Nano Banana 2 is Google's Gemini 3.1 Flash Image model. Google's current model page describes generation and conversational editing, 0.5K through 4K output options, image-search grounding, improved international text rendering, and additional aspect ratios. The broader image-generation documentation covers reference mixing, text rendering, resolution, and current limitations.
Those are provider capabilities—not a promise that every host exposes the same controls or that every output is correct. This guide documents Masonry's live route, adds one first-hand wide-format artifact, and turns the model's launch claims into checks you can repeat.
Evidence boundary: the banner below is one first-hand text-to-image output from one prompt, one seed chosen by the service, and the Masonry route described here. It is not a multi-seed benchmark, a reference-fidelity test, or a head-to-head comparison. Google capability statements are linked to Google; Masonry controls come from the live CLI contract checked August 4, 2026.
The model name you actually use
There are three identifiers worth separating:
| Context | Current identifier | Use |
|---|---|---|
| Google API model | gemini-3.1-flash-image | Google's direct API and documentation |
| Masonry browser route | /models/nanobanana-2 | Interactive model page |
| Masonry CLI model ID | gemini-3.1-flash-image-preview | masonry image and masonry models params |
This is not cosmetic. A first attempt with --model nanobanana-2 returned model not found, even though an alias schema was discoverable. masonry models list --type image returned the executable preview ID, and the same prompt succeeded with that ID. Check the live list before putting a model name into a script.
Current Nano Banana 2 controls on Masonry
Run masonry models params gemini-3.1-flash-image-preview to inspect the route. The checked CLI contract exposes:
| Input | Current contract | What it means in production |
|---|---|---|
| Prompt | Text generation or mixing instruction | Write observable requirements; capacity is not an acceptance test. |
| Output size | Required fixed dimension or alias | Choose placement and resolution before generating. |
| Seed | Optional integer from 0 to 2,147,483,647 | Record it for reruns, but still inspect the output. |
| References | Zero to 14 images | State the role and invariants for every source; references can still drift. |
| Search grounding | Not exposed by the checked CLI flags | Google's direct-API feature is not automatically a Masonry input. |
| Thinking level | Not exposed by the checked CLI flags | Do not paste provider-specific controls into this route. |
Masonry offers 1K, 2K, and 4K aliases for 1:1, 2:3, 3:2, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9, and 21:9. For example, 21:9:2k maps to 3168x1344, and 4:5:4k maps to 3712x4608.
Google's direct API documents a 0.5K output and extreme ratios including 1:4, 4:1, 1:8, and 8:1. Those are not present in the Masonry CLI contract checked for this guide; the provider's maximum feature set and a hosted route are different things.
One first-hand 21:9 text-layout test
The brief asked for a fictional cobalt-blue travel mug, a 21:9 ecommerce banner, two exact strings, left-side negative space, a right-side product, and no other copy.
Observed outcome: the candidate rendered “ONE BRIEF. THREE CROPS.” and “Square · Portrait · Wide” once each with no additional copy. It placed the product on the right, kept usable text space on the left, used a navy mug with restrained coral details, and produced a coherent camera-left light direction. The stone surface and warm neutral setting followed the brief broadly rather than literally. This run did not test reference preservation, localization, small type, or repeated identity.
| Check | Requested | Observed |
|---|---|---|
| Headline | Exact string once | Pass in this output |
| Subheading | Exact string once | Pass in this output |
| Extra copy | None | Pass in this output |
| Composition | Text left, mug in right 38%, generous negative space | Usable split; exact percentage not measured |
| Palette | Cobalt/navy with restrained coral | Pass in this output |
| Product fidelity | Fictional mug, no reference supplied | Not tested |
One clean result is useful evidence that the route can execute the brief. It is not evidence that exact text will pass across prompts, seeds, languages, or font sizes.
