# ComfyUI VLM Nodes

ComfyUI VLM Nodes adds vision-language model nodes to ComfyUI, covering image and video captioning, structured prompting, open-vocabulary object detection, and segmentation, across a curated set of current small and large models.

ComfyUI VLM Nodes provides production-oriented vision-language, structured-prompting, audio and utility nodes for ComfyUI, supporting NVIDIA CUDA, AMD ROCm, Apple Metal, Intel XPU and CPU without replacing ComfyUI's own PyTorch build, and using ComfyUI's model residency and offloading rather than its own startup installers.

Its Modern VLM node offers one stable interface with a small, curated selection of current vision-language models, including Qwen 3.5 and Qwen 3 VL, SmolVLM2, Liquid LFM2.5-VL, InternVL 3.5, Granite Vision and Gemma 3, defaulting to a smaller, faster Qwen model that still handles broad image and video understanding; a separate legacy-compatibility node keeps older, superseded or very large models available for existing workflows without cluttering the main picker. Maintained nodes elsewhere in the package cover Florence-2, Moondream2, JoyTag, GGUF models through llama.cpp, object detection, segmentation, tracking, hosted API models and video-intelligence tasks.

The current version streams decoded text live through ComfyUI's native progress channel and processes real image and video batches rather than one frame at a time. It is written in Python.

Website: https://registry.comfy.org/nodes/comfyui_vlm_nodes