Gemini prompt enhance

Gemini Prompt Enhance is what you reach for when your prompt is doing 20% of the work it could be. You wire in a rough description - maybe the output of a captioner node, maybe just a few words you typed - and it comes back from Gemini with the kind of texture a good prompt engineer would have typed by hand: quality descriptors, a lighting pass, a mood pass, skin detail if there's a person in frame. It's an LLM call embedded directly in your graph, not a static text template.

The mechanism is straightforward once you look at the default prompt field, which is really a system instruction disguised as a text input. It tells Gemini to add quality descriptors ("a high-quality photo, an 8K photo"), lighting description, scene mood, and skin realism if a person is present, then output only the rewritten text - no preamble, no explanation. You can leave that instruction as-is or rewrite it entirely; it's just a multiline string. The actual thing you want enhanced goes in text_input. The author credits this as a fork of ShmuelRonen's ComfyUI-Gemini_Flash_2.0_Exp, with gender-aware rewriting and better failure handling layered on top.

The inputs worth knowing: model defaults to gemini-2.5-flash-lite (fast, cheap) with gemini-2.0-flash and gemini-2.5-flash also selectable. temperature (0–1, default 0.4) and max_output_tokens (up to 8192) are the usual LLM knobs - keep temperature low if you want consistent, predictable rewrites across a batch. gender_prior (blank / M / F) plus gender_alternative let the rewrite adjust how it describes a person's gender, which shows up as a recurring theme across this pack - GenderControlOutput and the README's ModifyTextGender do the same thing at the text level without the LLM call. request_exception_handle is the one to actually think about: bypass (default) quietly swallows a failed API call and lets the graph keep running, raise_exception stops the queue dead, output_exception writes the error into the output string instead of your enhanced prompt. The single output, generated_content, is a plain string - wire it straight into a CLIPTextEncode or wherever your positive prompt normally lands.

Install: search "ComfyUI-utils-nodes" (or "Some Utils for ComfyUI") in ComfyUI Manager, or cd ComfyUI/custom_nodes && git clone https://github.com/zhangp365/ComfyUI-utils-nodes and restart. You need a Gemini API key - free from Google AI Studio (aistudio.google.com), drop it into the api_key input or a gemini_config.json file so you're not pasting it into every workflow you share.

Where this actually bites you:

• Leaving request_exception_handle on bypass during setup is the classic trap - a bad key or a network hiccup fails silently, and you won't notice until you're staring at a generation that used your raw, un-enhanced text instead of the rewrite you expected. Flip it to raise_exception while you're getting the node working, then decide whether you actually want silent failures in production.
• Google's content filters are famously aggressive on the image side of its stack, and the text side isn't exempt - if your source caption or your gender_prior choice pushes into territory Gemini's safety layer doesn't like, you'll get a refusal or a mangled rewrite back. That's Google's policy, not a bug in the node.
• Free-tier Gemini API keys have real rate limits. If you're running this inside a batch queue, expect throttling on anything beyond light, interactive use - there's no built-in retry/backoff beyond the exception-handling mode you pick.
• enabled (default true) is the pack's standard bypass toggle. Turn it off and the node presumably skips the API call, which is handy for quickly A/B-testing enhanced vs. raw prompts without rewiring the graph.