Don't stop here
Hand-picked guides our readers explore right after this one.
Stunning image generation with Midjourney prompt mastery
Read the guideExpert guide to Claude prompts with XML tags, artifacts, and complex reasoning
Read the guideAI prompts for bookkeeping, tax planning, financial analysis, and audit preparation
Read the guideNano Banana is Gemini's image model (the Pro version is Gemini's image generation model), and its headline strength is keeping a consistent likeness across edits. When it misbehaves, the usual complaints are: it changes your subject's face or details when you only wanted the background swapped, it garbles text in the image, it ignores part of a multi-part prompt, or it just returns the original image unchanged. Most of this traces to one thing: overloaded prompts. Nano Banana is far more reliable when you make one focused change at a time and explicitly tell it what to keep. There is also a real web-UI vs API behavior gap some users hit, and stricter refusals around real people. This guide covers the consistency, text, and ignored-edit failures and the prompt patterns that fix them.
The subject's face or key details change when you only wanted one part edited
Text in the image comes out garbled or misspelled
It returns the original image basically unchanged
It applies some of your requested edits but ignores others
Results are inconsistent between the Gemini app and the API
It refuses to edit photos of real or identifiable people
Asking for several changes at once (swap background, change outfit, fix lighting, add text) in a single prompt is the top cause of inconsistent results. The model tries to do everything and drifts on the parts you wanted preserved.
If you do not tell Nano Banana what to hold constant, it may re-interpret the whole image. Naming what must stay unchanged (the face, the pose, the product) anchors the edit.
Image models still struggle with in-image text. Longer strings and unusual fonts garble more often. Short, simple text placed with a clear instruction works best.
Some users see the same edit behave differently in the Gemini app versus the API, particularly locking a foreground subject while changing only the background. It is a known behavior gap, not just your prompt.
Editing photos of real or identifiable people (especially public figures) is restricted, so a refusal or a mangled result can be the safety layer rather than a quality failure.
When to try: First, for any inconsistent or ignored-edit problem
Break complex edits into steps: change the background first, confirm it, then change the outfit, then adjust lighting. Single-focus edits are far more predictable and let you keep the good result at each stage instead of losing it in one overloaded prompt.
When to try: Whenever the subject drifts during an edit
Tell it what must not change: 'Keep the person's face, pose, and clothing exactly the same, change only the background to a beach.' Naming the elements to lock is the single biggest lever for consistency.
When to try: For character/product consistency across edits
Nano Banana can keep key elements of an image unchanged while editing the rest. Lean on that: reference the original subject and instruct it to maintain the same likeness, rather than describing a brand-new image.
When to try: When text comes out wrong
If you need text in the image, keep it short, spell it out in quotes, and place it with a clear instruction ('add the text "SALE" in the top-left'). Expect to retry, long strings and fancy fonts are where garbling happens.
When to try: When edits are ignored entirely
If it hands back an unchanged image, re-state the edit as a clear command referencing the uploaded photo ('In the attached image, replace the sky with a sunset'). Re-uploading and giving one unambiguous instruction usually unsticks it.
When to try: For app-vs-API gaps or real-person edits
If the app and API diverge on an edit (for example locking a foreground subject), try the other surface. And for photos of real people, expect stricter behavior, use stylized or fictional framing to stay within policy.
One focused edit per prompt, then build on the result
Always name what to keep unchanged, not just what to change
Keep any in-image text short and quoted
Expect restrictions on real, identifiable people and frame accordingly
For persistent quality or consistency problems, report them through the Gemini app's feedback (thumbs down, with detail), which is how Google tracks model issues. Contact support only for account or access problems. Model quality quirks are addressed through updates, not per-user support tickets.
Almost always because the prompt did not tell it what to keep. Add an explicit instruction like 'keep the person's face, pose, and clothing exactly the same, change only the background', and make that one change on its own rather than bundling several edits.
In-image text is still a weak spot for image models, and longer strings and unusual fonts garble more. Keep text short, put it in quotes, place it with a clear instruction, and retry, short simple text is the most reliable.
Re-upload the photo and give one unambiguous command that references it, for example 'in the attached image, replace the sky with a sunset'. One clear instruction on a freshly attached image usually unsticks an ignored edit.
There is a known behavior gap between the web UI and the API for some edits, particularly locking a foreground subject while changing the background. If one surface drifts, try the other, it is not only your prompt.
Editing real or identifiable people, especially public figures, is restricted by policy, so you may get a refusal or a deliberately altered result. Use stylized or fictional framing, and describe attributes rather than naming a specific person.