I take this as a question about English usage within the constraints of scientific convention. Like some other questions on this site it relates to concise usage, lack of repetition, and clarity of expression. What follows is therefore good practice relating to most figures (but not all) and is not merely a matter of opinion.
Figures are usually shown to present or summarize those data or concepts that are described in the article's prose. They may illustrate examples of things alluded to in the prose. A caption should not duplicate the prose at any length or unnecessarily; this would waste space and would irritate some readers by diversionary repetition of ideas that are already systematically developed in the prose.
The caption should be sufficient to give immediate and concise (even if incomplete) understanding of the core significance of the figure. It should provide a succinct conceptual link between figure and prose by drawing attention to the main theme it illustrates that is developed in more detail within the prose.
Similarly the caption should not repeat unnecessarily any information that is given within the figure (such as the names of the variables already shown on the axes of a simple graph, an embedded title, or the embedded legends).