Nano Banana 2 Prompts: 50 Edits and Generations (Gemini 3.1 Flash Image)

The useful way to prompt an image model is to define what must remain stable, what should change, what the viewer needs to notice and how you will check the result. A long adjective pile is not a production brief. The templates below separate subject, action, setting, composition, light, constraints and review so you can adapt them to your own work.

First choose the right model and surface

Google’s current naming matters. Nano Banana 2 (Gemini 3.1 Flash Image) is an API model name, while “Nano Banana” is a product-family nickname. The Gemini app, Google AI Studio, Gemini API and Vertex AI may expose different controls, quotas and billing. Select the model shown by the product you are actually using, then confirm its current documentation before promising a resolution, reference count, grounding feature or price.

Use Nano Banana 2 as a practical starting point for general image generation, conversational edits, multiple references, text experiments and video-derived thumbnails. Its documented strengths do not guarantee perfect faces, geometry or small text, so keep an approval step.

A repeatable prompt structure

Subject and reference: say what the attached image is and what must stay unchanged.
Action: name one edit or generation goal instead of several competing changes.
Composition: provide aspect ratio, camera position, crop, focal point and safe areas.
Materials and light: describe surfaces, shadows, direction and colour temperature.
Text: provide exact approved copy, or request a blank panel for later layout.
Exclusions: list objects, claims, logos, identifiers or style constraints to avoid.
Review: ask for a transcription, checklist or uncertainty report when appropriate.

For an edit, start with the invariant list: “keep the face, pose, clothing and crop unchanged.” For a blend, label the inputs: “Image A is the subject; Image B is the background.” For a product, separate the object from the scene: “preserve the package exactly; change only the surface and light.” These small sentences reduce accidental creative decisions.

How to run a safe image workflow

1. Prepare references

Use images you own or have permission to use. Remove unnecessary personal information, school names, addresses, badges and metadata before uploading. If a person is identifiable, get consent for the intended use. Keep a copy of the original and note the model, date, prompt and output version.

2. Start with one variable

Change the background, then the light, then the crop. If you ask for a new outfit, a new pose, a new age, a new location and a new visual style in one request, you will not know which instruction caused drift. One-variable edits also make revisions cheaper and easier to explain.

3. Inspect the result

Check identity, object count, hands, edges, perspective, labels, numbers, logos, reflections, shadows and unintended claims. Generated text can look plausible while being wrong. If a chart or package contains important copy, use the image as a layout concept and place final text with a controlled design workflow.

4. Ask for diagnosis, not applause

When an output misses the brief, ask the model to classify the problem as reference drift, text error, composition, lighting, geometry or missing constraint. Then make the smallest next edit. Keep the successful parts of the previous prompt; rewriting everything can discard useful constraints.

5. Keep provenance

Google documents SynthID on generated images. That is useful provenance, not a substitute for disclosure, consent, copyright review or fact checking. Store the source references, prompt, model ID, output size, editor decision and final approved asset together.

50 Nano Banana 2 prompts

1. Controlled portrait edit

Prompt: Use the attached portrait as the identity reference. Keep the person's face, age, hairstyle and natural skin texture unchanged. Replace only the background with a softly lit warm-grey studio. Crop from the chest up and leave breathing room above the head.

Why it helps: Name the invariant parts before the change.

2. Reference-led outfit change

Prompt: Keep this person recognisable and in the same pose. Change only the outfit to a navy linen jacket over a white shirt. Preserve the face, hands, hair, body proportions and direction of light. Do not add jewellery or logos.

Why it helps: Ask for one controlled change at a time.

3. Four-season character set

Prompt: Create four matching 4:5 portraits from this reference: spring garden, summer beach, autumn park and winter street. Keep the same person, camera height, crop, facial features and pose in every panel; change only season, wardrobe colour and background details.

Why it helps: A repeated invariant list helps series consistency.

4. Pet identity edit

Prompt: Use this dog photo as the sole identity reference. Keep the muzzle, fur markings, collar, eye colour and size unchanged. Place the same dog on a quiet beach at sunrise with believable paws, shadow and wind in the fur.

Why it helps: Never imply the model can guarantee a perfect likeness.

5. Couples scene

Prompt: Combine Images A and B into one respectful couples portrait. Preserve both people's real facial features and relative height. Place them walking beside a rainy city canal at blue hour; match reflections and light, and add no text or invented accessories.

Why it helps: Label each uploaded image by role.

6. Family composition

Prompt: Use the supplied family photo as the reference. Keep every person recognisable, preserve their age and clothing colours, and place them at a simple outdoor picnic. Do not add or remove people, pets, jewellery or identifying signs.

