Load Llama Vision Model

Load Llama Vision Model is the middle sibling of the pack's three dedicated loaders. It scans ComfyUI/models/LLM and shows only models whose config.json architecture is MllamaForConditionalGeneration - that's the class Llama 3.2 Vision checkpoints use - then hands you a VISION_MODEL ready for the Llama-specific generation node.

One input (model_name, filtered dropdown), one output (VISION_MODEL). Nothing else to configure, and no trust_remote_code toggle to worry about - Llama Vision doesn't need custom code, and that's the point of the dedicated loader: fewer knobs, fewer ways to break.

Why you'd pick Llama Vision over Pixtral

The README's honest rundown: Llama 3.2 11B Vision is noticeably better than the nf4 Pixtral build at OCR - reading text off images - which the author flags as Pixtral's weak spot. It's also a fine general captioner. The tradeoffs: it only takes images in a single block at the start of the prompt (no inline placement), and multi-image understanding is weak - the README says image comparison is effectively unusable, with the model mixing up left/right and colors between two images. So: pick this loader for OCR and single-image work, not for "compare these two."

Like the other loaders, it's lazy: the processor loads here, and the actual model weights load when the generation node executes.

Model placement

Each model is a folder in ComfyUI/models/LLM with the standard HuggingFace layout - config.json, model.safetensors, tokenizer files. The nf4 quant is what most people run:

ComfyUI/models/LLM/Llama-3.2-11B-Vision-Instruct-nf4/
├── config.json
├── model.safetensors
└── ...

SeanScripts publishes Llama-3.2-11B-Vision-Instruct-nf4 for this pack. A full 11B model is a lot of VRAM; the 4-bit build is the realistic default.

Install

ComfyUI Manager search: ComfyUI-PixtralLlamaVision, or:

cd ComfyUI/custom_nodes
git clone https://github.com/SeanScripts/ComfyUI-PixtralLlamaMolmoVision

Needs transformers >= 4.45.0, accelerate, bitsandbytes, torchvision. Windows portable:

python_embeded\python.exe -m pip install -r ComfyUI\custom_nodes\ComfyUI-PixtralLlamaMolmoVision\requirements.txt

Restart and you're set.

Troubleshooting

• Empty dropdown - model not in models/LLM, or its architecture isn't MllamaForConditionalGeneration. Check config.json; if it says something else, this loader will never show it.
• OOM - 11B is heavy. Use the nf4 build and consider unload_after_generate on the generation node.

Load it, wire it into Generate Text with Llama Vision, and point it at text-heavy images - that's where it earns its keep.