一次调用返回多张图

candidates = 1,  finishReason = "STOP",  parts 长度 = 6

  parts[0]  text        "Here are three distinct logo design variations…"
  parts[1]  inlineData  image/jpeg  914,915 B  1408×768   sha256 前 12 位 f6ab6076c1e6
  parts[2]  text        "### Variation 2: The Modern Minimalist…"
  parts[3]  inlineData  image/jpeg  632,568 B  1408×768   sha256 前 12 位 e3d5eda56d83
  parts[4]  text        "### Variation 3: The Geometric Abstract…"
  parts[5]  inlineData  image/jpeg  796,069 B  1408×768   sha256 前 12 位 5fdf9d3fe811

"usageMetadata": {
  "promptTokenCount": 23,
  "candidatesTokenCount": 4132,
  "candidatesTokensDetails": [
    { "modality": "IMAGE", "tokenCount": 3360 }   // 3360 = 3 张 × 1120
  ],
  "totalTokenCount": 4155
}

parts = (response.get("candidates") or [{}])[0].get("content", {}).get("parts") or []

images, has_text_between = [], False
prev_was_image = False
for p in parts:
    if "inlineData" in p:
        images.append(p["inlineData"])          # 存 data 和 mimeType，后缀以 mimeType 为准
        prev_was_image = True
    elif p.get("text", "").strip():
        if prev_was_image:
            has_text_between = True             # 图后面又出现文字 = 交替生成
        prev_was_image = False

if not images:
    raise RuntimeError("未返回图片，检查是否被安全策略拦截")

# 交替生成 → 每张都是成品，全部保留；否则是思考中间稿 → 取最后一张
results = images if (len(images) > 1 and has_text_between) else images[-1:]

const parts = resp?.candidates?.[0]?.content?.parts ?? [];
const images = parts.filter((p) => p.inlineData?.data);
const hasTextBetween = parts.some(
  (p, i) => p.text?.trim() && i > 0 && parts[i - 1].inlineData,
);
const results =
  images.length > 1 && hasTextBetween ? images : images.slice(-1);

{
  "data": [
    { "b64_json": "…" },   // 变体 1
    { "b64_json": "…" },   // 变体 2
    { "b64_json": "…" }    // 变体 3
  ],
  "usage": { "completion_tokens_details": { "image_tokens": 3360 } }
}