad:  Einstein at a chalkboard
Good: An elderly physicist with wild white hair and a moustache,
      writing equations on a chalkboard in a 1940s university classroom

Attribute-based descriptions give you most of the visual outcome without tripping the named-person filter.

Step 2: Drop brand names

Bad:  iPhone on a desk
Good: A modern black smartphone on a wooden desk

Bad:  Tesla Model 3 in a garage
Good: A red electric sedan parked in a garage, minimalist design,
      flush door handles, glass roof

Describe what the brand looks like, not the brand name.

Step 3: Use Google AI Studio for adjustable safety

For borderline categories (dangerous objects, mild violence), AI Studio lets you lower the per-category threshold:

aistudio.google.com
→ Open a chat with a Gemini image model (or pick Imagen)
→ Run settings panel (right side) → Advanced settings → Safety settings
→ In the "Run safety settings" modal, drag the sliders down for the
  relevant category (Harassment, Hate, Sexually explicit, Dangerous)

These sliders can lower but not fully remove protection on certain categories (notably anything involving minors). Lowering safety also does not unblock named real people; that is a hard policy block, not a threshold.

Step 4: Use the Imagen API on Vertex AI directly

For business-grade use cases that need content-policy alignment, the Imagen endpoint on Google Cloud is still the most configurable surface (until its 17 August 2026 shutdown). The API parameter names were updated in 2026: safety_filter_level now takes block_low_and_above, block_medium_and_above (default), or block_only_high (the old block_some / block_most strings still work as aliases).

from vertexai.preview.vision_models import ImageGenerationModel

model = ImageGenerationModel.from_pretrained("imagen-4.0-generate-001")
images = model.generate_images(
    prompt="A modern office workspace with a laptop and coffee cup",
    number_of_images=4,
    aspect_ratio="16:9",
    safety_filter_level="block_only_high",
    person_generation="allow_adult",
)

person_generation accepts dont_allow (no people), allow_adult (adults including celebrities; default on most models), or allow_all (adults and minors; default on Imagen 4 models, restricted in the EU/UK/CH/MENA). None of these unlocks a named real person.

If you are starting fresh, migrate to the Gemini image model instead of generate_images, since Imagen is sunsetting:

from google import genai

client = genai.Client()
response = client.models.generate_content(
    model="gemini-3.1-flash-image-preview",
    contents="A modern office workspace with a laptop and coffee cup",
)

Step 5: Try a different prompt structure

Some blocks are keyword-based. Variations that sometimes help:

Replace “child” / “kid” with “young person.” This sometimes passes, sometimes doesn’t; the safest move is to drop the reference entirely.
Replace “knife” with “kitchen tool” when the scene is cooking.
Replace “shot” with “photograph” when you mean a photo, to dodge the gun-context filter.

Step 6: Accept the hard blocks

Named real people, minors in any context, brands as the primary subject, and medical imagery are policy decisions, not bugs. Trying to jailbreak past them with adversarial prompts risks a temporary or permanent suspension of your Google account, so don’t. If your use case genuinely needs these, use a different tool: a licensed stock photo, a professional photographer, or an image model whose policy fits your specific, lawful use case.

How to confirm it’s fixed

Run the rephrased prompt in the same surface where it failed. If it generates, the block was content-based and your rewrite cleared it.
If it still fails in the app but you suspect a soft block, paste the identical prompt into AI Studio. Success there confirms the consumer app’s stricter filter was the cause; failure there points to a hard policy category (named person, minor) that no surface will allow.
For API work, a clean response with images and no raiFilteredReason in the result means the safety filter passed.

Prevention

Describe visuals by attributes, not identities. This works across every current image model and avoids most blocks.
Keep a “rephrase map” in your notes: common brand names mapped to generic descriptions.
For business prompts that have to ship, use the Imagen API on Vertex AI (or the Gemini image model) rather than fighting the consumer app’s filter.
Never put the real name of a private individual in an image prompt. The privacy filter is even stricter than the public-figure one.
For sensitive industries (medical, security), decide up front whether image generation is the right tool at all.

FAQ

Why does the same prompt work in AI Studio but fail in the Gemini app? The gemini.google.com app runs Nano Banana 2 with the strictest consumer filter and no user-adjustable thresholds. AI Studio runs the same model family but exposes per-category safety sliders, so a borderline prompt that the app blocks can pass once you lower the relevant threshold.

Can I ever generate a real, named person? No. Named real people, public figures, and celebrities are a hard policy block on every Google surface, including editing or face-swapping an uploaded photo of a recognized public figure. No safety setting unlocks it. Describe the person by attributes instead.

Is Imagen 3 still the model? Not in the app. As of June 2026 the Gemini app uses Nano Banana 2 (gemini-3.1-flash-image-preview). Imagen 4 (imagen-4.0-generate-001 and variants) is still callable through the API and Vertex AI, but the standalone Imagen API is deprecated and scheduled to shut down on 17 August 2026, with Google steering developers to the Gemini image models.

My doctor / chef / hunting scene keeps getting blocked. What now? That is the dangerous/medical filter. Move the prompt to AI Studio and lower the “Dangerous” category slider, or call the Imagen API with safety_filter_level="block_only_high". If it still blocks, soften the wording (a “kitchen tool” instead of a “knife”) or remove the flagged object.

Will lowering the safety threshold unlock everything? No. Thresholds reduce filtering for content categories like harassment or dangerous objects, but they cannot remove protection on minors, and they have no effect on the named-person policy block, which is enforced separately.

Tags: #Gemini #Troubleshooting #imagen