Nconsistent PROHIBITED_CONTENT errors (False Positives) in Kids' Virtual Try-On Project using gemini-3.1-flash-image-preview

Hello,

I am currently developing a commercial “Kids’ Virtual Try-On” service using the gemini-3.1-flash-image-preview model in Google AI Studio.

This service is fully legitimate, operating strictly with parental consent, where parents voluntarily upload photos to see how clothes would look on their children. It has absolutely no malicious intent.

However, I am experiencing a critical issue with safety filtering inconsistency:

The Issue: The API frequently returns a PROHIBITED_CONTENT error.

Inconsistency (False Positives): The most frustrating part is that when I submit the exact same source image and the exact same prompt (e.g., a standard t-shirt) twice, it succeeds once but gets blocked the next time.

I understand that Google maintains a strict built-in policy regarding minor protection. However, since the system blocks the content randomly based on slight variations in the generated output rather than the input itself, it makes commercial service deployment impossible due to poor predictability.

Could the engineering team please review this over-blocking/false-positive issue for legitimate virtual try-on use cases? Is there any recommended prompt structure or pre-processing guideline to avoid this borderline filtering behavior?

Thank you.

Thank you for the clear explanation. I completely agree with your assessment regarding the strict safety filters and the model’s non-deterministic output. I would highly appreciate it if the engineering team could review this issue to help stabilize the predictability for commercial services.