Subject first, always
gpt-image-2 front-loads attention. The first clause owns composition. Put the subject anywhere past word 15 and first-try hit rate drops ~30 points.
Independent community tool · not affiliated with OpenAI.
▲800+ TESTS · FIRST-TRY 7/10
The order matters. In 800 side-by-side tests this template hit first-try 73% of the time: Subject · Text · Scene · Lens & Light · Style · Constraints. Everything else is finishing touches.
▲THE SIX PARTS
gpt-image-2 front-loads attention. The first clause owns composition. Put the subject anywhere past word 15 and first-try hit rate drops ~30 points.
Strong text rendering only kicks in when the model knows the exact string. Quotes are the signal; bare text gets paraphrased.
Give environment, time of day, weather, materiality. Specific beats generic every time.
"Shot on 50mm at golden hour, shallow DOF" > "cinematic". Technical cues produce technical images.
Reference an era, photographer, or artistic movement. One anchor is stronger than three hedged adjectives.
"no watermark, 3:2 aspect, no extra signage" — last in the prompt so it's freshest in the model's attention.
Independent third-party tool · Not affiliated with OpenAI.