These generations were made to help me test Custom Instructions. I set my Custom Instructions to output my exact inputs or minorly edit them for coherence. I wanted to see if that would help Dall-E 3 stay consistent with my desired style (anime in this case). I've noticed that ChatGPT likes to go off-the-rails with Dall-E 3 prompts related to anime-inspired images sometimes. Usually, the outputs don't match the style my inputs specified. This test is to show how using (near) exact prompts without ChatGPT's rewrites can help maintain a style between input and output!
(this is a very limited test and may not be applicable in every case)
Prompt: 1990s anime still of a couple sitting on a bench in the park. The image focuses closely on their faces, showing the details of their emotions as they interact with each other. The wintery scene is evident with snowflakes falling gently around them. The boy, with a hint of blush, is wearing a jacket and jeans, while the girl, with a warm smile, is dressed in a cozy sweater.