Ideogram 4 safety filter bypass: Describe the scene, not the clothing

A viral Reddit post reveals the mechanics behind Ideogram 4's safety filter for swimwear and beachwear. The author tested a ladder of prompts using clean JSON—'woman in a bikini at a pool,' 'lace lingerie,' 'wrapped in a sheet,' even 'fine-art nude from behind'—and every one was blocked, including the standard bikini prompt. The issue isn't the model's pixel-level classifier but keyword detection: explicitly naming a flagged clothing item triggers the filter, even for completely tame, non-explicit generations.

The fix is simple: replace clothing nouns with situational descriptions. Instead of 'a woman in a bikini,' use 'a cheerful young woman having fun at the beach on a sunny summer day' or 'enjoying a hot day at a resort pool.' The model infers context-appropriate attire (e.g., a bikini) without the trigger words, achieving a 4/4 pass rate in the test. This is not a jailbreak—Ideogram 4's post-training weights still suppress explicit anatomy. The method only corrects false-positive blocks for standard beachwear, letting the model naturally render what users wanted all along.