If you've used any AI image generation model at scale, you've hit the wall. Google's Gemini has strict content policies around real people in images — and YouTube thumbnails almost always include faces, which means frequent rejections.

We kept running into this with Thumbfa.st. Users would try to generate thumbnails featuring themselves or other creators, and Gemini would block the request. Frustrating, and a blocker for a core feature.

So we built a workaround: thumbnail anonymization.

What Anonymization Actually Means

The idea is simple. Before sending a thumbnail to the AI model, we strip out the real characters — faces, recognizable features — and replace them with generic placeholder silhouettes.

The AI sees a clean, policy-safe image with neutral shapes where the people used to be. No real faces, nothing that triggers content moderation. The model can then work freely with the composition, background, style, and text overlays — the parts that actually define a great thumbnail.

Why This Works

Gemini's restrictions aren't about thumbnails specifically — they're about detecting and processing real people's likenesses. By replacing characters with silhouettes before the request, we never send the sensitive content in the first place.

The model doesn't need to know who was in the original image to do its job. It just needs to understand the layout and style. Placeholders give it exactly that.

Before and After

Here's what the input looks like — a thumbnail with real characters replaced by neutral silhouettes:

And here's the placeholder used during the anonymization step:

The AI generates a new version based on that anonymized input. Then, the target person gets composited back in — whether that's you, a custom character, or anyone else. Here's the final result:

The quality is genuinely good. The model preserves pose, framing, and energy from the original while working cleanly within policy boundaries.

The Bigger Benefit: Any Character, Any Thumbnail

What started as a workaround turned into a real product improvement. Because we're already doing the anonymize → replace pipeline, it becomes trivially easy to swap in any person into any thumbnail.

Want to reuse a thumbnail layout but put yourself in it instead of the original creator? Done. Want to generate ten variations with different people? The anonymization step makes that clean and consistent every time.

This is something that used to require careful prompt engineering and a lot of luck. Now it's just how the pipeline works.

What We Learned

The lesson here isn't really about Gemini's policy — it's that working with AI content restrictions often forces you to think harder about your pipeline, and that thinking usually pays off.

Anonymization made our generations more reliable, more flexible, and easier to scale. The restriction was a constraint that pushed us toward a better architecture.

If you're building with image generation models and hitting similar walls, consider what the model actually needs vs. what you're sending it. There's often a way to preserve what matters while removing what triggers the policy.