Why Keyword Prompts Underperform Here
Gemini's image model was built for language-first interaction: it reads full sentences with syntax and referents, not weighted keyword lists. "A 35mm portrait of an elderly fisherman mending a net at dawn, warm side light, shallow depth of field" outperforms the same content chopped into comma fragments because the model uses the sentence structure to bind attributes to the right objects. The generator writes prose on purpose — it's the native format.
The Edit Prompt Formula: Change X, Preserve Y
The most common editing failure is under-specification: ask to "make the jacket red" and the model may also repaint lighting, skin tone, or the background, because nothing told it those were off-limits. Effective Nano Banana edit prompts name the change and then pin everything else: "change the jacket to deep red; keep the face, pose, lighting, and background exactly as they are." Every edit-style prompt this tool writes carries that second clause.
Consistency and Typography, Its Two Superpowers
Two things set Nano Banana apart from most image models. First, identity consistency: it can carry the same person or character through outfit changes, new scenes, and different poses when the prompt refers to "the same person" explicitly. Second, text rendering: it produces legible, correctly-spelled signage and titles far more reliably than diffusion models — put the exact wording in quotes. The generator uses both patterns whenever your idea calls for them.