Gemini Text to Cinematic Prompt

This is the text-only cousin of the pack's Gemini Image to Prompt node, and honestly the more generally useful one. Feed GeminiTextToCinematicPromptNode ("Gemini Text to Cinematic Prompt") a scrappy sentence - "a red car racing through a neon city at night" - and it returns a properly structured cinematic prompt: subject action, character and background detail, camera angle, lens movement, lighting, color palette, laid out in chronological order and written to describe a 5-second clip.

Why you'd want that: modern video and image models read prompts as instructions, and the best results come from prompts that describe a shot the way a director would storyboard it. This node front-loads that discipline so you don't have to. It's a single Gemini API call; nothing runs locally.

How it works

Same engine as the image node, minus the image. You give it input_text, an api_key, a modelo_select dropdown (the same 7 Gemini entries - and the same stale labels: "Gemini 2.5 Flash" points at a preview id, and "Gemini" plus "Gemini 1.5 Pro" both map to gemini-exp-1206), and prompt_base. The node concatenates prompt_base with your input, sends the whole thing to Gemini, and returns the finished prompt as a STRING named text.

Worth reading the default prompt_base once, because it's a real spec: start with the main action in one sentence, then movement, protagonist detail, environment, camera angles and lens movement, lighting type and color tones, sudden changes in sequence - "use technical filmmaking language, avoid metaphors, abstractions, or vague terms, keep everything in a single fluid paragraph, max 200 words." That last bit is the good part: a 200-word single paragraph is a genuinely useful length for video prompts, and the technical-spec framing is the kind of structure that transfers well to LLM-encoded models.

Where it wires in

The output feeds anything that takes a prompt string. The obvious partner is this pack's own DeepseekR1KlingAINode - use Gemini to turn a rough idea into a polished script, then Kling's DeepSeek assistant to adapt it for the video side. If you want to review before it goes downstream, route the text through ShowGeneratedText first.

Installing it and what bites

Install is pack-wide: ComfyUI Manager, search "ComfyUI-GeminiImageToPrompt", or

cd ComfyUI/custom_nodes
git clone https://github.com/santiagosamuel3455/ComfyUI-GeminiImageToPrompt

then restart. No models, no downloads - the requirements are google-generativeai, Pillow, and requests. The only real requirement is a Google AI Studio API key, and the free tier handles a lot of these calls. Failure modes are the same ones as the image node: a missing or invalid key, or a retired model id on the dropdown. If you hit "model not found," flip the dropdown rather than your whole install. It's an unmaintained single-commit pack from May 2025, so treat the dropdown labels as suggestions, not truth.