Nano Banana 2.1

Google's newest image model. Noticeably faster than Nano Banana 2, follows the prompt more closely, and holds on to a person, product, or character across edits and new scenes.

Nano Banana 2.1: Fast Image Generation and Edits That Keep Your Subject

Nano Banana 2.1 is Google’s newest image model. Describe an image and get it, or attach photos and say what to change. It’s noticeably faster than Nano Banana 2, follows the prompt more closely, and holds on to a person, product, or character across edits and new scenes. Layouts, posters, and graphics come out cleaner, styles stick to the reference, and fine detail and textures stay sharp up to 4K.

Use cases

• Generate images from a text brief
• Edit a photo with a plain-language instruction
• Keep a person, product, or character consistent across new scenes
• Combine several reference images into one composition
• Change only a marked area of an image using a mask
• Design posters, social graphics, and layouts with legible text
• Restyle an image to match a reference style
• Make banners and strips in extreme ratios (4:1, 8:1, 1:4, 1:8)
• Illustrate current events, products, or places with web search

Assistant instructions

• Runs as text-to-image when image_urls is empty and as an edit when it holds 1 to 14 images. Put every image the result should draw from into image_urls, in the order the prompt refers to them.
• Write prompt as a full brief: subject, setting, composition, lighting, style. For edits, name the change and what must stay (“Change the jacket to red. Keep the face, pose, and background”). Refer to references by position: “the person in image 1, in the kitchen from image 2”.
• Mask edits: there is no separate mask field. Add the black-and-white mask as an extra image in image_urls and say in the prompt which image is the mask and that only its white area changes.
• Put any text that must appear in the image in quotation marks and describe the type style next to it.
• aspect_ratio: use auto when editing so the output keeps the source shape. Pick an explicit ratio for new images or when the user asks for a different format. 4:1, 8:1, 1:4, 1:8 are for banners, headers, and vertical strips.
• resolution: 1K for drafts and iteration, 2K for finished work, 4K only when the user asks for print or maximum detail.
• num_images: 2 by default; 1 when the user wants a single result or the run is one step of a workflow.
• enable_web_search: turn on only when the image depends on real, current, or specific facts (a recent product, a landmark, today’s weather, a named public event).
• thinking_level: leave unset for most work. Use high for dense layouts, infographics, lots of text, or scenes with many subjects that must be placed precisely; minimal as a light middle ground.
• output_format: jpeg by default; png when the user needs lossless output for further editing.
• Keep each edit to one or two changes. For a series of changes, chain runs and feed each result back in as image_urls.

All AI models & tools

See pricing and start creating on Mitte