🍌 Gemini Banana 镜像图片编辑

The official Gemini Banana Image Edit node requires a Google API key, which means a Google Cloud/Vertex project and all the billing friction that comes with it. This node is the mirror twin: same image-edit job, but it routes through a reseller or mirror site - Comfly, T8, Kuai, API4GPT, OpenRouter, Comet, aabao - so you can pay with one key from a proxy instead of standing up a Google billing account. For a lot of users outside the US/EU, or anyone who wants looser billing and per-call rates, that's the entire reason this node exists (the reseller economics are covered in the external-api-nodes KB doc).

It's the "镜像图片编辑" node, the mirror sibling of the official editor, and mechanically it's the same Gemini generateContent call - just pointed at whichever endpoint you pick.

How it works

The mirror_site dropdown is loaded from Gemini_Banana_config.json (mirror_sites), and the node auto-selects the right endpoint per model. The README's NB2 update notes the important detail: Nano Banana 2 models go through Gemini-native generateContent, but some providers route NB1 and other models through OpenAI-style paths, so the node picks the format based on the model you chose. Pick custom if your provider isn't listed - but then you must fill both api_key and base_url yourself or it errors immediately.

The inputs that matter

• mirror_site - 15 presets: nano-banana官方 (official, needs a real Google key), comfly, Comfly-HK/US, Kuai API, the T8 variants, API4GPT, OpenRouter, Comet, aabao, custom. Default is comfly.
• api_key - the mirror's key. Leave blank and it falls back to the mirror site's config.
• image, prompt, negative_prompt - the edit job itself. Default prompt: "Can you add a llama next to me?"
• model - 12 options with an "Auto (Latest Gemini 3 Pro) 🤖" default that picks the newest model your mirror supports. The dropdown labels which provider each model needs (e.g. nano-banana [Comfly-T8], gemini-2.5-flash-image-hd [API4GPT]).
• aspect_ratio, response_modality, output_resolution (Auto/1K/2K/4K) - the imageConfig controls.
• upscale_factor / gigapixel_model - the Topaz Gigapixel hook (2x–6x).
• quality, style, detail_level, camera_control, lighting_control, template_selection - the photography presets, same as the official node.
• temperature / top_p / top_k / max_output_tokens / seed - sampling.
• timeout - seconds before the call gives up (default 300). Mirror calls can be slow; don't set it too tight.
• The optional block: custom_additions, safety_level, system_instruction_preset, custom_system_instruction, and the iterative-refinement trio (enable_iterative_refinement, keep_last_turns, reset_conversation, lock_seed, auto_use_last_image, conversation summary).

Outputs: edited_image (IMAGE) and response_text (STRING).

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 lives in Gemini_Banana_config.json - edit that file if you want to add your own providers.

Gotchas

Mirror providers come and go, and their model rosters drift from what the dropdown claims. The classic failure is "model not found" - check the model's provider label in the dropdown and confirm your mirror actually serves it. And a word of caution: a reseller still passes your prompt and image to Google's model, so the "looser filtering" some advertise is not a bypass, and your images are going to a third party. Use a throwaway key if you're not sure about a provider.