🍌 Gemini Banana 镜像图像生成

Text-to-image with Google's closed Nano Banana / Gemini image models, but without needing a Google billing account - that's this node in one line. "🍌 Gemini Banana 镜像图像生成" is the mirror-text-to-image twin of the official generation node. Same closed model behind it, same endpoint math, different door: you pick a mirror site (Comfly, T8, Kuai, API4GPT, OpenRouter, Comet, aabao…) and the node does the rest. The model can never be downloaded, so the API is the only path; the mirror exists to make that path cheap, region-accessible, and free of Google Cloud signup friction (the external-api-nodes KB doc walks through why that reseller layer exists).

The default prompt is a sunset mountain landscape, the default model is an "Auto" pick of the latest Gemini 3 Pro, and the output is an image plus the model's text response.

How it works

It's the same generateContent machinery as the mirror edit node, minus the input image. The mirror_site dropdown is loaded from Gemini_Banana_config.json, and the node selects the correct endpoint format per model (Gemini-native generateContent for NB2, OpenAI-style paths on some providers for NB1 - the README's NB2 update spells out the branching). The Auto model entry inspects what your chosen mirror supports and picks the newest usable model.

The inputs that matter

• mirror_site - 15 presets (nano-banana官方, comfly, Comfly-HK/US, Kuai, T8 variants, API4GPT, OpenRouter, Comet, aabao, custom). Default comfly.
• api_key - the mirror's key; falls back to the site config if blank.
• prompt / negative_prompt - the generation request. No image input here; this is pure text-to-image.
• model - Auto by default; otherwise pick from the 12 labeled options (note [Comfly-T8], [API4GPT], [OpenRouter] tags - the model must be served by your mirror).
• aspect_ratio, response_modality, output_resolution (Auto/1K/2K/4K) - imageConfig controls.
• upscale_factor / gigapixel_model - Topaz Gigapixel hook (2x–6x).
• quality, style, detail_level, camera_control, lighting_control, template_selection - the photography presets.
• temperature / top_p / top_k / max_output_tokens / seed - sampling.
• enable_google_search (optional) - grounding; only meaningful on providers that support it.
• timeout - seconds before giving up (default 300).

Outputs: image (IMAGE), response_text (STRING), and grounding_info (STRING) when Google Search grounding is enabled. The image goes to preview/save; the text and grounding info are diagnostics.

Install

ComfyUI Manager → ComfyUI_LLM_Banana, or:

cd ComfyUI/custom_nodes
git clone https://github.com/xuchenxu168/ComfyUI_LLM_Banana
cd ComfyUI_LLM_Banana
pip install -r requirements.txt

Restart. The mirror list is editable in Gemini_Banana_config.json.

Gotchas

Same family rules as the mirror edit node: providers churn, "model not found" is the #1 complaint, and the Auto model pick is only as good as your mirror's roster. Keep timeout above a minute for busy providers - some of these mirrors queue. And remember the data-leaves-the-machine truth of the whole category: your prompts go to a third party whose logging you don't control, and no mirror removes Google's own content filter from the model, whatever they claim.