Gemini Enhance Prompt

You type "cyberpunk alley, rain, neon" and GeminiEnhancePrompt hands your sampler a paragraph about wet asphalt reflections, a glowing ramen sign, and a lone figure in a long coat. That's the whole job. It's a prompt enhancer that runs on Google's servers instead of your GPU - one of two nodes in the small comfy-gemini pack (the other, GeminiImageToPrompt, does the reverse and reads an image).

This fits a very real slot in the graph. The 2026 ComfyUI stack reads prompts like instructions, not tag bags (prompt-engineering.md), and prompt enhancement is now a routine upstream step - literally prompt enhancer shows up 250+ times a year in the community (llm-in-comfyui.md). The interesting nodes run an uncensored local 8B model for free. This one is the API alternative: frontier-chat quality, per-call cost, and Google's content filter attached. For SFW stuff where you want Gemini-grade wording, it does the job with zero VRAM.

How it works

The node is a thin HTTP client, not a model. It takes your prompt, tacks on a system_prompt (the instruction telling Gemini to expand on scene, subject, style, lighting, and "output only the enhanced prompt, no longer than 700 characters"), and calls generate_content on whatever model you pick. The text comes back and flows out the enhanced_prompt (STRING) output - straight into a CLIP Text Encode, or a PreviewText/ShowText node if you want to read it before committing.

The inputs that matter

Four inputs, and you really only touch two:

• prompt - the rough text you want expanded. Multiline, no default required.
• model - a dropdown, and it's hardcoded. The source only offers gemini-1.5-pro-latest and gemini-2.0-flash-exp. Note the README name-drops gemini-2.5-pro-latest - that's not in the actual list. The -exp model is the sensible pick (2.0 Flash has a history in this community as the uncensored-and-cheap captioner), but "exp" means Google can retire it whenever.
• api_key - masked field. Leave it blank and the node reads the GEMINI_API_KEY environment variable, which is the better habit.
• system_prompt - the enhancer's instructions, editable if you want tag-style output or a different length cap.

Installing it

ComfyUI Manager (search comfy-gemini), or the manual route:

cd ComfyUI/custom_nodes
git clone https://github.com/abdelrahman-habib/comfy-gemini
cd comfy-gemini
pip install -r requirements.txt

Then restart ComfyUI. No model downloads - everything happens on Google's servers - but requirements.txt pulls google-genai, which drags in google-auth and httpx. Grab a free API key from aistudio.google.com and you're live.

Where people get burned

The silent fallback is the trap. The enhance method wraps everything in a try/except that returns your original prompt unchanged on any failure - missing key, wrong model, dead network, quota. So the workflow "runs," produces an image, and you have no idea the enhancement never happened. If output looks un-enhanced, open the ComfyUI console and look for Error calling Gemini API in Enhance Prompt. It's the #1 confusion and it's by design, not a bug.

The dropdown rots. Both models are API-side names, not files, and Google has deprecated 1.5 Pro. If you start hitting model-not-found errors, there's no fix in the UI - the choices are baked into the code. You may be better off in the other direction: a local GGUF/Qwen LLM node for enhancement is free, offline, and uncensored, and for most workflows that's the better default (llm-in-comfyui.md walks through it).

Treat it like the credential holder it is. This is a small, one-commit pack with a tiny install base, in the exact category that shipped the LLMVISION malware once (external-api-nodes.md). The reassurance here is concrete: the whole thing is about a hundred lines of readable Python that only calls Google's client. Skim it before your first run, and remember your prompts and key are leaving your machine to a service that logs and filters them.

One workflow note: the default template produces flowing natural language, which suits LLM-encoded bases like Flux and Z-Image. If your model is Illustrious or Pony - a tag-tuned anime lineage - you'll want to rewrite the system prompt to emit comma-separated tags, or the prose will fight the encoder.

It's not the most powerful prompt tool on the registry, but for a one-node job - "make this rough idea sound like a real prompt" - it's honest, small, and does exactly what it says. Just remember it's silent when it fails.