Possible regression in IMAGE_SAFETY filtering for Nano Banana Pro affecting normal fashion use case

Gemini API
,
1

Dear Gemini team,

This is my first time submitting feedback in the Google community. I am a long-time Google user and currently work in the lingerie and sleepwear industry.

Over the past few months, Nano Banana Pro has provided significant value to our workflow. It has greatly reduced the cost of physical model photoshoots while substantially improving the efficiency and realism of fashion presentation images.

However, in recent days, image generation using Nano Banana Pro has been consistently rejected. Even images showing models wearing lingerie—when clearly non-NSFW and intended solely for legitimate product display and fashion presentation—can no longer be generated.

To further isolate the issue, I tested text-only prompts without uploading any images. When prompts contained very common daily fashion terms such as “lingerie styling,” “bikini,” “beach swimwear,” or even “nightgown / sleepwear,” image generation was still blocked.

From a real-world perspective, lingerie styling, beach bikinis, and sleepwear are all normal and widely accepted forms of clothing presentation in everyday life. They are commonly seen in advertising, e-commerce platforms, fashion catalogs, and lifestyle photography. Completely blocking these categories makes Nano Banana Pro nearly unusable for many legitimate professional and personal use cases.

If this behavior is the result of a recent safety or policy update, I fully understand and appreciate the team’s efforts toward responsible AI usage. However, in its current state, this change represents a functional regression for normal fashion, e-commerce, and everyday lifestyle applications, and is likely caused by overly aggressive IMAGE_SAFETY filtering or false positives.

I sincerely hope the team can consider refining the safety criteria to better distinguish NSFW content from normal daily fashion imagery. At present, blocking even sleepwear generation feels excessively restrictive and significantly reduces the practical value of the model.

Thank you for your time and for your continued work on Gemini.

1 Like
2

I’m in a similar situation where I am using Nano Banana Pro to build a Virtual Try On feature that can be used to pick and choose clothing items.

I’m sending model images based on user’s picture to generate studio-style professional images, but gemini-3-pro refuses to generate try-ons that aren’t even remotely NSFW.

It works fine for regular use cases, like office wear and hiking, but fails to work very often if sleepwear or beachwear is involved.

I’ve tried modifying the prompts to explicitly prevent any hints of sexualization, but nothing works.

This is quite frustrating.