🧠 VRGDG Qwen GGUF 🧠

Every music-video workflow VRGameDevGirl ships needs a writing step: rough idea in, cinematic LTX or Wan prompt out, or a caption for a LoRA training set. The VRGDG Qwen GGUF node is that brain - it loads a Qwen GGUF model inside the graph and runs it through llama.cpp. No API key, no external Ollama server, nothing leaving your machine. That "run any GGUF directly as a Comfy node" pattern is exactly what people keep asking for, and this is a working version of it.

It's worth knowing the name is a slight lie in the other direction: nothing here is Qwen the image model. This is the language model - Qwen3-class - being used as a text worker. It doesn't generate frames, it writes the prompts and captions that generate frames.

How it works

Under the hood it's a thin wrapper around llama-cpp-python. Pick a model, and the node resolves it to a .gguf file (local path, local folder, or Hugging Face repo), hands it to llama.cpp's Llama() loader, and runs create_chat_completion. The loaded model is cached per file + context + GPU-layer + thread combo, so a long workflow doesn't reload it on every pass.

Two details separate this from a raw llama.cpp call. First, it's Qwen-aware: the chat template uses Qwen's native <|im_start|>/<|im_end|> format, and when enable_thinking is on it opens a <think> block so the model can reason - then strips that block out of what it returns, so you get the answer, not the monologue. Second, when you feed it images it builds a multimodal handler around the same loader, using a separate mmproj projector file the way vision Qwen GGUFs require.

The inputs that matter

• model_preset - defaults to unsloth/Qwen3.8-27B-GGUF. That's a ~27B model; it will be the VRAM hog of your workflow. The other option is custom, which lets you point custom_model_id at any local .gguf, folder, or repo.
• gguf_filename - set this. The preset repo ships thirty quant files; if you don't name one, the node can't pick. Qwen3.8-27B-Q4_K_M.gguf is the usual compromise, Q8_0 if it fits.
• task_preset - text_to_image, text_to_video, image_to_video, image_edit, captioner_training, or custom. Each injects a long, well-built instruction block, so all you type in user_input is the rough idea. trigger_word only matters for the captioner preset (LoRA training).
• image_count - how many of the 24 image1–image24 inputs to show. Needed for image_edit/captioning; the projector file is only required when images are actually attached.
• download_if_missing - lets it fetch a missing GGUF into ComfyUI/models/LLM/GGUF. The first run of a 27B quant is a multi-GB download; don't mistake that for a hang.
• enable_thinking - off is faster and cleaner for prompt writing. Flip it on when the task genuinely benefits from reasoning.

Flip advanced to reveal n_ctx (default 8192), n_gpu_layers (default 99), n_threads, temperature, top_p, and max_new_tokens. Leave them alone until something misbehaves.

Outputs: text (the prompt/caption - wire it into your text encoder or a display node), used_model (the exact file it loaded, handy for multi-model setups), and status (ok or error: …).

Installing it

ComfyUI Manager is the easy path - search vrgamedev and install, or clone manually:

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

Restart ComfyUI and hard-refresh the browser page. This pack's requirements.txt includes llama-cpp-python and huggingface_hub, which is where the real friction lives: on Windows portable builds you'll want Cython and scikit-build-core installed first, and llama-cpp-python may need a working CMake/Ninja and C++ compiler when no prebuilt wheel exists. Python 3.12 beats 3.13 for older portable setups.

Where people get burned

• VRAM math. A 27B Qwen plus LTX or Wan on one card is a lot. The node has an unload_after_run toggle that frees the model after each pass - use it, or your sampler will start OOMing on the first frame.
• The projector filename. The default preset expects qwen-mmproj-BF16.gguf, but the file actually in the repo is mmproj-BF16.gguf. If vision-mode download 404s, set mmproj_filename to the real name.
• Download errors. Missing files surface as clear errors; a 403 means a gated repo and a missing hf_token, not a broken install.
• Multi-quant repos refuse to auto-pick - that's the gguf_filename requirement above, not a bug.

One last thing: an LLM node is exactly the shape of the thing that got weaponized in the ComfyUI_LLMVISION malware incident, so the usual rule applies - it's open AGPL source, and you should glance at what you're running. Once it's up, it's the most useful prompt writer in the pack, and it never bills you per call.