How to Fix Nano Banana IMAGE_SAFETY Errors

You wrote a perfectly reasonable prompt - a clean apparel shot, a product on a model, nothing unusual - and Nano Banana Pro kicked it back with an IMAGE_SAFETY error. It's one of the most common frustrations brand creators run into with this model, especially anyone working in fashion, apparel, jewelry, or cosmetics, and it's rarely obvious why a specific generation got flagged.

The error itself is unhelpfully vague. You typically see a flat rejection - sometimes labeled IMAGE_SAFETY, sometimes surfaced as a generic "unable to generate" message - with no explanation of which part of the request tripped the filter: the text prompt, the reference image, or the combination of the two. That ambiguity is exactly why this guide exists.

What IMAGE_SAFETY errors mean and why they happen

IMAGE_SAFETY is Google's content-safety filter stepping in before an image is returned. It's part of the same responsible-AI infrastructure documented in Google's Gemini API safety settings documentation, which governs how models in the Gemini family - including the image generation stack Nano Banana Pro runs on - screen output before it reaches the user.

It's designed to catch genuinely problematic content, but it's tuned conservatively - meaning it also catches a meaningful number of completely legitimate commercial requests, particularly anything involving people, clothing, or bodies in ways the filter's training data associates with risk. The filter has no awareness of commercial intent; it can't distinguish an ecommerce product listing from anything else it's trained to flag, which is why straightforward apparel and jewelry photography gets caught as often as it does.

Two things make this filter especially frustrating to work around. First, it's inconsistent - the same prompt can pass on one generation and get blocked on a near-identical retry, since the filter's scoring has some variance built in. Second, it gives no diagnostic detail, so you're often left guessing whether the text, the reference image, or some combination triggered the block.

The most common triggers

Four categories account for the overwhelming majority of IMAGE_SAFETY blocks brand creators run into. Recognizing which one applies to your specific rejection is the first step to fixing it efficiently, rather than rewording at random.

Apparel and fashion prompts

Requests involving models wearing clothing - especially close-fitting, swimwear, or undergarment product categories - trigger the filter disproportionately, even for straightforward ecommerce use cases. This is the single most common trigger category for brand creators using Nano Banana Pro for fashion work.

Jewelry, cosmetics, and on-model beauty shots

Close-up shots of jewelry worn on the body (necklaces, earrings framed close to the face, rings on hands) and cosmetics applied to skin can trigger the same filter as apparel, since both involve close framing on a person. The filter doesn't distinguish "close-up of a product on a body" from other close-framing requests it's trained to flag more broadly.

Pose and framing issues

Certain poses, camera angles, or crops - close-in framing on a person, for instance - can push a request over the line even when the intent is entirely standard product photography. This trigger is independent of product category; it can affect a home goods lifestyle shot with a person in frame just as easily as a fashion shot.

Reference image flags

If one of your uploaded reference images contains something the filter interprets as risky, it can flag the output even when your text prompt is completely clean. This is the trigger category most often missed, since creators assume a rejection must be a prompt-wording problem and spend time rewording text that was never the issue.

Step-by-step fixes

Work through these in order. Most blocks resolve at step 1 or step 2; the later steps are for cases that keep failing after an initial reword.

Step 1 - Reword the prompt to be more clinical and specific

Vague or suggestive-adjacent language ("stunning," "sultry," "alluring") gets flagged more than plain, descriptive commercial language ("model wearing the product, front-facing, studio lighting, neutral pose"). This single change resolves the majority of first-time blocks, since it's the fastest fix to test before investigating the reference image or pose.

Step 2 - Swap or crop your reference images

If a reference image is the trigger rather than the text prompt, try a cropped or alternate reference - sometimes background elements or framing in the source image are the actual issue, not your product. Test this by generating the same prompt with the reference image removed entirely; if it passes without the reference, you've isolated the trigger.

Step 3 - Adjust pose and framing language explicitly

Specify neutral, standing, front-facing poses rather than leaving pose interpretation open - ambiguity tends to push the model toward outputs more likely to get flagged. Explicit framing language ("full body, arms at sides, direct camera angle") removes the interpretive gap that can push a generation toward a riskier pose.

Step 4 - Break the shot into steps

For fashion, jewelry, and cosmetics product photography specifically, some creators find more success generating the garment or product alone first (flat lay or on a form) and compositing separately with Nano Banana Pro, rather than a full on-model shot in one pass. Splitting the generation into a clean product shot plus a separate compositing step avoids triggering the model/pose-related filter categories entirely.

Common error messages and what they mean

Not every rejection surfaces the same way. Knowing which message you're seeing narrows down which fix to try first.

Direct IMAGE_SAFETY rejection

A rejection explicitly labeled IMAGE_SAFETY almost always points to the text prompt or the combination of prompt and reference image. Start with Step 1 (reword to clinical language) before touching the reference image.

Silent or generic "unable to generate" failure

A vaguer failure message with no specific safety label is more often a reference-image issue - something in the uploaded photo the filter is reacting to independently of your text. Start with Step 2 (test without the reference image) to isolate the cause faster.

Intermittent pass/fail on identical retries

If the exact same prompt and reference pass on one attempt and fail on the next, you're seeing the filter's built-in scoring variance rather than a hard content violation. Retrying once or twice with no changes is a legitimate first move here, though it shouldn't replace addressing the underlying triggers if failures are frequent.

When to switch tools instead of fighting the filter

If you've reworded the prompt multiple times and you're still getting blocked, that's a signal to move that specific shot to a different tool rather than keep fighting the filter - Midjourney, for instance, doesn't apply the same restrictions to standard apparel photography. Reserve Nano Banana Pro for the shots where its precision genuinely matters (product-only shots, packaging, labels) and lean on other tools for on-model apparel work if the filter keeps getting in the way.

Reserve Nano Banana Pro for the shots where its precision genuinely matters, and lean on other tools when the filter keeps getting in the way.

Stuck rewording the same blocked prompt? The Prompt Fixer diagnoses vague or suggestive language layer by layer and rewrites it into the clinical, specific phrasing that clears the filter more reliably.

Fix my prompt →