🍌 Gemini Banana 镜像多图像编辑

Take the multi-image edit node, route it through a mirror instead of the official Google API, and you get this: "🍌 Gemini Banana 镜像多图像编辑," the mirror multi-image editor. Up to four images in, one edited image out, powered by Gemini's image model through Comfly, T8, Kuai, API4GPT, OpenRouter, Comet, or aabao - whichever mirror you set. For anyone who can't or won't set up Google billing, this is the way to get multi-image consistency-style edits.

The one behavioral difference from the official MultiImageEditBananaNode (besides the mirror plumbing and the mirror_site dropdown) is that the model list here is the mirror-flavored set - the Auto entry plus labeled options like nano-banana [Comfly-T8] and gemini-2.5-flash-image-hd [API4GPT] - and there's an extra timeout control, which mirrors genuinely need because they're slower than the official API.

How it works

Same Gemini generateContent with multiple inline images, but the endpoint and auth are chosen by mirror_site. The node loads the site list from Gemini_Banana_config.json, and for the newer Nano Banana models it uses the Gemini-native format, with fallbacks to OpenAI-style paths for NB1 on providers that require them. Choose custom and you're on your own for both api_key and base_url - the node refuses to guess.

The inputs that matter

• mirror_site - 15 presets, default comfly.
• api_key - the mirror's key (falls back to site config).
• image1 … image4 (optional) - the images to edit.
• prompt / negative_prompt - the instructions; the Chinese default wants rewriting to be specific.
• model - Auto by default, or the labeled mirror models.
• aspect_ratio, response_modality, output_resolution - output controls (1K/2K/4K).
• quality, style, detail_level, camera_control, lighting_control, template_selection - the photography presets.
• upscale_factor / gigapixel_model - Topaz hook.
• temperature / top_p / top_k / max_output_tokens / seed - sampling.
• enable_iterative_refinement, keep_last_turns, reset_conversation, lock_seed, conversation summary - session refinement.
• timeout - seconds before giving up (default 300).

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, edit Gemini_Banana_config.json to add mirrors.

Gotchas

The mirror reality applies harder here than single-image: multi-image requests are bigger, so they're slower through a proxy and more likely to hit that timeout. Raise it before you blame the node. "Model not found" remains the standard mirror failure - check the provider label on your chosen model. And on multi-image consistency, treat the output as a strong draft; VLM multi-subject attribution is still shaky, so eyeball the result before you build a batch pipeline on it.