🧠 VRGDG Qwen 3.5 🧠

If the rest of this pack is the movie crew, this node is the director. VRGDG_Qwen3.5 is an LLM node that runs a Qwen model inside your ComfyUI graph, and in the VRGameDevGirl music-video workflow it's what turns "make me a music video about a girl at a neon diner" into an actual scene plan: split lyrics, storyboard prompts, image prompts, captions - the whole creative layer the other nodes then execute. No API key, no cloud round-trip, no Ollama sidecar. The name is the giveaway: it's a Qwen 3.5 runner with presets tuned for this specific pipeline.

How it works

Pick a model_preset - there are 23, ranging from Qwen/Qwen3.5-0.8B up to Qwen/Qwen3.5-35B-A3B, plus the older Qwen2.5-VL-7B-Instruct. The default is Qwen3.5-4B, and that's the one to start with. 4B fits comfortably on a mid-range card and handles structured JSON-ish output fine; the 35B-A3B is a Mixture-of-Experts model and is only worth the download if you're on serious hardware. Don't reach for the biggest model first - for prompt drafting, the 4B is genuinely the sweet spot.

task_preset is where the magic lives. It's an enum with text_to_image, text_to_video, image_to_video, image_edit, captioner_training, and custom - each one ships built-in instructions that shape how the model writes. The custom_instructions field is used only when you pick custom; the tooltip says so explicitly, so don't waste time typing in there while a preset is active. user_input is your actual creative direction, the thing that gets run through the preset. There's also trigger_word, which matters only for the captioner_training preset - that's how you teach it your LoRA's trigger token.

What makes this more than a chat wrapper is image_count (0–24). Set it and the node sprouts that many optional image inputs - a vision-language model, so you can feed it your reference character shot or storyboard frames and have it describe or build on them. Handy when you want the LLM to actually look at your reference before writing prompts.

The controls that matter: download_if_missing will pull the model into ComfyUI/models/LLM/Qwen on first run (a 4B is a few GB - expect the first queue to be slow), device is auto/cuda/cpu, and dtype should stay auto. temperature defaults to 0.6, top_p 0.95, max_new_tokens 800 - lower temperature if the output keeps drifting from your direction.

Outputs are text (the generated output), used_model (which model actually ran, after any override), and status. Wire text into whatever needs the prompt or the storyboard JSON.

Install

Nothing special beyond the pack itself:

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

or search "vrgamedev" in ComfyUI Manager. The pack's requirements.txt pulls in kornia, librosa, and imageio - on a Windows portable install run that from python_embeded. Then restart, and let the node download its model on first use.

Where people trip

First run is slow because the model downloads and loads - that's not a hang. And because this pack's README barely documents the newer nodes, most of what you know about task_preset behavior comes from the YouTube tutorial the author ships and from trial and error. If download_if_missing is off and no model is present, you'll get a status string telling you nothing loaded - flip that toggle and requeue.