Gemini Prompt Generator-JT

Stuck staring at an empty prompt box, again? This is the fix that costs you zero VRAM: Gemini Prompt Generator-JT is a tiny ComfyUI node that calls the Google Gemini API, hands it a theme, and returns a ready-to-use image prompt as a string. No local model, no download - just an API key and an internet connection. It's a fork of the older "Magifactory" prompt generator with enough extras bolted on that it's worth a look on its own, especially if you're already paying Google nothing for the Gemini free tier.

Honest framing up front: this is a thin wrapper around a cloud API, not a clever pipeline. But thin is fine when the alternative is running a 7B+ LLM locally just to write a sentence. The one genuinely useful thing it adds over a bare API call is multimodal input - wire in up to three reference images and Gemini captions or expands on them. That makes it a passable poor-man's image-to-prompt tool too.

How it works

The node reads your key from config.json (it must exist - more below), builds a canned English instruction - "Generate me a prompt for image generator. The theme of the prompt is {theme}." plus a length cap and a blunt "do not explain, only output the prompt" suffix - and sends it to Gemini at temperature 0.9. The request runs on a background thread with a timeout you control, so a hung call errors out instead of freezing your queue.

Two settings change what gets sent. enable_memory keeps the last 15 generated prompts in memory and feeds them back so the model doesn't repeat itself across a batch - genuinely nice for keeping variety when you're generating a whole set. override_system_prompt is the secret weapon: leave it empty and this behaves like a normal prompt generator, but fill it in and the node turns into a general-purpose LLM call with template variables {theme}, {prompt_history} (or {memory}), {prompt_length}, and {seed}. The README's own example is "generate an image prompt for {theme}, history: {prompt_history}" - you can also just ask it to do something unrelated, since it's now a full chat-model call.

The inputs that matter

• theme - the only thing you must fill in. A phrase, a sentence, a vibe. This is what gets expanded.
• model - dropdown ships with gemini-2.5-flash-lite, gemini-2.5-flash, gemini-2.5-pro, and gemini-3-flash-preview. For day-to-day use start with 2.5-flash; the pro model is slower and throttled harder.
• seed - the gotcha the README calls out: it is not sent to Gemini. It just controls when ComfyUI re-runs the node, so keeping it random is your "give me a new prompt" button.
• image_1 / image_2 / image_3 (optional) - IMAGE inputs; wire a Load Image (or a VAE decode) in and the model sees it.
• prompt_length (default 120, 0 = unlimited) and timeout (default 15s) are the only other ones worth touching early.

The single STRING output is your generated prompt. Wire it into a CLIP Text Encode node's positive input and you're done.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/LiJT/ComfyUI-Gemini-Prompt-Generator-JT

Then restart ComfyUI. Its only Python dependency is google-generativeai, which the install handles. You can also try ComfyUI Manager by searching "Gemini prompt generator JT". After restart, the step everyone forgets:

cd ComfyUI-Gemini-Prompt-Generator-JT
cp config.json.default config.json
# then edit config.json and put your key in "GEMINI_API_KEY"

Grab a free key from Google AI Studio - no credit card needed for the free tier.

Common issues

• "API key is required" - you skipped the config step. The node reads config.json from its own folder; without it, it refuses to run.
• Stale model errors - Gemini 1.5 models were retired in September 2025; if you're still pointed at one, everything 400s. The dropdown's 2.5 line is the fix.
• Rate limits. The free tier is genuinely small - on the order of tens to low thousands of requests per day depending on the model - and the README's numbers already disagree with the node's own tooltip, because Google keeps moving them. A batch run can blow a day's budget fast. When you hit 429, the fix is waiting, not retrying harder.
• Timeouts - the default 15 seconds is fine for flash models, tight for 2.5-pro. Bump timeout if you see ComfyUITimeoutError.
• Memory isn't persistent - history lives in RAM, so it resets every restart. That's by design, not a bug.

One last security note, since a prior node in this exact niche (a "LLM vision" helper) turned out to be malware: your Gemini key sits in plaintext inside custom_nodes, so read what you install and keep that folder trusted. This one is a plain API wrapper, but the habit matters.

If you'd rather keep prompt-writing local and private, the community's usual answer is a Qwen-VL node or Ollama - but that's VRAM and setup. This node is the zero-setup-on-the-GPU option: grab a free key, type a theme, get a prompt.