🧠 VRGDG Qwen 2.5 🧠

This is the node that makes the whole "fully automated" music video workflow actually work. VRGDG_Qwen2.5 runs a Qwen instruction model locally inside ComfyUI and uses it to turn your lyrics and creative direction into the per-scene prompt maps that the rest of the pack consumes. No API, no key, no cloud - it downloads a model and runs it on your GPU.

The name is a small lie you should know about up front: it says "Qwen 2.5," and the presets are all Qwen2.5 variants - but the default model_preset value shown in the schema is Qwen/Qwen3.5-4B, a string that isn't in the preset list at all. That's a leftover default from a newer model line, and it's a sign this node is evolving fast. Pick a preset from the actual list rather than trusting the default, and you'll be fine.

How it works

Pick a model, pick a task, give it your input, and read the text output. The inputs that matter:

• model_preset - Qwen/Qwen2.5-VL-3B-Instruct, Qwen/Qwen2.5-VL-7B-Instruct, Qwen/Qwen2.5-3B-Instruct, Qwen/Qwen2.5-7B-Instruct, Qwen/Qwen2.5-14B-Instruct, or custom. The VL (vision-language) models can look at images - that's how the workflow keeps your reference character in mind. The 3B models fit comfortably on midrange cards; 14B wants serious VRAM.
• custom_model_id - override with any Hugging Face repo id or a local model folder path. If you have a quantized local model, point at it here.
• task_preset - text_to_image, text_to_video, image_to_video, image_edit, captioner_training, or custom. Each ships built-in instructions tailored to that job; the captioner one is for building training captions and is where trigger_word (a LoRA trigger token) gets used.
• custom_instructions - your own full instruction block, used only when the task is custom.
• user_input - your actual task details and creative direction.
• image_count - how many optional image inputs to reveal, 0 to 24. Set 1–4 and the node shows image1…imageN ports for reference images (the VL models use them).
• download_if_missing - allow the node to fetch missing models into ComfyUI/models/LLM/Qwen. First run with this on is a big download; plan accordingly.
• device (auto/cuda/cpu), dtype (auto recommended), temperature (0.6 default), top_p (0.95), max_new_tokens (800 default, up to 32000).

Outputs: text (the generated response), used_model (which model actually ran - the honest answer to "what am I really running"), and status (a string describing the run).

The workflow fit

In the shipped music video workflow, the flow is: lyric transcription → the template builder and prompt creator hand creative direction to this node → it writes the scene prompt map → the fixer repairs it → the splitters hand scenes to the video generators. The community's reports on the pack confirm the LLM-driven flow: "it utilizes LLM for prompts, fully automated," and the automated prompts come out surprisingly coherent until you hit character consistency. If you're going to review anything before a long render, review what this node produced.

Gotchas

• VRAM is the whole game. 14B is a lot; if you're on 8–12GB, stick with 3B or 7B and let dtype: auto pick the precision. Watch ComfyUI's console on first run - a missing-model error means download_if_missing was off.
• The default preset string is stale (see above). It says Qwen3.5-4B which isn't in the list - choose explicitly.
• 24 image inputs is overkill for any real workflow. Set image_count to what you actually have (usually 1) to keep the node sane.

Installing it

It's the flagship node of the pack, so install the pack:

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

restart ComfyUI, or use ComfyUI Manager → search vrgamedev. Then enable download_if_missing and let the first run pull the model into ComfyUI/models/LLM/Qwen - or download it manually to that folder. The pack's requirements.txt (kornia, librosa, imageio) covers the workflow's audio and image side; the node itself uses ComfyUI's bundled transformers stack.

The honest take

This is the highest-impact node in the pack and the heaviest one. It's the difference between "I typed a brief" and "I typed a brief and got thirty coherent scene prompts." If you only care about the video-enhancement half of this repo, skip it; if you're here for the music video workflow, it's the engine.