Banana2-提示词模板

Gemini's image models (the "Nano Banana" line, covered in the closed-source-models KB doc) respond dramatically better to a specific style of structured prompt: composition, style, lighting, camera language, laid out as labeled sections. Writing that by hand every time is a chore, and generic English prompts underperform. This node - Banana2-提示词模板 - is a template engine for exactly that. Pick a template like 生成-逼真场景 (photoreal scene) or 编辑-风格迁移 (style transfer), type your idea, and it calls Gemini to produce a Banana-ready prompt plus the raw response.

It's the structured-prompt sibling inside ComfyUI_LLM_Banana, built on the same endpoint logic as the Gemini-Multimodal node and the NanoBanana-GeneralAPI node. Because it's Gemini-native (generateContent), it works through Google's official API or through the mirror/proxy providers the pack is known for.

How it works

The template_type dropdown carries 30+ templates covering the Google-style image categories: generation (photoreal scenes, stylized illustration/stickers, text-in-image, product photography, minimal/negative-space, comic panels), editing (add/remove elements, local repaint, style transfer, multi-image composition, detail preservation, character consistency, TVC storyboard grids), Sora-style video storyboard templates, Webtoon 9-panel grids, and a custom option. It reads your user_prompt (and up to four optional image wires for edit templates), calls the model, and returns the assembled prompt as banana_prompt plus the untouched raw_response if you want to see what the model actually said.

The inputs that matter

• template_type - the core control. 32 options, roughly half generation half editing. Start with 生成-逼真场景 if you're unsure.
• model - defaults to gemini-3-pro-preview [all]; options include the -thinking variant (tagged for which providers carry it) and the older 2.x models.
• api_provider - 11 choices: google, comet, the T8/Comfly mirrors, aabao, custom. The mirror options exist so you can run the same endpoint logic through a proxy/reseller.
• api_key, base_url, version, auth_mode - the unified endpoint/auth controls. version=Auto picks v1beta on Google domains and v1 on mirrors; auth_mode=auto uses x-goog-api-key for Google and Bearer otherwise. Custom providers must supply both key and base_url or the node refuses to run.
• thinking_level (high/low) and media_resolution (Auto/low/med/high) - quality knobs; media_resolution=Auto is what bumps the version to v1beta on Google.
• temperature, max_output_tokens - sampling.

Optional inputs: image through image_4 (for edit templates), custom_template to write your own, system_instruction for a system prompt override.

Install

ComfyUI Manager, search 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

The pack needs google-genai>=1.51.0 (from requirements.txt) for the newer image endpoints. Restart, and if you've updated and don't see the new base_url/version/auth_mode fields, hit "Reload Custom Nodes" or restart ComfyUI - the README says that explicitly.

Gotchas

If you pick a mirror provider, make sure the provider actually carries the model you selected - the dropdown labels which providers each thinking model is available on, and a model the provider doesn't have will 404 or return "model not found." And the output is a prompt, not an image: wire banana_prompt into a Gemini Banana image-generation or edit node next. Missing keys on a custom provider error out immediately by design, so fill both api_key and base_url together.