🍌 Gemini Banana Text to Image

The mirror image-gen node is the pragmatic pick when you can't deal with Google billing. This one is the opposite choice: "🍌 Gemini Banana Text to Image" goes straight to Google's official API with your own key - the cleanest, least-flaky route to Nano Banana / Gemini 3 Pro Image generation that this pack offers. If you have a Google AI Studio key, start here instead of a mirror; fewer variables, no reseller in the middle, and you keep the full model feature set.

It's text-to-image only (there's a separate edit node if you need image input), and it brings the pack's usual control surface: aspect ratio, 1K/2K/4K output resolution, quality/style/detail presets, camera and lighting controls, a negative prompt (a nice extra the raw Gemini API doesn't really have), and Google Search grounding via the optional enable_google_search flag.

How it works

Calls Gemini's generateContent with imageConfig (aspect ratio + image size) and responseModalities. The camera/lighting/template presets are the pack's own prompt-injection layer - they get baked into the request as structured instructions, so "Professional Portrait" + "Golden Hour" actually steer the model toward that look. The negative prompt is likewise folded into the instruction side of the call.

The inputs that matter

• api_key - your Google AI Studio key. This is the official path, so no mirror_site dropdown here.
• prompt / negative_prompt - what to draw, and what to avoid. Default prompt is a landscape; replace it.
• model - gemini-3-pro-image-preview (default) or the 2.5-flash image models. Pro is the quality pick; flash is cheaper and faster.
• aspect_ratio - Auto or one of 10 ratios.
• response_modality - TEXT_AND_IMAGE or IMAGE_ONLY.
• output_resolution - Auto / 1K / 2K / 4K. 4K costs more; Auto is a sensible default.
• upscale_factor / gigapixel_model - Topaz hook (needs the optional face-restore extras to function).
• quality, style, detail_level, camera_control, lighting_control, template_selection - the presets. A short list you'll actually touch: style (natural/vivid/artistic/cinematic…), quality (standard→ai_ultra), and template_selection (portrait, landscape, product, anime…).
• temperature / top_p / top_k / max_output_tokens / seed - sampling.
• custom_instructions (optional) - free-form extra guidance.
• enable_google_search (optional) - grounding; you'll see grounding_info on the output when it's on.
• Optional refinement block: enable_iterative_refinement, keep_last_turns, reset_conversation, lock_seed, conversation summary.

Outputs: generated_image (IMAGE), generation_text (STRING), grounding_info (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

Requires google-genai>=1.51.0. Restart.

Gotchas

Cost: Pro image calls are metered per image and resolution, and this pack's love of defaults (4K, Pro) will quietly inflate your bill if you don't dial it back to flash + 1K for drafts. Also, the same filter reality as every closed model: whatever Google refuses, this node refuses - there's no bypass, and a "negative prompt" won't unblock a content block. And the 4K/upscale options are dead weight unless you've installed requirements_face_restore.txt for the Gigapixel path.