🧠 VRGDG SuperGemma GGUF Chat 🧠

VRGDG_SuperGemmaGGUFChat loads a Gemma model as a quantized GGUF file and runs it for text generation directly inside ComfyUI. "SuperGemma" is the author's branding - you won't find it on a model card, and that's fine; underneath it's a plain llama.cpp-backed Gemma chat node with task presets bolted on.

Why Gemma specifically? Gemma 3 is the text encoder inside LTX-2 and LTX-2.3, and its full fp16 weights are a 22GB monster that causes most of the OOM errors in the LTX community. Running a GGUF quantization of it - Q8 is basically fp16 at half the size, as the GGUF crowd likes to say - is the low-VRAM path. This node gives you that same model for prompt enhancement and captioning without a separate llama.cpp setup.

The gotcha you'll hit first

The model_file dropdown defaults to [No Gemma GGUF found in models/LLM]. That's not a bug - it's the node telling you, honestly, that you haven't put a model where it looks yet. Drop a Gemma .gguf into ComfyUI/models/LLM/ (only files with "gemma" in the name are listed), restart or refresh, and the dropdown fills in. If you want to send images to the model, you also need an mmproj_file - the multimodal projector GGUF, also under models/LLM - but that's only required when you actually use image inputs.

The inputs that actually matter

This node has a lot of fields, but you can ignore most of them.

• task_preset - the one to understand. Pick text_to_image, text_to_video, image_to_video, image_edit, or captioner_training, and the node injects built-in instruction blocks tuned for that job. custom uses your own custom_instructions. For the music-video workflow, text_to_video and image_to_video are the scene-prompt writers; captioner_training is for captioning LoRA training data.
• user_input - your actual task details and creative direction.
• image_count - how many of the optional image slots (up to 24) to show on the node. Turn it up if you're handing the model a batch of reference frames.
• advanced - flips on the GGUF runtime controls. n_ctx defaults to Gemma's full 262144-token training context, which is enormous; if RAM or VRAM pressure is high, lower it hard (16384 is plenty for prompt work). n_gpu_layers 99 means offload everything to GPU; -1 also offloads all supported layers. unload_after_run frees memory after each generation - worth enabling if the workflow then wants those GBs back for video.

Outputs: text (the response), used_model, and status.

Installing

Shared pack install:

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

or Manager → search "vrgamedev", restart. No extra Python deps beyond the pack's shared ones - the GGUF runtime is bundled with the node.

The honest verdict

This is the right tool when you want Gemma-class prompt enhancement or captioning inside ComfyUI without juggling an external process, and the task presets genuinely save time. Where people get burned: forgetting the model file (the empty dropdown), leaving n_ctx at 262144 on a 16GB machine, and expecting the mmproj to be optional when they've wired up images. Mind those three and it's a smooth, fully-local Gemma. It's also the rare node where "SuperGemma" being an in-house name is worth knowing - so when you google it and find nothing, you're not missing a secret model, you're just looking at a well-nicknamed wrapper.