Smart Image Banana

Problem

• Generic text-to-image APIs can be unstable when the image needs readable in-image text, consistent people or objects across references, or a high-resolution layout.
• This skill targets agent workflows: generate or edit an image first, then retrieve the returned url for downstream processing.
• It wraps Nano Banana / Gemini image generation with a fixed model and image_generation capability, reducing trial and error around model names and endpoint differences.

How it works

• Async job: submit POST /api/v1/llm/invoke, take the task_id, poll until status = succeeded, then read data.result.data[].url.
• Reference editing: pass reference images through images as url or b64; state what to preserve and what to change in prompt, such as preserving facial features and outfit while changing the background.
• Paid call: if the first request returns 402, complete WeChat Pay; retry with the same JSON body, carrying WeixinPay-Required and X-Payment-Id.
• Output control: use size for ratios like 1:1, 16:9, or 9:16, or pixel values; specify exact in-image text and position in prompt.

Boundaries and notes

• Not a general illustration API: for ordinary images without special requirements, a more general image-gen capability may be better.
• Payment capability required: if the agent does not have weixinpay, say paid creation is unavailable and stop.
• Async timeout: poll every 2–5 seconds and time out around 3–5 minutes; inspect error_message on failure and do not invent upstream errors.
• Quality checks: verify in-image text, multi-reference consistency, preserved regions in edits, and aspect-ratio fit.