Start from a photograph instead of a blank canvas. Upload a source image, describe what should change, and let a Nano Banana model rebuild it while the original framing, pose and product geometry stay where they are. Three tiers are registered here, and you pick the one that fits the job.
Three tiers, all from Google · 14 reference images on two of them · Tier chosen by you, never routed
Every number below is read from this site’s own model configuration, not copied from a marketing page. Google registers four Nano Banana tiers; this workbench runs three of them, and Nano Banana 2 Lite is not among them. Google also publishes an output size this workbench does not offer. Both gaps are stated separately rather than merged into one list.
The entry tier. Output is fixed at 2K with no size selector at all. Eight frames are available, covering 1:1, 2:3, 3:2, 3:4, 4:3, 16:9 and 9:16 plus an automatic frame. A single prompt accepts up to 5 reference images. One credit per generation.
Fits:single-image retouching where the frame is already decided
The middle tier, and the one with the widest frame list here: fifteen options, including the tall and wide 1:4, 4:1, 1:8 and 8:1 extremes plus 21:9. Output is selectable at 1K, 2K or 4K and defaults to 1K. Up to 14 reference images per prompt.
Fits:edits that mix a subject shot with several reference plates
The top tier. Same 1K / 2K / 4K selector and the same 14-image reference ceiling, with eleven frames. Google positions this tier for the most complex visual tasks and describes it as the best model for complex and multi-turn image generation and editing.
Fits:long edits that run over several conversational turns
Image-to-image is a different contract from text-to-image. Instead of asking a model to invent a scene, you hand it pixels you already approved and ask for a change inside them. Google’s own wording for this mode is narrow and specific: provide an image and use text prompts to add, remove or modify elements, change the style, or adjust the color grading. That list is short on purpose. Framing, perspective and the subject’s placement are meant to survive the edit.
The practical consequence is that the source image carries most of the art direction. You do not re-describe the lighting you already shot, the wardrobe you already styled, or the logo placement you already approved. You describe the delta. This is why the same workflow that retouches a product packshot also handles seasonal variants, background swaps for regional campaigns, and style transfers where the original composition has to be recognisable afterwards.
This workbench runs three Nano Banana tiers from Google. The tier is a choice you make in the panel, not something the system assigns, so the difference between a 5-image reference ceiling and a 14-image one stays visible while you work. Image-to-image and text-to-image share one entry point and one credit balance, which means a draft generated from a description can be edited in the next step without moving to another tool.
These are not invented prompt formulas. Google documents seven named editing patterns with a fill-in-the-blank template for each, and they map cleanly onto the edits people actually run. The templates are reproduced here so you can lift them directly.
Template: “Using the provided image of [subject], please [add/remove/modify] [element] to/from the scene. Ensure the change is [description of how the change should integrate].” Google notes that the model will match the original image’s style, lighting and perspective, so the integration clause is where you describe how the new element should sit in the existing light.
Template: “Using the provided image, change only the [specific element] to [new element/description]. Keep everything else in the image exactly the same, preserving the original style, lighting, and composition.” Google calls this semantic masking, and the mechanism is conversational rather than a painted mask — you define the region by naming the object, then explicitly lock everything else.
Template: “Transform the provided photograph of [subject] into the artistic style of [artist/art style]. Preserve the original composition but render it with [description of stylistic elements].” The composition clause is doing the real work here; leave it out and the model can redraw the scene instead of restyling it.
Template: “Create a new image by combining the elements from the provided images. Take the [element from image 1] and place it with/on the [element from image 2]. The final image should be a [description of the final scene].” This is the pattern that needs the reference ceiling, which is why the 14-image tiers matter for composite work.
Template: “Using the provided images, place [element from image 2] onto [element from image 1]. Ensure that the features of [element from image 1] remain completely unchanged. The added element should [description of how the element should integrate].” Google’s guidance is blunt: to keep critical details such as a face or a logo, describe them in great detail alongside the edit request.
Template: “Turn this rough [medium] sketch of a [subject] into a [style description] photo. Keep the [specific features] from the sketch but add [new details/materials].” Two clauses again: what to carry over from the drawing, and what to add that was never in it.
