Prompt Rejection Diagnoser

Why Gemini over-blocks harmless prompts

The safety check that runs before image generation is not the same system that understands your prompt. It is a fast classifier that scores surface-level signals — individual words, word combinations, name lookups — and it is deliberately tuned so that a false negative (letting something bad through) costs far more than a false positive (blocking your beach photo). The result: perfectly innocent prompts get refused because one word also appears in genuinely bad prompts. Google has acknowledged stretches where this over-blocking spiked noticeably; we documented one of them, with sources, in Why Gemini Omni blocks your prompts — the real rules vs the current bug.

The practical consequence cuts both ways. Because the filter reads words rather than intent, a neutral rewording of the trigger word usually unblocks a harmless prompt — community threads consistently report roughly seven or eight out of ten false positives clearing after a rewrite. And because it reads words rather than intent, no rewording makes disallowed content allowed: the same classifier runs again on the generated image itself. That is why this page only helps with the first case.

The eight false-positive categories

These are the trigger patterns the community has documented most often, and the ones the diagnoser above checks for:

1. Real people and celebrities. “Taylor Swift on a stage” blocks; “a famous pop star performing on a stadium stage” renders. Photorealistic depictions of identifiable real people are partly a genuine restriction (likeness and misinformation risk), so the reliable fix is a generic role description — not a workaround for the name.
2. “In the style of” a living artist. Naming living artists trips style-mimicry rules. Describing the aesthetics instead — “bold ink outlines, flat pastel palette, dreamlike composition” — passes and gives you more control. Long-dead artists whose work is public domain (Van Gogh, Monet) rarely trip it.
3. Brands and trademarks. Logos, product names, and franchise characters: “a Coca-Cola can” blocks more often than “a red soda can with white cursive lettering” — which renders the same object.
4. Age-ambiguous words plus photorealism. The single most common false positive. “Girl”, “teen”, or “school uniform” combined with “realistic photo” puts the classifier on maximum alert. If you mean an adult, say so: “a young adult woman in her twenties”. For legitimate family or children’s content, use an illustration-style qualifier instead of photorealism. Anything that actually sexualizes minors is a hard policy line — not a false positive, and not something any rewording should or will touch.
5. Medical and anatomy terms. “Surgery”, “syringe”, “anatomy” read as gore or adult signals out of context. Clinical-educational framing — “medical textbook diagram, labeled, educational illustration style” — tells both the model and the filter what you actually want.
6. Violence-adjacent action words. Fantasy battles, sparring poses, and action scenes are renderable; words like “kill”, “gun”, and “blood” still spook the filter. Stylization plus an explicit “no gore” usually clears it. Realistic graphic violence is a genuine policy line.
7. Ambiguous homonyms. “Photo shoot”, “shot on DSLR”, “killer look”, “execute this concept”, “comic strip” — everyday phrases whose other meaning is violent or illicit. These are the easiest fixes: “photo session”, “captured with”, “striking look”, “carry out”, “comic panel sequence”.
8. Skin, body and wardrobe words. “Nude lipstick”, “flesh tones”, “bare shoulders”, “naked eye” — cosmetic and idiomatic uses that trip adult-content keywords. Swap to “neutral beige”, “warm skin tones”, “off-shoulder neckline”, “unaided eye”.

A worked before / after example

A real-world composite of the most common triggers — the “before” version gets blocked, the “after” version renders the identical harmless image:

Five swaps, five categories: the age-ambiguous “girl” becomes an explicit adult; the firearm homonyms “shot” and “killer” become “captured” and “radiant”; the trademark becomes a generic description of the same object; and the living photographer’s name becomes a description of her signature look. Nothing about the image changed — only the words that were being misread.

Related reading

• Why Gemini Omni blocks your prompts — the real rules vs the current bug — how to tell a genuine policy block from a filter misfire, with sources.
• Gemini Omni failure modes — when the problem is not the filter but the model itself: text, hands, drift, and prompt length.
• Photo editing without changing the face — preserving identity through edits once your prompt gets through.

More free Gemini tools

• Nano Banana prompt generator — build a clean prompt from tested phrase blocks instead of debugging one.
• Image fix prompt picker — repair blurry, plastic, or warped output once generation succeeds.

FAQ