🤖 VRGDG LLM Multi 🤖

If the VRGameDevGirl workflow has a brain, it's this node. VRGDG_LLM_Multi is the thing that takes your reference image and transcribed lyrics and turns them into the per-scene video prompts that eventually drive LTX generation. The README's music-video flow calls it the "Prompt Creator" - "customize it to define style, mood, and scenes" - and this is the node wearing that job title. It's one prompt box you can point at any major model provider, with vision input and even image output, all in a single node.

Here's the thing to know up front: it's an API node, not a local model. It calls hosted providers and needs your own API key. No key, no output. That's not a bug, it's the design - the pack leans on frontier models for the hard creative reasoning (scene planning, style direction) and leaves the heavy video lifting to local LTX.

What it does

The inputs that matter:

• api_key (STRING) - your key for the selected provider. Keep it in an environment or your preferred key manager if you can, because it lives in the workflow file otherwise.
• provider (enum) - openai (default), anthropic, google, xai, deepseek, openrouter, or apifreellm. The last one is the "free tier" option people use for testing; everything else needs a paid key.
• model (enum, 25 choices) - the big list: OpenAI (gpt-4o, gpt-4.1, o4-mini, gpt-5-nano, gpt-image-1), Claude (claude-sonnet-4-20250514, claude-3-7-sonnet, claude-3-5-haiku), Gemini (gemini-3-pro-preview, gemini-2.5-pro, gemini-3-flash-preview, and image-capable gemini-3-pro-image-preview), plus a dozen more.
• custom_model (STRING) - if your model isn't in the list, type its id here. This is your escape hatch for anything the enum hasn't caught up with.
• prompt (STRING, multiline) - the actual instruction. Default: "Write a concise answer."
• image1..image4 (IMAGE, optional) - up to four images you can attach. This is how the reference character shot gets in front of the model for vision models.

Outputs: text (STRING) is the answer you actually use. used_provider and used_model report what really ran - invaluable when you've wired an OpenRouter key but left the provider on openai and you're getting 401s. status carries the request result, and image (IMAGE) comes back populated when you use an image-generation model like gpt-image-1.

Wiring it into the pipeline

The canonical setup: reference image → image1. Lyrics string (from VRGDG_LoadAudioSplit_HUMO_TranscribeV3 or a text loader) → concatenated into the prompt. Out comes per-scene prompt text that feeds the video side. Because the node returns plain strings, you can tee the output to VRGDG_LLM_OutputSaver to persist what the model wrote, which matters - if a good run happens at 2am you want the prompts saved.

Installing it

It's part of the comfyui-vrgamedevgirl pack - ComfyUI Manager, search "vrgamedev", install, restart. Or:

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

then install the README's requirements:

pip install -r custom_nodes/comfyui-vrgamedevgirl/requirements.txt

Common gotchas

• 401s are almost always a provider mismatch. You set the key for one provider but the provider dropdown says another. Check used_provider in the output - it tells you the truth.
• Image inputs only work on vision models. Feeding image1 to a text-only model wastes the input or errors out. Match the images to a model that advertises vision.
• The enum is a snapshot. Models get renamed and deprecated constantly (that's the state of the industry). When your model id isn't listed, use custom_model instead of fighting the dropdown.
• It costs money. You're burning tokens on every queue run, and the music-video loop calls this node a lot. Set status up in a monitor if you want to see what's being spent on which provider.