An image model can render an effect description literally. In the taxi short, the prompt 'mushroom-shaped explosion' produced an actual mushroom with smoke instead of an explosion.
Fix
Check Look at every effects sheet for literal readings of metaphors before the sheet is used as a reference. The same literal-reading habit is behind Audit the references before rewriting the prompt: when the output is wrong, look at what the model was actually given.
Descriptive effect wording can be taken literally. "Mushroom-shaped explosion" produced an actual mushroom with smoke. Rewording it as "nuclear blast" mostly fixed it. Replace shape metaphors with the name of the real phenomenon.