Text to video prompts from concepts

This is the node that does the "thinking" in VRGameDevGirl's AI music video workflow. You feed it a concepts file and it hands you ready-to-run text-to-video prompts. The name says it: Text to Video Prompts From Concepts - it takes the broad ideas you defined for your video and lets a local LLM expand them into the kind of scene prompts a video model actually listens to.

The twist that makes it worth reaching for: the model is a Gemma GGUF running entirely on your machine. The name "VRGDG_T2VPromptsFromConcepts" sounds like it should be calling some API, and it isn't. No key, no cloud, no per-prompt billing. If you've ever priced out prompt-generation APIs while iterating on a multi-scene video, you know exactly why that matters - a music video can burn dozens of prompt drafts before one scene lands.

How it works

The node has exactly one input worth knowing about:

• model_file - a dropdown listing Gemma GGUF models it finds in models/LLM. That's the whole config surface. Pick your model and go.

Under the hood it reads a ConceptPrompts.txt (the concepts you wrote - song theme, character, scenes) and runs it through the Gemma model with a prompt template tuned to spit out structured text-to-video prompts. Those prompts then flow into the rest of the pack: the VRGDG_ThemeSplitter node splits a context block into environment, lighting, camera motion, and the other per-scene ingredients, and the VRGDG_UpdateLatestCombinedJsonPrompts node manages the batch of prompts that drive generation.

If you're not running the full music video workflow, this node is still a nice standalone pattern: a local LLM that turns a short brief into 10-20 scene prompts for LTX, Wan, or whatever T2V model you use. The reddit-verified pattern of running a Gemma GGUF inside ComfyUI purely for prompt generation is well established, and this is that idea packaged as a single node.

Installing it

This node ships inside the VRGameDevGirl Video Enhancement Nodes pack. Easiest path: ComfyUI Manager → Install Custom Nodes → search vrgamedev (or paste the repo URL). Manual install is the standard dance:

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

Restart ComfyUI. The pack's requirements.txt pulls in kornia, librosa, and imageio - Manager handles these, but if you cloned by hand and get import errors, run pip install -r requirements.txt against your ComfyUI's Python.

The one extra thing this specific node needs is a model:

1. Grab a Gemma GGUF (a 2-4B quant works fine for this job - you don't need the 27B for prompt templating).
2. Drop it in ComfyUI/models/LLM/ (create the folder if it isn't there).
3. Restart ComfyUI so the dropdown picks it up.

Common issues

The dropdown shows [No Gemma GGUF found in models/LLM]. That's the node's own default, and it's a message, not a bug - it means exactly what it says: no Gemma GGUF is in models/LLM. Drop one in and restart.

Nothing happens and there's no output wire. Note this node has no outputs in its schema - it's a side-effecting node that writes its generated prompts (or updates the concepts state) for downstream nodes to pick up. Don't hunt for a missing output socket; wire the pack's other prompt nodes and they read what this one produced.

Slow first run. The first load of any GGUF LLM caches and quantizes, so the first generation takes noticeably longer than later ones. That's normal.

It's a small node with a single real job, and it does it locally, which is honestly the right call for a workflow that regenerates prompts constantly.