Reproduce the wide test from the CLI
masonry image "Create a premium 21:9 editorial ecommerce banner for a fictional cobalt-blue travel mug on pale limestone, warm-gray studio background, soft window light from camera-left, restrained navy and coral accents. Exact headline on the left: 'ONE BRIEF. THREE CROPS.' Exact small subheading below: 'Square · Portrait · Wide'. Keep the mug on the right 38 percent and leave generous negative space around the text. Render those two text strings exactly once each. No other words, logos, badges, prices, or claims." \ --model gemini-3.1-flash-image-preview \ --aspect 21:9:2k \ --output nano-banana-wide.png
For a repeatable comparison, add a recorded seed:
masonry image "Your fixed test brief" \ --model gemini-3.1-flash-image-preview \ --aspect 4:5:2k \ --seed 41217 \ --output candidate-41217.png
A seed helps rerun the route; it does not make a weak brief or unreviewed output production-ready.
Prompt patterns that are easy to verify
Exact displayed copy
Quote every required string, specify its hierarchy, and prohibit competing text:
Headline: “SUMMER SET.” Subheading: “Three daily essentials.” Render each string exactly once. No other words, numbers, logos, prices, badges, or legal copy.
Reject the output for one wrong character. Add price, ingredient, dosage, legal, or brand-critical copy with deterministic typesetting after generation.
Reference-based product scene
Attach the approved product image and separate changes from invariants:
masonry image "Change only the environment to pale limestone with cool morning light. Keep the supplied bottle's silhouette, closure, proportions, color, label placement, and every printed character unchanged. Add no props or copy." \ --model gemini-3.1-flash-image-preview \ --ref ./approved-bottle.png \ --aspect 4:5:2k \ --output bottle-scene-candidate.png
“Keep unchanged” is an instruction, not a lock. Compare the candidate with the source; composite the approved product back into the generated scene when exact packaging matters.
Multiple references
Google documents up to 10 object images plus four character images for this model, while the Masonry route accepts up to 14 references. Give each source one role in the prompt: product geometry, approved label, person identity, wardrobe, or environment. Avoid fourteen unexplained images and hoping the model infers the hierarchy.
An acceptance sheet for commercial work
| Area | Pass condition |
|---|---|
| Required text | Every character, punctuation mark, line break, and occurrence count matches the brief. |
| Product or identity | Silhouette, color, geometry, face, wardrobe, and approved artwork match the source. |
| Composition | Subject occupancy, negative space, crop, horizon, and camera angle suit the final placement. |
| Scene truthfulness | No invented ingredient, claim, price, certification, accessory, or functional detail appears. |
| Technical output | Correct dimensions, legible delivery crop, clean edges, and no unwanted text or marks. |
| Repeatability | Prompt, model ID, size, seed, references, and reviewer decision are recorded. |
Nano Banana 2, Pro, or Lite?
Google's current model-selection guidance positions:
- Nano Banana 2 as the general-purpose balance of capability, cost, and latency.
- Nano Banana Pro for complex professional asset production, stronger brand consistency, localization, and precision control.
- Nano Banana 2 Lite for lower-latency, lower-cost generation where scale is the main constraint and multiple-reference work is not the priority.
Those are portfolio positions, not proof that one route wins your task. Run the same source, prompt, output size, and acceptance sheet when the decision matters. The Nano Banana 2 vs GPT Image 2 comparison and product-photography model guide cover narrower comparisons without treating one model as a universal default.
The bottom line
Nano Banana 2 is a current generation-and-editing option with official support for high-resolution output, reference mixing, text rendering, and broad aspect-ratio control. On Masonry, the actionable details are more specific: use gemini-3.1-flash-image-preview in the CLI, choose a required 1K/2K/4K size, optionally record a seed, and attach up to 14 references.
The first-hand banner passed its exact-copy and broad layout checks, but it was one output without references. Treat that as a reproducible example, not a reliability score. Define the acceptance sheet before generating, inspect every required detail, and select the model by accepted deliverables rather than launch language.
For the lower-cost 1K route, use the Nano Banana 2 Lite guide, which separates current price and latency claims from one first-hand output and its visible limits.
