🐠香蕉-banana全能图像@炮老师的小课堂

"香蕉-banana" is the pack's nickname for Google's Nano Banana image model (the Gemini Image family) - closed, no local weights, and one of the most relentlessly discussed generators in the community per closed-source-models.md. DapaoBananaAllroundNode is the dapaoAI-relay wrapper for it: text-to-image and multi-image editing in one node, with the model menu covering four flavors including "官方稳定版" (official-stable) variants.

The inputs that matter

• 🔑 API密钥 + 🤖 模型 - dapaoAI key, and the model choice: bananaPRO (default), bannana-2, plus the two 官方稳定版 (official-stable) entries - 香蕉pro官方稳定版 and 香蕉2官方稳定版. The stable variants are the author's reliability picks; the plain ones are the cheaper/faster relay routes.
• 📝 提示词 - plain-text prompt, same default commercial-photography example as the sibling image nodes.
• 📐 图片尺寸/比例 - 15 presets (1:1 default, 2:3, 3:2, 3:4, 4:3, 9:16, 16:9, 21:9, 1:4… plus 模型默认).
• 🧩 清晰度 - 1K / 2K / 4K.
• 🖼️ 出图数量 - 1 to 10 per prompt.
• ⚡ 异步模式 - tooltip is the important part: it only controls whether multiple images from the same prompt run concurrently. Upstream prompt lists are auto-fanned-out by ComfyUI no matter what this says. And every image is billed as one API call - no bundle discount for cranking 出图数量.
• 🎲 随机种 - cache-only, never sent to the API.
• 🖼️ 图像1–图像12 - reference images, up to twelve. Same mechanic as the GPT Image node: plug in any reference and it auto-switches to image-to-image/editing mode. Twelve inputs is a lot of edit context - enough for multi-angle product edits or character-consistency work.

How it works

Same relay architecture as DapaoGPTImage2AllroundNode: submit to api.dapaoai.com, poll the task, decode the result into an IMAGE tensor. ⌛ 请求超时 here defaults to 900s (max 1800). Outputs are 🖼️ 图像, 🔗 图片链接, and 📋 响应信息.

Where Banana beats GPT Image, and vice versa

Per closed-source-models.md: Nano Banana is the consumer-accessible Gemini image model, and Nano Banana 2 (the "bannana-2" here) is the speed-quality hybrid; the Pro line does 4K native and 14-reference "thinking mode" editing. In practice people reach for Banana when they want strong multi-image editing and decent text rendering, and GPT Image when text-in-image fidelity is the whole job. This node gives you both doors from the same key.

Gotchas

• The relay is the single point of failure. The README is frank that dapaoAI is the maintained endpoint and other channels get dropped over time. If 妙笔AI工坊 is having a bad day, your Banana calls are too - that's the whole reseller-genre fragility (external-api-nodes.md calls this the node-and-vendor ecosystem where community signal is thin).
• Async mode ≠ list concurrency. Don't read "异步" as "parallel everything." Prompt lists fan out automatically; the switch only affects same-prompt multi-image runs.
• Auto-switch to edit mode when any reference is connected - unplug references for pure T2I.
• Per-call billing. 出图数量 10 = ten paid calls, and no retry of paid POSTs.
• Chinese-only labels, standard for this pack.

Install: ComfyUI Manager → "dapaoAPI", or git clone https://github.com/paolaoshi/ComfyUI-dapaoAPI.git into ComfyUI/custom_nodes/, pip install -r requirements.txt, restart. dapaoAI key from api.dapaoai.com (register → wallet → redeem code → apply key in the default group).