心宝❤Banana

This node is Google's Nano Banana - the Gemini image models - sitting inside ComfyUI as if it were a local sampler. You type a prompt, it phones a paid API, and the returned image drops back into your graph like any other IMAGE tensor. No GPU load, no checkpoint download, and also no free lunch: every run is metered.

It's the flagship of the comfyui-banana-li-linux pack from 李心宝 (Xinbao), a Chinese e-commerce designer who built a whole kit around Gemini generation for product work. The "Banana" in the name is the community's nickname for Gemini-native image generation, which Google itself later adopted as a brand. What makes this pack different from Google's own flow is that it routes through Xinbao's reseller channel, so there's no Google Cloud project to set up - you buy a key, paste it in, and go. The KB's external-api-nodes doc covers exactly why this reseller ecosystem exists: cost, region access, and billing friction that the official Gemini API route is bad at.

What it actually does

Text-to-image is the base case. You get prompt, a multiline box that takes a real description of content, style, and composition. Then the things that matter:

• banana_api_key - leave it blank and the node falls back to the [gemini] section of the pack's config.ini. That fallback is the intended workflow: set the key once, never bake it into a shared workflow.
• model_type - five choices, defaulting to gemini-3-pro-image-preview (Nano Banana Pro). gemini-2.5-flash-image is the cheaper original Nano Banana; gemini-3.1-flash-image-preview is the speed tier.
• aspect_ratio - Auto plus a dozen presets. Note the tooltip: the extreme 1:4 / 4:1 / 8:1 ratios only work on gemini-3.1-flash-image-preview; other models silently fall back to Auto.
• batch_size - 1 to 8 images per request. Keep it small. The author is blunt about why: at high batch the upstream API sometimes reports success, bills you, and returns no image.
• seed - -1 randomizes, a fixed value reproduces the same output. top_p (default 0.95) is the sampling knob: lower is more conservative.

For image-to-image, wire up to five reference images into image_1 through image_5. That's the multi-image fusion Gemini is good at - feed it a product shot plus a style reference and it'll merge them. image_size (1K/2K/4K, default 2K) only takes effect on the gemini-3-pro-image* models.

The outputs are images (an IMAGE tensor, batch or single) and text. The text output is worth wiring to a preview node: when the request fails, the node hands back a visual error image and puts the reason in the text port instead of just dying.

The fiddly bits

There are three toggles a beginner will be tempted to flip and mostly shouldn't. 线路 (route) picks the channel - Hong Kong dedicated, direct US, CF dedicated, or the default 心宝新渠道 - and they share the same balance, so you can switch if one is slow. 绕过代理 (bypass proxy) forces a direct connection when your proxy is flaky. 禁用SSL验证 disables certificate checks to improve success rate, and the tooltip says it plainly: only on a trusted network, or your key can leak to whoever's intercepting your traffic.

There's also an optional binding_context input - skip it for plain generation; it only matters for the pack's local-crop chain (see the binding and crop nodes).

Installing

This pack ships per-platform repos - Linux (this one), Windows, and macOS. Manager search "comfyui-banana-li", or:

cd ComfyUI/custom_nodes
git clone https://github.com/98624017/comfyui-banana-li-linux

Then restart ComfyUI. The pack ships precompiled binaries, but you still need its Python deps: requests, opencv-contrib-python, scikit-image, segment-anything, blend_modes, timm, Pillow, psd-tools. No model files to download - the model lives on Google's servers.

The traps

• It costs money every single run. The KB's honest framing: a local generation is free after electricity, a cloud call is metered forever. Validate with batch 1 before committing to 8.
• Your images and prompts leave the machine. That's the mechanism, not a bug - and it's why the key security and SSL warnings are worth respecting.
• Batch 1–4 for stability. The "billed but no image returned" failure is a real upstream behavior, and this node retries with exponential backoff but can't always recover.