Why it helps: Review every face before sharing.

7. Professional headshot

Prompt: Turn this selfie into a restrained professional headshot. Keep my actual face, hair and body shape. Use soft window light, a pale neutral background and a natural expression. Do not beautify, slim, age, widen or invent credentials.

Why it helps: Good for editing, not identity fabrication.

8. Product on five surfaces

Prompt: Use the attached white ceramic mug as the exact product reference. Create five separate 1:1 scenes: marble, oak, linen, cafe table and windowsill. Keep its shape, colour, handle, logo placement and scale identical; change only surface and light.

Why it helps: Use a single product reference across the set.

9. E-commerce listing

Prompt: Create a clean 4:5 product listing image from this reference. Keep the item and its packaging exact. Use a warm-white background, soft shadow, even light and generous negative space. Do not invent ingredients, certifications, measurements or claims.

Why it helps: Proofread labels against approved artwork.

10. Lifestyle product scene

Prompt: Place this exact skincare bottle on a bright bathroom shelf beside one folded white towel and one small green plant. Preserve the bottle, cap, label placement and proportions. Use morning light and do not add health or performance claims.

Why it helps: Separate visual mood from product facts.

11. Logo mockup

Prompt: Place the supplied logo on a folded heather-grey T-shirt lying on a wooden table. Keep the logo artwork unchanged and readable; show realistic fabric texture, soft daylight and a natural fold. If the logo is unclear, leave it blank rather than redraw it.

Why it helps: Treat generated text as a draft.

12. Packaging correction brief

Prompt: Inspect this packaging concept and return two parts: the visual mockup and a plain-text checklist of every word that needs proofing. Do not claim small text, barcodes, ingredients or legal marks are accurate unless they came from the supplied artwork.

Why it helps: The checklist is often more useful than a confident-looking label.

13. Video thumbnail

Prompt: Use this uploaded video as visual context and create a 16:9 thumbnail for a tutorial. Show the clearest subject, strong contrast, one focal point and empty space for verified headline copy. Do not add a fake quote or unreadable text.

Why it helps: Google documents video-to-image support for Nano Banana 2.

14. Video summary infographic

Prompt: Review the supplied video and make a simple 16:9 visual summary with five empty labelled boxes: problem, context, steps, warning, next action. Use only information visible in the video and keep labels large and separate from the body copy.

Why it helps: Check the generated summary against the source video.

15. Reference blend

Prompt: Image A is the subject; Image B is the background; Image C is the lighting reference. Put the subject naturally into B, borrowing only the light direction and colour mood from C. Preserve the subject's identity and avoid adding unrequested objects.

Why it helps: Explicit roles reduce ambiguous blending.

16. Interior redesign

Prompt: Use this room photo as the structural reference. Keep walls, windows, floor plan and camera position unchanged. Replace only the rug, seating and pendant light with simple mid-century pieces. Keep scale believable and do not remove architectural features.

Why it helps: Lock structure before styling.

17. Room declutter

Prompt: Edit this room photo by removing only the loose boxes, cables and empty cups marked with circles. Preserve the furniture, walls, windows, shadows and perspective. Fill the cleared areas with plausible existing surfaces, not new decor.

Why it helps: Mark removal targets when precision matters.

18. Restoration

Prompt: Restore this scanned photograph by removing scratches, dust and creases while preserving the original faces, composition and period character. Where detail is missing, make the repair conservative and flag uncertain areas in a separate note.

Why it helps: Restoration should not silently rewrite history.

19. Black-and-white colourisation

Prompt: Colourise this archival photo with restrained, period-appropriate tones. Preserve every face, expression, garment shape and background detail. Do not add modern objects or infer a precise colour when the source cannot support it.

Why it helps: Call uncertain colour choices out for review.

20. Readable process diagram

Prompt: Create a clean 16:9 process diagram with exactly five stages: Brief, Draft, Check, Approve, Publish. Use high-contrast typography, one simple icon per stage and no extra words. Leave a margin around every label.

Why it helps: Short labels are easier to verify.

21. Data chart concept

Prompt: Turn the supplied verified table into a simple editorial chart. Preserve every number and category exactly. Use a clear legend and accessible contrast. If any label cannot be rendered confidently, leave an empty label box instead of guessing.

Why it helps: Rebuild final charts from data when precision matters.

22. Recipe card

Prompt: Design a vertical recipe card using only this supplied ingredient list, quantities and steps. Keep every number exact, use large readable type and reserve a blank footer for a human editor. Do not invent nutrition or preparation claims.

