Nano Banana negative prompt: describe what you want instead

Nano Banana has no negative-prompt field to fill in; Google's advice is to say what you want in positive words. Its Gemini image guide recommends "semantic negative prompts": instead of "no cars", describe the intended scene, such as "an empty, deserted street with no signs of traffic". Sume's image request matches that: POST /v1/images takes a required prompt string and no negative-prompt field, so the exclusion goes in the sentence.

Google's tip is from its Image generation with Gemini page and Sume's request fields from the Sume Image API page, both read 2026-10-02.

What is a semantic negative prompt?

It is Google's term for an exclusion written as a description of the thing you do want. The guide's example turns "no cars" into a deserted street with no signs of traffic. The reasoning is in the page's list of best practices, next to being hyper-specific, giving context and intent, and iterating with follow-up prompts.

Does the Sume image request have a negative field?

Not in the Image API request table. Its fields are model, prompt, n, resolution, aspect_ratio, size, quality, output_format, mask_url, background, output_compression, seed, stream, input_references, the provider object, metadata, mode, webhook_url and wait_timeout_seconds. The docs say a parameter the selected model does not list is rejected with 400 unsupported_parameter rather than silently dropped, so a made-up negative_prompt field is a request you will have to fix.

How do I handle a stubborn extra object?

Check the wording first: is the thing mentioned in your prompt at all, even as a "no"? Remove it and describe the empty space instead. If it persists, make an edit pass with the first image as an input_references entry and a short instruction naming what should be left clear. Google's best practices also suggest iterating with follow-up prompts rather than expecting a perfect first image.

Does this apply to other image models on Sume?

The Sume request is the same for every catalog model, so the missing field is not specific to Nano Banana. Google's tip is about Gemini, and this post does not claim it holds for other vendors' models; for those, read the vendor's own prompt guide, then test.