Search before you style the subject.

Nano Banana 2 can retrieve visual context before it renders. The prompt works better when evidence and art direction have separate jobs and separate acceptance checks.

When the subject has facts, research them before styling them. Ask Nano Banana 2 for cited observable traits, freeze those traits as requirements, then compose an original frame around them.

Google's current Gemini image documentation describes Grounding with Google Search for Nano Banana 2, technically Gemini 3.1 Flash Image. The copied prompt alone does not switch search on: the Google Search tool and its search types must be enabled in the API request or supported surface. The model can then use Web Search and Image Search as context before generating. Google says Image Search grounding is available only for this image model. The distinction matters because Nano Banana 2 Lite is a different, faster model and does not carry the same capability.

This is useful when the subject has observable facts that style words can easily overwrite: a species, artifact, historical garment, machine, landmark, or regional material. The model can retrieve context, but retrieval is not art direction. It answers what should be true. The creative brief still decides what the new image should do. Put both jobs in one paragraph and the strongest mood word may quietly push a factual trait out of the frame.

Google's current Gemini image guide documents Grounding with Google Search using Web and Image Search for Nano Banana 2, distinguishes model support, and explains that applications must display returned search suggestions. Grounded response steps carry inline URL citations and attribution metadata.

Generate and edit images with Gemini — Google AI for Developers

Write the grounding job as a list of five visible questions, not research this subject. For a bird, ask for crown and throat color, breast color, bill color, wing pattern, and tail-feather proportion. For an object, ask for silhouette, joining method, surface material, period-correct hardware, and scale. A closed list tells the model what evidence must return and gives you a finite set of traits to inspect in the generated image.

Require the trait list in an accompanying text block with inline source links, and preserve the returned search-suggestion attribution metadata. The point is not to decorate the result with links; it is to keep the factual layer reviewable. Prefer authoritative natural-history, museum, maker, or institutional sources appropriate to the subject. If the retrieved evidence disagrees, stop and describe the disagreement instead of averaging it into an invented feature. A beautiful synthesis of conflicting traits is still an unreliable picture.

After the evidence list, start a visibly separate art-direction block. This is where the frame becomes yours: aspect ratio, subject position, camera height, pose, environment, lighting, depth, palette, and mood. Repeat that all grounded traits must remain visible. The separation prevents diffused mist, dramatic silhouette, vintage print, or another stylistic idea from hiding the very feature the research established.

Do not ask the model to choose a retrieved image and improve it. Ask for an original composition that uses the facts, then forbid reproduction of any source photograph's camera angle or arrangement. Facts can travel; somebody else's composition should not be the invisible template. If one source dominates the retrieved context, make the new spatial brief especially explicit: change viewpoint, crop, subject relation, and background while holding only the factual traits.

Inspect the output in the same order the prompt was written. First, can you see all five grounded traits? Is the complete tail, handle, edge, or other scale-bearing feature inside the frame? Do the returned citations support the visible claims? Only after that pass should you judge light, atmosphere, and emotional pull. Beauty is a poor tie-breaker when one candidate quietly drops a required fact.

For a clean test, run the same prompt once with Web Search and once with Image Search grounding also enabled. Do not change the art direction. Compare only the five factual traits, citation quality, and whether the composition remains original. This is not a contest between pretty frames. It asks whether visual retrieval adds evidence that text retrieval missed, and whether the model can carry that evidence into a new image without tracing one source.

Use this when a real subject has visible facts that cannot be sacrificed to style. Replace the five traits and art direction for your own subject, but keep the two-block structure.

Use Grounding with Google Search with both Web Search and Image Search to ground the appearance of an adult male resplendent quetzal before generating. GROUNDING JOB: identify and cite five observable traits shared by authoritative natural-history references—crown and throat color, breast color, bill color, wing pattern, and tail-feather proportions. ART DIRECTION: create an original 3:2 editorial wildlife image of one adult male resplendent quetzal on a mossy cloud-forest branch after rain. Use a side-on three-quarter pose, diffused canopy light, and shallow mist behind the bird. Make all five grounded traits clearly visible. Keep the entire tail inside the frame. Do not reproduce any source photograph's camera angle or composition. Return a short evidence note with inline source links in accompanying text, and preserve the returned search-suggestion attribution.

The load-bearing line: “Make all five grounded traits clearly visible.” It prevents the evidence block from disappearing behind atmosphere. A trait that was correctly retrieved but hidden by crop, pose, or shadow has not survived into the deliverable.

Choose one real subject you have previously rendered from memory. Write five observable traits and run the prompt with Web Search only. Then enable image search grounding as the only variable and run it unchanged. Score only trait visibility, citation support, complete framing, and whether the composition remains independent of any one source image.

CHANGE ONLY: IMAGE SEARCH GROUNDING

I am building a grounded-image worksheet that separates the evidence list, source check, art direction, visibility rubric, and originality pass.

Want it when it is ready? Reply with send me the grounded-image worksheet and I will share the finished version.

Which subject have you been styling from memory when you should have grounded its visible traits first?

Reply with the subject and five facts you would want a source to settle before the image model begins composing the frame.

If this was useful, forward it to someone making convincing images of subjects they have not checked closely enough.

Until next time,

Luxe Prompting