ComfyUI_VLM_nodes

Why we included this project

If you run ComfyUI and want vision-language features without adding a separate runtime, this node pack brings image captioning, open-vocabulary detection, and segmentation into your existing pipelines. It wraps a curated set of modern VLMs, including Qwen3-VL, SmolVLM2, InternVL, and Gemma 3, behind a single node interface, with a deliberately small production picker and a legacy tier that keeps older saved workflows working. The same install also covers grounding-style detection, SAM2/SAM3 segmentation, and video temporal reasoning, so one setup handles several vision tasks instead of requiring separate tools. The maintainers favor official model repositories and native Transformers interfaces over unverified community quantizations, which helps if reproducibility matters to you. And because the nodes respect ComfyUI's existing device and model-residency handling, they run on CUDA, ROCm, Apple Metal, or CPU without forcing a separate runtime.