What you want to see, written as a plain description.
Your words are converted to numbers and given to the model on every pass, not only at the start. That is why every part of the prompt affects the whole picture.That is also why order matters. Words near the start have the most effect, so begin with the subject. A workable order is:
the subject;
the attributes that make it particular;
where it is;
how it was made.
Written out, that is “a red fox asleep on a mossy log, morning light, 35mm photograph”.Concrete nouns and adjectives work better than mood words. The training images were captioned with things like “mossy log”, so the model has an average for that; almost nothing was captioned “epic”.If you do not name the background, the lens or the colours, they are still decided: the model uses whatever was most common in its training images. Name anything you care about.Which model you are on decides those defaults, which is why one prompt gives a different picture on two checkpoints. A checkpoint is one trained model file.