MiniCPM-V GGUF

The low-VRAM captioner, minus the tuning. MiniCPM-V GGUF is the plain version of this pack's GGUF node: it loads a quantized MiniCPM-V-4.0 through llama.cpp and spits out one caption, with no sampling parameters exposed. On a 4–6GB card this is often the only way you get a real vision-language model running at all, and it's about as simple as this category gets.

Mechanically it's llama.cpp doing the work: the node loads a GGUF of MiniCPM-V-4.0 plus the separate mmproj vision projector, resizes your image to 336×336, and runs it as a chat completion. Both files auto-download into ComfyUI/models/LLM/GGUF on first use.

The inputs that matter

• image / video - same as the other nodes in the pack, with a caveat: on video it only captions the first sampled frame. Treat this as an image node that happens to accept video, not a video analyzer.
• model - 10 quantization levels of MiniCPM-V-4.0, from Q4_0 (~2.08GB) up to Q8_0 (~3.83GB). The default is Q4_0, which is fine, but the pack's own README recommends Q4_K_M as the quality/size sweet spot - it's a K-quant at nearly the same size, and it's the one I'd pick. Q8_0 is essentially fp16 quality if you have the room.
• preset_prompt (21 built-ins) and custom_prompt (overrides the preset when non-empty), device (Auto/GPU/CPU), memory_management, seed. That's the whole surface - all other generation settings are locked to the pack defaults (1024 tokens, temp 0.7, top_p 0.9).

Output

A single STRING - the caption. Wire it into a text display or caption-saver and done. There's no PROMPT echo here; the Advanced GGUF node adds that plus every parameter, if you ever want them.

Installing - read this before you get confused

pip install -r requirements.txt will not give you GGUF support - llama-cpp-python is deliberately left out of that file. Without it, the GGUF nodes don't even show up in your node list. Run the pack's helper:

cd ComfyUI/custom_nodes/ComfyUI-MiniCPM
python llama_cpp_install.py

That compiles llama-cpp-python from source with CUDA if it detects a GPU, and on Windows portable it needs Visual Studio Build Tools plus the CUDA Toolkit first (guide included in the pack's llama_cpp_install/ folder). Two more things: only MiniCPM-V-4.0 has GGUF builds right now - V4.5 GGUF needs a newer llama.cpp than most llama-cpp-python releases ship, so the node will error if you try - and if you only want the transformers nodes, you can skip all of this entirely. On a big enough GPU, the transformers version is higher quality for the same effort.