Turn the vague request “beautiful mug” into a structured prompt. Return JSON ONLY: {"prompt":"...","style":"...","details":"..."} - in prompt collect the subject, environment and light in one phrase, in style indicate the visual style, in details add material, angle and format.
The image generator does not draw like a human. It is trained on millions of “picture + caption” pairs and remembers what objects, styles and light look like. When you provide text, the model starts with random noise and removes it step by step, bringing the image closer to the meaning of the description. This has two consequences. First: the more specific the words, the less the neural network thinks for you - it fills out a vague request with the “average Internet user.” Second: the start is always from random noise, so the same prompt gives different pictures. This is not a bug, but a way to get options. Working levers: anchor nouns (what exactly is in the frame), adjectives of material and light, indication of style and format. Insider technique: write the prompt in “layers” - subject, environment, light, style; The model reacts more strongly to the first words, so put the main thing at the beginning. Second trick: if you like the picture, fix its seed and change one word at a time. A typical beginner mistake is to load emotions (“beautiful, wow”) instead of attributes. The model does not know your “beautiful”; she knows “soft diffused light”, “ceramics”, “close-up”.