Project Details
- GCP Project ID: provador-virtual-49421..
- GCP Project Number: 7839727832..
- Target Models: gemini-3.1-flash-image, gemini-3.1-flash-lite-image, gemini-3-pro-image (and any dedicated Virtual Try-On successors)
1. Executive Summary & Business Impact
We operate a legitimate fashion e-commerce platform (Zavemi / Naked, nakedsw.com.br ) that generates high-quality, on-model product photography for apparel retailers. Our core category is swimwear, beachwear, and resort wear. These are standard, commercial retail product listings (comparable to swimwear catalogs shown on Amazon, Zalando, or major department stores).
We received a deprecation notice for virtual-try-on-001 and gemini-2.5-flash-image. However, the recommended migration path (gemini-3.1-flash-image and gemini-3-pro-image) breaks our core business. The newer Gemini image models apply a non-configurable, post-generation safety filter (Layer 2 IMAGE_SAFETY) that blocks generated apparel try-on outputs when the garment is swimwear (e.g., bikinis, one-pieces), despite the request working perfectly for non-swimwear items.
The virtual-try-on-001 model handles swimwear correctly without false-positive blocks. Because swimwear is our core business, the forced deprecation without an allowlist or functional successor will break our production platform.
2. Technical Diagnosis & Observed Behavior
We have thoroughly tested the recommended models via the :generateContent endpoint (with location set to global / regional equivalent) and verified the following:
- API Configuration Applied:
- safetySettings set to BLOCK_NONE across all four configurable categories (HARM_CATEGORY_HARASSMENT, HARM_CATEGORY_HATE_SPEECH, HARM_CATEGORY_SEXUALLY_EXPLICIT, HARM_CATEGORY_DANGEROUS_CONTENT).
- imageConfig.personGeneration set to “allow_all” (or “allow_adult”).
- A minimal, compliant edit prompt with two input images (a standard model photo and a swimwear garment).
- Behavior on Non-Swimwear Apparel (e.g., dresses, tops, trousers):
- finishReason = STOP. The request completes successfully, returning a high-quality generated image with subject identity preserved.
- Behavior on Swimwear (Bikinis / One-pieces):
- The API call runs for approximately 40–47 seconds (indicating the image is successfully generated in the backend), but then fails with a safety finish reason (IMAGE_SAFETY block) at the output stage.
- Comparison with Legacy Model:
- The exact same swimwear inputs processed via virtual-try-on-001 (:predict) return a flawless, compliant output with zero blocks.
This indicates that the block originates from an internal, non-configurable image-safety classifier that cannot be managed via client-side safetySettings.
3. Compliance & Responsible AI Position
- Policy Compliance: Our application strictly complies with Google Cloud’s Acceptable Use Policy and Generative AI Prohibited Use Policies. We do not generate sexually suggestive, adult, or non-consensual content.
- Commercial Context: All generated content is strictly for commercial e-commerce retail display.
- Application Guardrails: We enforce application-level content moderation and input filtering to ensure our platform is used solely for legitimate apparel retail listings.
- Ready for Review: We are fully prepared to complete a formal Google Responsible AI use-case review, provide sample assets (retail garments and model inputs), or schedule a call with Google’s trust and safety engineers.
4. Requested Action
To prevent a complete disruption to our production service, we request one of the following resolutions:
- Project Allowlisting: Please add our project (provador-virtual-494213) to the allowlist for sensitive-content / person-generation on the newer Gemini image models (gemini-3.1-flash-image, gemini-3-pro-image) to prevent false-positive blocks on legitimate swimwear.
- Deprecation Extension / Successor Access: Grant us an extension of service for virtual-try-on-001 or early access to a dedicated Virtual Try-On successor that supports swimwear retail categories.