One Prompt, Different NightCafe Models: What Actually Changes?

The witch, castle, book, and cat appeared in all three images. Their relationships were another matter.

An elderly witch stands beside an enormous open spellbook. A miniature castle rises from its pages, a black cat watches from the table, and a brass lantern supplies warm light beside a glowing blue crystal.

That was the scene I asked three NightCafe models to create. Each result included the main subjects, but the images differed in more than their castle designs. The witch’s expression changed, the cat looked in different directions, and the miniature world connected to the book differently.

Those differences gave me something specific to examine. Instead of choosing only by appearance, I could compare what I requested with what each image actually showed.

Three NightCafe outputs based on the same scene description. The replacement SeedDream prompt also included “Square format.”

What I Compared

I compared images recorded as Flux 3, GPT Image 1.5, and SeedDream 4. Those names follow my supplied file labels. The comparison uses one selected output from each, with an additional SeedDream attempt because its first image was portrait.

For the SeedDream replacement, I added “Square format” to the prompt. The three images used below are square, but this was not…