Why it helps: Numbers need a manual check after generation.

23. Map-style route

Prompt: Create a simple illustrated route map from the supplied place names and directional connections. Use no external geography, logos or unverified distances. Include a blank note area for a human to add checked travel information later.

Why it helps: Do not let visual plausibility become factual evidence.

24. Editorial illustration

Prompt: Illustrate a librarian helping a student organise books in a bright community library. Use flat editorial shapes, warm paper texture, inclusive characters, accessible contrast, no logos and no readable text.

Why it helps: Describe the audience and visual constraints.

25. Illustrated explainer

Prompt: Create a four-panel explainer about checking an AI answer: ask, inspect sources, compare evidence, approve. Use one consistent character and the exact four short labels. No extra claims, speech bubbles or decorative words.

Why it helps: One visual metaphor per panel keeps it clear.

26. Social crop set

Prompt: Create matching 1:1, 4:5 and 9:16 crops from this product reference. Keep the product scale and identity consistent, place the focal point inside every safe area and leave the top third empty for approved copy.

Why it helps: Design for crops before adding text.

27. Film still mood

Prompt: Create a cinematic 3:2 editorial photograph of a quiet Lagos street after rain, warm shop lights reflected in the road, natural human proportions and restrained colour. Avoid readable brand names and invented signage.

Why it helps: Location mood is not a claim about a real street.

28. Architecture concept

Prompt: Create three concept views of the same small library: street elevation, courtyard and reading room. Preserve the supplied dimensions and material palette; label each image as concept only and do not imply planning approval or structural feasibility.

Why it helps: Separate ideation from professional documentation.

29. Typography-first poster

Prompt: Create a minimalist poster with exactly this headline: [paste approved headline]. Use a single bold type treatment, high contrast and generous margins. Render no other words; return a separate transcription for manual comparison.

Why it helps: Never publish generated text without checking it.

30. Colour palette variants

Prompt: Create four versions of this same illustration using only these palettes: charcoal and lime, navy and coral, cream and forest, plum and gold. Preserve composition, subject count, line weight and all approved text.

Why it helps: Constrain the variable you are exploring.

31. Object removal

Prompt: Remove only the red suitcase beside the chair. Keep the chair, floor, wall, lighting, perspective and shadows unchanged. Fill the area with the existing floor pattern. Do not add another object.

Why it helps: Specific removal instructions beat 'clean this up'.

32. Background replacement

Prompt: Replace only the background behind the subject with a softly blurred botanical garden. Keep the subject's hair edges, skin texture, clothes, pose and lighting direction unchanged. Match the depth of field without changing identity.

Why it helps: Edges and light are the quality checks.

33. Fashion lookbook

Prompt: Create three lookbook images from this garment reference: studio, street and cafe. Keep the garment's cut, colour and details consistent. Use diverse adult models, natural poses and no logos or unverified lifestyle claims.

Why it helps: Use the reference to anchor the product, not the model's imagination.

34. Food editorial

Prompt: Create a top-down editorial image of this supplied dish on a neutral table with one linen napkin and soft side light. Preserve the food's actual ingredients and portion appearance; do not add health claims or unprovided garnishes.

Why it helps: Visual styling must not change the recipe facts.

35. Infographic from notes

Prompt: Convert these verified notes into a five-part infographic. Use only the supplied wording, one idea per panel, large labels and a clear reading order. Flag any note that is ambiguous instead of resolving it creatively.

Why it helps: Source fidelity matters more than decoration.

36. Child-safe story scene

Prompt: Use this child's photo only as a private family reference. Keep the child fully clothed and recognisable, place them in a gentle illustrated storybook scene, and add no school, location, name or identifying text.

Why it helps: Obtain consent and avoid unnecessary identifiers.

37. Pet story scene

Prompt: Use this pet photo as the reference. Keep markings, muzzle, collar and size unchanged. Place the animal in a softly illustrated garden story scene with no unsafe food, dangerous situation, owner name or location clue.

Why it helps: Cute does not remove safety and privacy checks.

38. A/B creative directions

Prompt: Create two distinct campaign visual directions for this exact product: quiet editorial and bright playful. Keep the product unchanged and list the visual differences separately. Do not add a price, review, guarantee or performance result.

Why it helps: Ask for alternatives without inventing evidence.

39. Meeting visual

Prompt: Turn these supplied meeting themes into a neutral one-page visual summary with three columns: decisions, open questions, next actions. Use the exact names and dates supplied; leave missing owners blank.

Why it helps: Blank is safer than a guessed owner.

