Missing generated image data in output after update to Nano Banana 2

[zgojin]

Description / 描述:

[English]
Following the Gemini Web update to Nano Banana 2, there is a discrepancy in the response behavior for image-to-image tasks. When an input_image is provided with a transformation prompt (e.g., "Style: figurine"), the model returns a reasoning block but no generated image data in the final output field.

The API returns a status: "completed", but the output array contains only output_text. There is no image_url object or Markdown image link present in the assistant's response. The input field correctly reflects the source image Base64, but the output is limited to the reasoning text.
[Chinese]
在 Gemini 网页更新为 Nano Banana 2后，图生图任务的返回行为发生了变化。当提供 input_image 和转换指令（如“Style: figurine”）时，模型返回了 推理块，但最终 output 字段中未包含生成的图片数据。

API 返回状态为 status: "completed"，但 output 数组仅包含 output_text（推理文本）。助手的回复中缺失了 image_url 对象或图片链接。虽然 input 字段正确回显了源图片的 Base64，但 output 字段仅包含纯文本内容。

---

To Reproduce / 重现步骤:

[English]

1. Provide a payload with a Base64 encoded input_image and a prompt for style transformation (e.g., "Style: figurine").
2. Inspect the output array in the JSON response.
3. Observe that only output_text (reasoning) is present, with no newly generated image.

[Chinese]

1. 在请求中提供 Base64 编码的 input_image 以及风格转换提示词（例如 "Style: figurine"）。
2. 检查响应中的 output 数组。
3. 观察到结果仅包含 output_text（推理文本），而没有任何新生成的图片数据。

---

Expected Behavior / 预期行为:
The captured response should include the generated image data (URL or Base64) that Nano Banana 2 produces on the web side.
捕获到的数据应包含 Nano Banana 2 在网页端成功生成的图片数据（URL 或 Base64）。

Actual Behavior / 实际行为:
Although the web UI generates the image successfully, the captured JSON response in the output field only contains the reasoning text and misses the image payload.
虽然网页端已成功生成图片，但项目捕获到的 JSON 响应中 output 字段仅包含 推理文本，未能捕获到图片数据。

---

Environment / 环境信息:

• Model: gemini-3.0-pro
• Platform: Gemini Web Update

Notes on Attachments / 附件说明:

[English]
I have attached the Python test script used for this request and the complete raw JSON response (debug_response.txt). These files provide the exact payload structure and the specific model output containing the thinking process without the generated image.

[Chinese]
我附上了用于发起请求的 Python 测试脚本以及完整的原始 JSON 响应（debug_response.txt）。这些文件提供了准确的 Payload 结构，以及包含推理过程但缺失生成图片的具体模型输出，供进一步分析。

test.py

debug_response.json

[luuquangvu]

@zgojin Without detailed error logs, it's hard to pinpoint the cause, but from your JSON file, I suspect you're using an outdated version. The image generation timeout problem is tied to the gemini-api library. You could try my fork, which includes a PR awaiting merge and comes with numerous new features and bug fixes, to see if it works for you.

[zgojin]

@luuquangvu thanks for getting back to me! You are spot on—I am indeed using a fairly outdated version. I've held off on updating for a while because I made some custom modifications to the code locally.
​I will definitely check out your fork and see if the PR resolves the image generation timeout issue on my end. I'll let you know how it goes. Thanks for the help and for sharing your work!