Google Nano Banana 2: request settings and preview delivery

Prepare an image request in Google Nano Banana 2 version 1.0.0, starting with the subject and composition you want. The resource description says the recipe generates an image using Google's Nano Banana 2 Comfy Partner Node and returns it through PreviewImage without a duplicate ComfyUI save. This establishes the intended task and delivery, while the retained interface supplies the request choices.

The example below is a suggested prompt, not an executed example. Before relying on a returned image, establish execution access as described in the next section.

Preparation classifies this resource as API execution. Its component inventory is empty and explicitly incomplete; it does not establish that no dependencies are needed. No exact package, model-file, credential or account requirements are retained, and no backend or platform is named. Use the running environment's setup information to establish access before depending on this recipe. There is no supported prerequisite resource to link here.

The retained workload is labeled Standard configuration at 1024 × 1024. It is a configuration record, not a measured run or proof that the 1K option produces those exact dimensions. GPU capacity, system RAM and duration have no scoped estimates, so this record cannot determine hardware sizing or turnaround. API classification alone does not resolve those requirements.

All five retained fields are optional and marked advanced. Their interface defaults and accepted values are known, but their graph bindings are unresolved and their current native values are unknown. An optional form field is not evidence that the provider accepts every omitted or empty value.

Field	Interface default	Accepted values	Useful choice and evidence limit
Prompt	Empty string	Text	Describe the requested content. No prompt-length constraint is retained.
Aspect ratio	auto	auto, 1:1, 2:3, 3:2, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9, 21:9	Select a ratio to make the requested shape explicit. How auto resolves and which dimensions are delivered are unspecified.
Resolution	1K	1K, 2K, 4K	Select among the available resolution labels. Exact pixel dimensions, cost and duration consequences are not retained.
Thinking level	MINIMAL	MINIMAL, HIGH	Select the named thinking setting. Its implementation and effects on quality or timing are not characterized.
Seed	-1	Integers from -1 through 4294967295, inclusive	Enter an accepted integer. The meaning of -1 and conditions for repeatability are unspecified.

For a square request, select 1:1. An illustrative configuration is Aspect ratio 1:1, Resolution 1K, Thinking level MINIMAL, and Seed 42, paired with your prompt. These values fit the retained interface; they are not a tested configuration or a guarantee of a particular image. 2K and HIGH are available alternatives, but the pack does not establish their benefits or tradeoffs.

When exploring settings, keeping the prompt and other choices recorded makes the requested configurations easier to compare. Even with the same seed, the retained evidence does not establish that a difference in images is caused solely by the setting you changed. No executable graph adaptation can be supplied from these unresolved bindings.

The supported account is a request interface for image generation, with documented delivery through PreviewImage. The native graph is absent, node definitions are empty, and execution-effect coverage is incomplete. Consequently, the pack cannot trace how the prompt and settings reach a provider, how generation is performed, or how the resulting image reaches the preview.

Interface metadata names GeminiNanoBanana2, node 1, and targets such as /1/inputs/prompt. These are declared interface targets, not confirmed native nodes or executable inputs. They cannot justify graph edits or establish connections. The defaults in the settings table are form defaults rather than saved graph values.

Image generation is the useful task classification because the resource description explicitly names it. The operation record's alias:txt2img signatures do not supply additional mechanism evidence; its capability, rule-binding and adaptation lists are empty. Nothing here establishes reference conditioning, masked editing, compositing, a sampler or a VAE decoder. The generic compositing and tiled-decoding sources elsewhere in the pack therefore do not establish controls or remedies for this recipe. A prompt asking to keep a subject unchanged also does not establish original-pixel preservation.

PreviewImage is the only named delivery node in the resource description. The description says the generated image is returned without a duplicate ComfyUI save. This is documented delivery intent; no PreviewImage node ID, output port identifier or captured output confirms the wiring.

No encoding, filename, download contract or persistence behavior is retained. A preview can meet an immediate viewing need, but a downstream integration or saved-file requirement needs an output contract from the running environment. The absence of a duplicate ComfyUI save does not establish that the image is never stored elsewhere.

Only text and configuration inputs are exposed in the retained interface. There is no supported image-upload or mask path to use for editing an existing image. If the task requires those inputs, or unchanged original pixels, establish a suitable recipe and its exact output guarantees separately; this pack supplies no executable fallback recipe.