Template: “A studio portrait of [person] against [background], [looking forward/in profile looking right/etc.].” Google recommends including previously generated images in later prompts to hold the subject steady, and adding a reference image of the pose when the pose is unusual.
Google’s second prompting rule is positive framing — describe what you want rather than what you want to avoid, with “empty street” instead of “no cars” given as the example. On this workbench that is also the only option available, because the image panel has no field dedicated to exclusions. Write the replacement, not the removal.
Editing someone else’s photograph is not the same as generating a new one, so the rights question deserves a straight answer rather than a slogan.
Read the terms of serviceIn image-to-image the model receives two inputs: an image and an instruction. The instruction is usually written text, and in the workflows Google documents it is conversational rather than parameter-driven — you describe the change, read the result, then describe the next change against that result. Multi-turn conversation is Google’s recommended way to iterate, which is why the editing loop matters more here than a single lucky prompt.
What the instruction can cover is bounded and worth knowing precisely. Google groups editing into adding, removing or modifying elements, changing style, and adjusting colour grading. Alongside those, the documented patterns include combining elements drawn from several supplied images, holding a face or a logo steady while something else changes, and turning a sketch into a finished frame. What is not in that list is the more ambitious claim that any part of an image can be rewritten without consequence.
Subject consistency has a published ceiling: Google states that the models maintain the resemblance of up to five characters and the fidelity of up to fourteen objects within one workflow. That is a real number to design around, because a brand shoot with nine recurring models is outside it, while a product edit with one hero object and a handful of reference plates sits comfortably inside.
Three steps, and the middle one is where the quality lives.
Pick the frame you want to keep. Attach reference images if the edit needs a style, a material or an object that is not in the source. The reference ceiling depends on the tier you select, so decide the tier before you build a long reference set.
Name what changes and, explicitly, what stays. Locking the parts you want untouched is the single most useful habit for this workflow, and it is the mechanism behind Google’s inpainting pattern. Write the target state rather than the excluded one, because the panel has no separate field for negatives.
Review the result, then send the next instruction against it. Google’s own guidance is that multi-turn conversation is the recommended way to iterate, and that critical details should be described in detail alongside the request. Export at the size you selected for the tier.
Each item below is a control or a limit you can verify in the panel itself.
Three Nano Banana tiers sit side by side and you pick one. Nothing is routed automatically, so the reference ceiling and the size selector you get are the ones you asked for.
Nano Banana 2 and Nano Banana Pro accept up to 14 reference images per prompt. The entry tier accepts up to 5, which is a real constraint for composite edits and is stated here rather than averaged away.
Nano Banana 2 and Nano Banana Pro offer 1K, 2K and 4K, defaulting to 1K. The entry tier has no size selector and outputs at a fixed 2K, so “4K” is never an option there.
The frame list runs from 1:1 through 16:9 to 21:9, and the middle tier adds extreme ratios such as 1:4 and 8:1 for strips and banners.
The result stays on the page so the next instruction can build on it, which is the loop Google recommends for iterating on an image.
Image-to-image and text-to-image open from the same workbench with the same credit balance, so a generated draft can be refined without exporting it somewhere else.
Product work is the clearest fit. A packshot that was lit and approved once can be re-lit, re-surfaced or placed against a different environment while the product geometry holds. Because the edit starts from approved pixels, the output keeps the silhouette buyers recognise, which is exactly what a reshoot from a description would put at risk. The same source frame then serves seasonal variants and regional backgrounds without a second photoshoot.
The second cluster is creative iteration. Concept sketches become finished frames, one hero image becomes a small family of look-alikes, and a photograph is restyled into a different artistic language while its composition survives. A third, quieter cluster is correction: removing an object from a frame, fixing a distracting detail, or swapping one element while explicitly locking the rest. All three live on the same panel, and all three benefit from the same habit of naming what must not move.
Open the workbench, upload a source image, and describe the change you want while keeping the rest of the frame locked.
Open the image workbench