Precision object editing lets you modify specific elements within an image using natural language instructions — "replace the red car with a blue truck" or "remove the person in the background" — without manually drawing selection masks or layer boundaries. The model internally identifies the target region, reasons about how the edit should look given the surrounding context (lighting, perspective, shadows), and composites the change seamlessly. This contrasts with mask-based editing (Photoshop generative fill) where you must outline the target area first, and with full-image regeneration where the entire image is re-generated.
Describe the object to edit in plain English: "the coffee cup on the left table" or "the red car behind the fence." Nano Banana Pro uses Gemini 3 Pro's language understanding to locate the exact object — no coordinate input, no bounding box, no mask.
The model uses a chain-of-thought reasoning step before editing: it identifies the target object, analyzes the surrounding context (lighting angle, perspective, shadow direction), and plans the edit to match. This reasoning step reduces physically impossible edits like mismatched shadows or broken perspective.
Three edit types: replace an object with something else ("swap the lamp for a candle"), remove an object entirely ("remove the watermark"), or modify an attribute ("make the car blue instead of red"). Each operation preserves the untouched portions of the image pixel-for-pixel.
When editing images with people, Nano Banana Pro maintains identity consistency for up to 5 characters and 14 distinct objects in the scene. Edit one character's outfit without affecting the others' appearances.
Edits respect the image's existing lighting, perspective, and material properties. Replace a wooden table with a glass one and the reflections, shadows, and surrounding objects update to match the new material physically.
Go to Sunra Image and select Nano Banana Pro from the model dropdown. Switch to edit mode.
Upload the image you want to modify. Supported formats: PNG, JPG, WebP. Higher resolution images give the model more context for precise edits.
Describe the edit in natural language. Be specific about the target and the desired change: *"Replace the red backpack on the chair with a brown leather messenger bag. Keep everything else the same."*
Click Generate and compare the result with the original. Zoom in on the edited area to check for edge blending, shadow consistency, and perspective accuracy.
If needed, make additional edits on the result — change another object, adjust a color, or refine the first edit. Each edit builds on the previous output.
Whether you're a solo creator, an agency, or a brand — every model adapts to how you work.
Studio editorial portrait
Replace the plain gray studio backdrop with a floor-to-ceiling window overlooking a rainy cityscape. Match the soft key light already on the subject. Keep the model's pose, outfit, and expression unchanged.
Industrial grunge editorial
Remove the graffiti tags on the concrete wall behind the subject. Fill with clean raw concrete texture matching the existing grain and weathering. Do not alter the person or their clothing.
Rainy Shibuya night portrait
Replace the umbrella the person is holding with a translucent neon-pink umbrella that casts a soft pink glow on their face. Keep the rain, reflections, and Shibuya background unchanged.
Replace the white sneakers with the same model in matte black. Keep the laces white. Maintain the same lighting, angle, and surface texture. Everything else in the image stays identical.
Model: Nano Banana Pro · Mode: Edit · Reasoning: Thinking
Remove the water bottle from the desk. Fill the area with the same wooden desk surface, matching the grain direction and lighting. Do not change anything else in the image.
Model: Nano Banana Pro · Mode: Edit · Reasoning: Thinking
Replace the potted succulent on the windowsill with a small bonsai tree. Match the pot style to a dark ceramic. The bonsai should cast a shadow consistent with the window light direction.
Model: Nano Banana Pro · Mode: Edit · Reasoning: Thinking
Change the woman's hair color from blonde to dark auburn. Keep the hairstyle, length, and texture exactly the same. Do not change her face, clothing, or any other part of the image.
Model: Nano Banana Pro · Mode: Edit · Reasoning: Thinking
Generate product color variants from a single photo shoot. "Change the sofa upholstery from gray to navy blue" or "swap the wooden legs for chrome." No re-shooting, no 3D rendering — just describe the variant.
Modify staged photos without re-staging: swap furniture pieces, change wall paint colors, or remove personal items. Each edit preserves the room's lighting and perspective so the result looks photographed, not composited.
Adapt marketing images for different markets: "Replace the English sign with Japanese text" or "swap the local currency symbol." Edit specific text and signage without regenerating the entire scene.
Designers can rapidly explore variations: try different furniture in an interior, test color schemes on a product, or experiment with wardrobe options on a fashion shoot — all from a single source image.
| Nano Banana Pro (NL Editing) | Other Editing Tools | |
|---|---|---|
| Target selection | Natural language — describe the object to edit, no mask needed | GPT Image 2: also NL-based. Photoshop generative fill: requires manual mask drawing |
| Reasoning before editing | Gemini 3 Pro "Thinking" mode — chain-of-thought plans context-aware edits | GPT Image 2: implicit reasoning. Photoshop: no reasoning — fills mask region independently |
| Character consistency | Maintains up to 5 characters and 14 objects — edits one without affecting others | GPT Image 2: strong multi-turn editing. Photoshop: no character awareness |
| World knowledge for replacements | Google Search grounding — knows what specific products, brands, and objects look like | GPT Image 2: strong general knowledge. Photoshop: limited to training data |
| Best for | Precise single-object edits with physical accuracy (shadows, reflections, perspective) | GPT Image 2: multi-turn iterative editing sessions. Photoshop: pixel-precise manual control |
"The cup" is ambiguous if there are multiple cups. Use distinguishing details: "the blue mug on the left side of the table" or "the person wearing the red jacket." Ambiguous targets may cause the model to edit the wrong object.
Explicitly protect important elements: "Do not change the background" or "Keep the person's face and hairstyle unchanged." Without constraints, the model may make broader changes than intended, especially on complex scenes.
If you need to change 3+ objects, edit them one at a time. Each edit builds on the previous result, giving you a checkpoint to verify before proceeding. Batching too many changes in one instruction reduces precision.
If an object is fully occluded (hidden behind something else), the model has no context for what it looks like from this angle. Edits work best on visible objects where the model can analyze shape, lighting, and context from the existing pixels.
Join thousands of creators, agencies, and brands who use Sunra every day.