40. Document cover

Prompt: Create a clean report cover using this exact title, subtitle and date. Use a restrained abstract visual, high contrast and generous margins. Render no additional text, logos or statistics.

Why it helps: Add final text in a design tool when possible.

41. Product comparison layout

Prompt: Create a three-column comparison mockup for the supplied objects A, B and C. Keep each object exact, equalise scale, use labels A/B/C only and leave the feature rows blank for verified copy.

Why it helps: A layout is not a completed comparison.

42. Visual variation

Prompt: Generate five compositions from this same subject using these camera positions: eye level, overhead, low angle, three-quarter and close crop. Preserve subject identity, object count and colour palette; change only camera position.

Why it helps: Controlled variation makes selection easier.

43. Prompt refinement

Prompt: Rewrite the following visual brief into a concise image prompt with sections for subject, action, setting, composition, light, constraints and review checks. Do not add facts or visual elements missing from the brief: [paste brief].

Why it helps: Use the model to structure a brief before generating.

44. Output review

Prompt: Review this generated image against the brief below. Return a table with requirement, visible evidence, uncertain or failed, and suggested next edit. Do not praise the image generally or infer facts that are not visible.

Why it helps: A review prompt is not a substitute for a human decision.

45. Text transcription check

Prompt: Read the text in this image and return an exact transcription, preserving line breaks where possible. Mark unclear characters with [unclear]. Do not silently correct spelling, punctuation, numbers or names.

Why it helps: Compare transcription with the approved source.

46. Localisation edit

Prompt: Translate only the supplied poster text into [language], preserving the hierarchy, tone and meaning. Keep names, numbers, URLs and brand terms exactly as provided. Return the translated text separately for review.

Why it helps: Localisation still needs a fluent human reviewer.

47. Batch naming sheet

Prompt: Create a contact-sheet concept for these eight supplied product references. Use labels 01 through 08 only, consistent crops and a neutral background. Do not invent product names or features.

Why it helps: Use index numbers to prevent naming drift.

48. Accessibility concept

Prompt: Create a high-contrast educational visual for a reader with low vision. Use large simple shapes, strong contrast, no tiny text and one short approved label per section. Return an accessibility review checklist separately.

Why it helps: Design constraints are part of the prompt.

49. Safe creative edit

Prompt: Restyle this owned photograph as a paper-cut collage while keeping the subject recognisable and the composition intact. Remove all brand marks and identifying text. Do not imitate a living artist's exact signature style.

Why it helps: Describe visual properties instead of naming a living artist.

Prompt adaptation examples

For a thumbnail, keep the subject simple and reserve text space: “show the clearest moment, one focal point, high contrast, 16:9, empty right third for approved headline.” For a catalogue image, lock the object: “same product, exact shape and label placement, neutral surface, soft shadow, no claims.” For an infographic, do the opposite of improvisation: supply the verified labels, numbers and units, then ask the model to leave unclear items blank.

For a portrait, define dignity and identity: “preserve the consenting adult’s actual face, age, skin texture and body shape; change only the background.” For a multi-image composition, state which reference controls which property. For a series, repeat the invariant list in every prompt and compare the outputs side by side. Consistency is a process, not a single magic phrase.

Limitations worth planning for

Exact identity, geometry and small text can drift even when the prompt is specific.
A generated image may add plausible objects, reflections or claims that were never supplied.
App and API controls are not automatically identical; access, quotas and prices can change.
Grounding can help with current context, but sourced facts still need editorial checking.
Image generation is not a substitute for engineering, medical, legal, financial or accessibility review.
Do not upload private or sensitive images unless the people involved understand and consent to the workflow.

FAQ

Is Nano Banana 2 the same as Nano Banana Pro?

No. Google documents them as different models: Nano Banana 2 is Gemini 3.1 Flash Image and Nano Banana Pro is Gemini 3 Pro Image. Their controls, positioning and pricing can differ.

Can these prompts guarantee readable text?

No. They can constrain the request and ask for a transcription, but important copy should be checked against an approved source and often composited outside the model.

Can I use these prompts in the Gemini app?

They are written as portable briefs. The app, AI Studio, API and Vertex AI may expose different upload, model, size, grounding and billing controls. Adapt the prompt to the surface you are using.

How should I compare outputs?

Define acceptance criteria before generation: identity preserved, object exact, text correct, composition usable, no unsupported claims and suitable output size. Compare against those criteria rather than choosing the most attractive image by instinct alone.

For a broader library, see Nano Banana prompts, Gemini image prompts and AI image prompts.