How to use Nano Banana for image generation and editing

Define the job in one sentence: create a product hero, restyle a room, adapt a portrait, build a social graphic, or explore a character direction. That sentence becomes the priority for every later instruction. If the request contains several deliverables, generate them separately so each composition has a clear review standard.

For a new scene, describe the subject before style. For an edit, identify the source image and say which elements must be preserved. Phrases such as keep the package silhouette, preserve the face, or retain the camera angle are more useful than asking the model to make the image similar without explaining the protected features.

Reference quantity is not a quality setting. A small, coherent set usually communicates identity and art direction better than many unrelated images. Remove screenshots with irrelevant interface elements, crop around the useful subject, and avoid including conflicting versions of the same product unless the prompt clearly says which version is authoritative.

CreateForge uploads references through its private authenticated flow and hands temporary provider access to the generation task. The references are inputs, not public gallery assets. Still, only upload material you have the right to process, and avoid placing confidential information in an image when it is not needed for the creative result.

• Use one reference for identity and another for style only when both roles are necessary.
• Describe every reference role explicitly instead of expecting file order to explain the brief.
• Crop distracting borders, controls, and unrelated background material before uploading.

Pick the output frame before generation. A vertical story image benefits from foreground-to-background depth, while a wide hero needs deliberate horizontal balance and space for neighboring interface or copy. Selecting a ratio changes the canvas; describing the same framing in the prompt helps the model populate that canvas intentionally.

Choose the available output format according to the next step. PNG is useful when a lossless working file matters, while JPEG is often more compact for review and delivery. The current control labels and quote on the Nano Banana model page are the source of truth because provider capabilities and prices can change after an article is published.

Review the image at full size and separate factual failures from taste. Wrong object count, lost identity, broken text, and an unusable crop are factual problems. Color mood or camera energy may be preferences. Correct the factual issue first and preserve the instructions that already worked, then make a second pass for art direction.

Successful output is registered in the private CreateForge Library after the provider task and storage step complete. Compare saved versions side by side, note the change between them, and stop when the asset meets the original requirement. More variation is not automatically more progress when the delivery decision has already been made.

When the first image is close, freeze the successful parts in writing before changing anything. Note the approved camera angle, subject placement, background, palette, material response, and negative space. Then alter one variable, such as the surface finish or the amount of headroom. If the prompt changes composition, lighting, styling, and copy at the same time, the next output cannot show which instruction improved the result and which instruction caused a regression.

Use a simple review log for repeat work: attempt number, prompt change, selected controls, visible result, and decision. The log does not need private customer text or a long narrative. Its purpose is to reveal when repeated low-cost attempts have become more expensive than switching models or repairing an accepted image manually. A practical Nano Banana workflow optimizes for the first usable asset, not for the largest number of visually different generations.

Before publishing, place the accepted image in its real layout and check the crop at desktop and mobile sizes. Confirm that important details remain visible behind headings, buttons, safe areas, and platform overlays. Recheck logos, small text, subject edges, and contrast after compression rather than approving only the large Library preview.

Keep the durable Library result as the source and make delivery derivatives from that file. Name exports by campaign, placement, ratio, and revision so reviewers do not approve an obsolete version. This final discipline is simple, but it prevents a successful generation from becoming an avoidable production or handoff error.