🖼️ Local Image Analysis (GGUF)

The local GGUF vision loaders in this pack all funnel into one place, and this is it. Local Image Analysis is where you take the VISION_MODEL config from the 🖼️ Local Vision Model Loader, hand it an image - or a video, which is the rare trick - and get back actual text describing what's in front of it. It's the node that makes the whole local-vision half of the pack worth installing.

What it takes

Required inputs are the model config plus the usual generation knobs:

• model (VISION_MODEL) - the output of the Local Vision Model Loader. This is what makes it a local analysis: everything runs through llama-cpp-python on your machine.
• prompt - default "Describe this image in detail." Your actual instruction lives here.
• max_tokens - default 512, cap 4096. Fine for captions; raise it for longer reasoning.
• temperature (0.7), top_p (0.9), top_k (40), seed - standard sampling controls. For consistent captions, drop temperature toward 0.1 and fix a seed; for varied creative description, leave them.

The optional inputs are where it gets interesting. image and video are marked "choose one" - the video input is a frame sequence (IMAGE type), and this is the local path's video-analysis capability, which works with Qwen2.5-VL's video understanding. There's also an optional system_prompt defaulting to a describe-images-and-videos persona, which you can override or replace with the System Prompt Config node.

What comes out

A single context string - the model's full response. It's an output node, so it displays in the UI, and the string flows onward for Save Text nodes, prompt builders, or further processing.

The setup that matters

Because this runs locally, the heavy dependency applies: you need llama-cpp-python with CUDA. Plain pip install llama-cpp-python gives you a CPU build, and you'll wonder why a 7B model crawls. Install the CUDA wheel matching your driver:

cd ComfyUI/custom_nodes
git clone https://github.com/walke2019/ComfyUI-GGUF-VLM.git
cd ComfyUI-GGUF-VLM
pip install -r requirements.txt
pip install llama-cpp-python --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cu126

Then the model pairing: drop a Qwen2.5-VL GGUF and its mmproj file into models/LLM/GGUF/, load it in the loader node, and connect. Remember the pack's local vision support is Qwen2.5-VL specifically - a Qwen3-VL GGUF here will misbehave, and you'd want the Transformers loader for that family.

Where people get burned

Two classics. Forgotten mmproj: if analysis errors or returns nothing, the loader probably failed to auto-match the projector file - check the naming (see the loader node) or set mmproj_file manually. And CUDA not installed: if everything runs but takes forever, you're on the CPU build. Both are one command or one filename away from working, which is why this node is so often the last one people get running - and then it's smooth.