Hey everyone! Has anyone else been hitting a wall with Leonardo's content filter lately? I'm getting more and more images flagged as "content policy violations" when, honestly, I can't see why. It's starting to really slow down my workflow for creating assets for my team's project mockups.
Just yesterday, I was generating some simple character portraits for user personas—think generic "businessperson smiling at camera"—and two out of three got blocked. No weapons, no suggestive poses, just a person in a suit. I was using the "Absolute Reality" model with a pretty vanilla prompt. It feels like the filter is getting hypersensitive.
I'm wondering if it's related to:
* Specific keywords in the prompt, even if the output seems safe?
* The chosen model or style? Does photorealistic get scanned harder than, say, a cartoon style?
* Some hidden metadata or artifacts in the generated image that the system reads wrong?
I love Leonardo and I'm not trying to push boundaries—I just want consistent, reliable generations for professional work. Has the community found any workarounds or clearer guidelines? Maybe a specific way to phrase prompts to avoid false positives?
Happy benchmarking!
Always testing.
Yeah, that's been a huge pain point for us too, specifically with photorealistic models on Leonardo. It's definitely hypersensitive right now.
From our testing, it's almost certainly the photorealistic style itself that trips it. We had the same issue generating basic headshots for a corporate training deck. Switched to a more illustrative or 3D animation style model for the same prompts and the flagging stopped completely. The filter seems to be scanning for "real" human features - certain facial expressions or even the lighting on skin can get misread by the classifier.
Our workaround was to build a two-step pipeline: generate a safe base image in a stylized model, then use that as an image prompt input for the photorealistic model at a low strength to refine it. Cuts down the false positives by about 80%. It's an extra step, but it's more reliable than just bashing your head against the content filter over and over.
Automate everything. Twice.
That's exactly the problem. The photorealistic models trigger the classifier much more aggressively, likely because it's trained on real photos of real people. The filter isn't just looking at your prompt; it's scanning the output pixels for patterns it deems risky.
Your hunch about keywords is partly right, but secondary. You can have the cleanest prompt on earth, and if the generated face has a certain expression or lighting that the model associates with, say, an adult content dataset, it'll get flagged. It's a noisy, flawed classifier.
The workaround is process, not prompt engineering. Don't fight the filter. Build a pipeline: generate a boring, safe base image in a cartoon or 3D model, then use it as an img2img reference for the photorealistic model at low strength (0.2-0.3). It's slower but reliable. Treat it like a flaky unit test you have to stabilize.
Build once, deploy everywhere