🧠 VRGDG General GGUF 🧠

The VRGameDevGirl music video workflow doesn't ask you to write 40 scene prompts by hand - a local LLM writes them, and VRGDG General GGUF is that LLM. It's a llama.cpp-based GGUF text-generation node with built-in task presets so the model knows whether it's writing an image prompt, a video prompt, or a training caption. No API, no key, no cloud. If it can fit in your VRAM, it runs entirely on your machine.

How it works. Pick a model_preset - it ships with unsloth/gemma-4-26B-A4B-it-GGUF and a "supergemma" community repack as options, or set custom_model_id to point at any local .gguf file, folder of GGUFs, or Hugging Face repo id. gguf_filename disambiguates when a repo has multiple quant files - and you'll want to use it, because GGUF repos are quant ladders (Q4 through Q8 and friends). The node then runs the model through llama.cpp with a task_preset telling it what kind of output to produce: text_to_image, text_to_video, image_to_video, image_edit, or captioner_training - each one carries its own instruction block. custom lets you write your own instructions from scratch. user_input is where your actual creative direction goes.

The vision side. Set image_count to a number up to 24 and that many optional image inputs appear; for vision-capable GGUF models you'll also want mmproj_filename - the multimodal projector - or the model can't actually see the images. That's how the workflow's "here's my reference character, write scenes that keep her consistent" trick works: the LLM looks at your frames and writes prompts that stay on-model.

The knobs that matter. For a beginner, the defaults are fine: temperature 0.6, top_p 0.95, max_new_tokens 800. n_ctx (8192 default) is the context window. n_gpu_layers defaults to 99, which just means "offload everything"; on a 12GB card you may want to drop it and let llama.cpp split across CPU and GPU. download_if_missing will fetch a missing GGUF into ComfyUI/models/LLM/GGUF - handy, but a 15GB+ download without warning isn't fun, so keep an eye on it. unload_after_run frees the model from cache when you're done - worth turning on if this node shares VRAM with your video generation.

Outputs. text is the generated prompt - the one you wire into your sampler's positive prompt. used_model tells you which file actually loaded (the lifesaver when a preset points at something you don't have), and status reports what happened.

Installing. The pack:

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

Then install requirements - this node needs llama-cpp-python, which is the one dependency that actually fights back on Windows portable installs. If pip install llama-cpp-python refuses to build, grab a prebuilt wheel for your Python version rather than forcing a compile. On the author's own workflow requirements, you also want FFmpeg in PATH for the surrounding video steps.

Where people get burned. The "gemma-4-26B-A4B" preset is an active-parameter model - 26B total, ~4B active - so it runs far smaller than its name implies, but the A4B naming still confuses people into thinking it's a toy. It's not. Second: if you swap models, change both the repo preset and gguf_filename; leaving a stale filename is the classic "loaded a 40GB repo file I never wanted" move. And if the node outputs gibberish or refuses images, check mmproj_filename first - a vision model without its projector is a model that can't see.