🌀 VRGDG_MusicVideoPromptCreatorJson

VRGDG_MusicVideoPromptCreatorV3 is the node that turns lyrics, a character, and a pile of style lists into a full set of scene prompts for a music video. It's an LLM call with a fancy front end - and that means it needs a Gemini API key, which is the first thing people miss about it.

Let's be upfront about the architecture, because it changes everything: this node does not generate prompts locally. It sends your inputs to Google's Gemini API and returns the assembled scene prompts as the single concatenated_string output. The name and the "V3" and the "Json" in the display name hint at a long iteration history; what you get today is the orchestration hub of the whole pack. No key, no output - it just errors. Get one from the Google AI Studio / Gemini API console; it's a paid API, though cheap for text.

Inputs that matter

The defaults are thoughtfully pre-filled, which is rare and good. You'll change a handful:

• character_description - who the video stars. Default "The Woman." Keep it tight; this gets repeated into every scene prompt.
• song_theme_style - the aesthetic envelope: mood words, genre feel, color language.
• pipe_separated_lyrics - your lyrics as line1 | line2 | line3. This is where cleaned lyric output from the extractor/cleaner nodes lands, and it's what makes the video actually follow the song.
• word_count_min / word_count_max - how long each scene prompt should be (defaults 30–50 words). LTX-style video models read best at these lengths; don't crank them up thinking longer is better.
• list_handling_mode - how the lists below get used: Strict Cycle (each item used once, then repeat), Reference Guide (LLM writes variations inspired by the list - the default), Random Selection, or Free Interpretation.
• prompt_structure_mode - the sentence skeleton: character-first, environment-first, action-first, rotating, or LLM's choice.

Then come the content lists, all multiline: environment, lighting, camera_motion, physical_interaction, facial_expression, shots, outfit_rules, character_visibility. Each ships with a solid default pool, and these are the dials that actually shape the look - swap the environment list and the whole video relocates.

The story mode

Toggle story_mode and the node tracks continuity across runs using summary files. The optional Summary_File_Path + summary_index load a previous run's summary (so a multi-part video stays consistent), and total_sets / groups_in_last_set tell the LLM which run is last and how many prompts to leave room for. The tooltip is blunt: story mode is advanced and "using Gemini PRO is recommended" - meaning expect heavier API usage and better results on the bigger model. It's the difference between "sixteen disconnected scenes" and "a sixteen-scene video that remembers what happened in scene one."

Where it sits in the pack

This is the node everything else feeds: lyrics upstream, character and style here, and the concatenated_string output flows into the chunker (VRGDG_IndexedPromptChunker) which splits it into per-scene prompts for generation. The whole music-video pipeline exists to serve this one call and distribute its result.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/vrgamegirl19/comfyui-vrgamedevgirl

Restart ComfyUI; it's under VRGDG/Prompt Tools. ComfyUI Manager: search vrgamedev. No extra local models - the dependency is the API key, not a download.

Honest notes

Three things. First, you're at Gemini's mercy for censorship and tone - Google's image models are famously filtered, and while this is text-only, the prompt text you get back still inherits the model's sensibilities; the KB's coverage of closed-source models calls Gemini "heavily censored" and "unimaginative" for image generation, and the same guardrails color its text output. Second, it costs money per run, and story mode multiplies that. Third, the README doesn't document this node at all - trust the tooltips, and test one song's worth of prompts before committing to a long run. Despite the caveats, it's the most genuinely useful node in the pack: the defaults are good, and the result is a shot list you'd otherwise spend an hour hand-writing.