Choose image specifications and references

Start from the deliverable

For a product card, social post or campaign variant, write down the intended crop, output format and acceptance criteria before choosing a model. Resolution and aspect ratio solve different problems: one selects a size tier while the other defines the frame shape. A label such as 1k is not a guarantee of a particular width and height for every aspect ratio.

The Nano Banana 2 page already contains the calculator, examples, API and specifications. Use those controls rather than a separate or stale price list. This guide explains the workflow; it does not rank image quality.

Validate the whole input

Read GET /v1/media-models, find the exact model and inspect request_schema. Validate the entire input object, including required fields, enum spelling, maximum lengths, reference limits and combination constraints. A field appearing in properties does not make every combination valid. Do not infer support from a neighboring model or provider name.

For Nano Banana 2, the existing guide uses the following task envelope. This is illustrative request JSON, not an automatic call; confirm these selections against the current schema before submitting a billable request.

{
  "model": "nano-banana2",
  "input": {
    "prompt": "A simple studio product photograph of a ceramic cup on a neutral background",
    "resolution": "1k",
    "aspect_ratio": "1:1",
    "output_format": "jpeg",
    "images": []
  }
}

Choose format and references deliberately

JPEG is useful when an opaque photographic result is acceptable; choose PNG or WebP only where the model accepts that enum and the delivery workflow needs it. Selecting PNG does not by itself request a transparent background. Check the actual returned file rather than assuming the extension or resolution label proves its properties.

For an edit or reference-based request, supply only authorized input assets in the shape accepted by the model. Do not embed credentials in URLs. Keep original assets, prompt and request parameters together so reviewers can assess whether the intended subject, text, layout and proportions were preserved. Reference input guides the task; it is not a guarantee of exact reproduction.

Approve a sample before creating variants

Review one approved sample for subject details, legible text, crop and brand requirements before expanding the batch. Change one input at a time when diagnosing a problem. If the output is visually wrong, a transport retry does not solve the creative issue; revise the prompt or specification and treat the next generation as a new budgeted attempt.

Image generation uses POST /v1/tasks, followed by GET /v1/tasks/{id}; do not treat the accepted task as the final image. Read the model API guide, task learning path and batch budget guide. These pages complement the existing model calculator rather than replacing current account pricing.