Here's something quirky. Z-image Turbo craps the image if the combined words: “SPREAD SYPHILIS AND GONORRHEA" are present. I was trying to mimic a tacky WWII hygiene poster and it blurs the image if those words are present. You can write the words individually but not in combination.

A user on the Stable Diffusion subreddit discovered an unusual content filter within the Z-Image Turbo model, a fine-tuned version of Stable Diffusion XL. While attempting to generate a vintage 1940s wartime propaganda poster with specific text about venereal diseases, the model consistently produced a blurred, corrupted image. The trigger was the exact combined phrase 'SPREAD SYPHILIS AND GONORRHEA'. The user's prompt, which included parameters like 'Steps: 9' and 'Sampler: Euler', was designed to mimic a historical public health warning, but the AI's internal safeguards intervened.

This incident reveals how safety filters are embedded in AI image models, sometimes in unexpected ways. The filter is phrase-specific; writing the words 'syphilis' and 'gonorrhea' separately or in a different order does not trigger the censorship. This suggests the model's developers implemented targeted content blocking, likely to prevent the generation of harmful medical misinformation or explicit content. The discovery has sparked discussion about the transparency of such filters and their potential to inadvertently censor legitimate historical or educational content, as in this case of historical poster recreation.