NanoBanana-GeneralAPI

If you only install one node from ComfyUI_LLM_Banana, this is the one to consider. NanoBanana-GeneralAPI is the pack's universal image-generation-and-editing node: it can generate an image from text or edit up to four input images, through Google's official API or any of ~14 mirror sites, with a unified endpoint/auth system that the rest of the pack has now been aligned to. The README is explicit that Banana2-提示词模板 and Gemini-Multimodal share its design. Learn this node and the others start to look like variants.

It's a closed-model API node, so the usual rules apply (see the external-api-nodes KB doc): per-call cost, your data leaves the machine, Google's filter still applies no matter which mirror you use.

How it works

The config file Gemini_Banana_config.json defines the mirror_sites (nano-banana官方, comfly, Comfly-HK/US, Kuai, T8 variants, API4GPT, OpenRouter, Comet, aabao, custom), each with a URL and optional key. The node resolves settings by priority: your input → mirror-site config → global config → provider default. So you can leave api_key and base_url blank and let the selected mirror fill them in, or override everything yourself.

The endpoint builder and auth are the pack's core logic: if your base_url is already a full endpoint (.../models/<model>:generateContent) it's used as-is; otherwise it's assembled as base_url/<version>/models/<model>:generateContent (or appended to a /v1-style base). auth_mode=auto uses x-goog-api-key on Google domains and Authorization: Bearer elsewhere; google_xgoog and bearer force one or the other. version=Auto picks v1beta/v1alpha on Google and v1 on mirrors.

The inputs that matter

• prompt - the generation/edit request (default: "生成一张清晰的香水产品图" - generate a clear perfume product image).
• mirror_site - the provider. Default nano-banana官方 (official Google). Pick custom and you must fill api_key + base_url or it errors.
• api_key / base_url - override or let the mirror config supply them.
• model - gemini-3-pro-image-preview or custom (then fill custom_model).
• version, auth_mode - endpoint/header control, described above.
• response_mode - TEXT_AND_IMAGE / IMAGE_ONLY / TEXT_ONLY.
• aspect_ratio + image_size - geometry and 1K/2K/4K output.
• upscale_factor / gigapixel_model - Topaz Gigapixel hook (2x–6x).
• temperature / top_p / top_k / max_output_tokens / seed - sampling.
• strict_native - force the Gemini-native format (helps with mirrors that drift).
• system_instruction, image_mime (png/jpeg/webp), timeout, use_system_proxy - the rest.
• Optional: image … image4 (for editing), extra_payload_json (raw JSON to merge into the request).

Outputs: text (STRING) and image (IMAGE).

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

Needs google-genai>=1.51.0. Restart.

Gotchas

The priority fallback can be confusing when things silently "work" with a config key you didn't intend - check the logs (the README tells you to watch Request URL and Request Headers) to confirm which endpoint and auth actually fired. strict_native is the escape hatch if a mirror returns format errors on the OpenAI path. And the image_size 4K option is the bill-inflator: drop to 1K for drafts. Mirrors churn, so if a site 404s, the fix is usually "pick a different mirror," not a config deep-dive.