Qwen2.5-VL Inference (Image+Prompt→Text)

This is the one node in the pack you'll actually reach for every day. Feed it an image, give it a prompt, and Qwen2.5-VL-7B reads the picture and writes back a text answer - no API, no key, no cloud. If you're on a laptop with an Intel CPU and nothing else, it's also about the only way to run a 7B vision-language model in ComfyUI that isn't a ten-minute wait.

The pack is the Qwen2.5-VL half of blackmeat1225's OpenVINO experiments - a hobbyist who's made it his whole thing to drag AI generation onto Intel iGPUs (think i5-1135G7-class machines). It was assembled with Claude/DeepSeek/Gemini doing the heavy lifting, which explains both what it gets right (clean, working glue around a great model) and the rough edges in the README. Expect a show-and-tell level of polish, not a production SDK.

How it works

Behind the scenes it's a small shim over Intel's OpenVINO GenAI pipeline. Your image gets scaled down to a 512×512 thumbnail, and the model's vision tower + language model run the show: pipe.generate(prompt, image=..., max_new_tokens=...) returns whatever Qwen says, and that one string is your output.

The model is helenai/Qwen2.5-VL-7B-Instruct-ov-int4, an INT4-quantized OpenVINO conversion of the 7B instruct model. First run downloads it automatically into ComfyUI/models/openvino/Qwen2.5-VL-7B-Instruct-ov-int4 - that's a multi-gigabyte download, so don't panic when the first queue run looks stuck on "downloading model". If you already have the files somewhere, point model_path at the folder instead; the node accepts an absolute path and skips the download.

Inputs and outputs

The three you'll touch:

• image - any IMAGE tensor. Batch images get collapsed to the first frame, and the node thumbs everything down to 512, so detail is capped.
• prompt - your question, multiline. Default is Can you describe the image?, which is exactly the boring-but-useful thing to start with.
• max_new_tokens - how many tokens the model is allowed to write (default 500, hard-capped at 512 in the schema).

model_path and device are there too - leave model_path on the default unless you know better, and know that OpenVINO's "GPU" means Intel iGPU, not your NVIDIA card. The single STRING output is the model's answer; wire it into a Show Text node to read it, or save it and feed it forward into a CLIP text encoder for the classic "describe the image, then regenerate it" workflow.

Install

ComfyUI Manager is the easy route - search ComfyUI-Qwen2.5-VL-7B-OPENVINO and install. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/blackmeat1225/ComfyUI-Qwen2.5-VL-7B-OPENVINO
cd ComfyUI-Qwen2.5-VL-7B-OPENVINO
pip install openvino-genai pillow

Then restart ComfyUI. The author explicitly recommends a fresh virtual environment, and that's not paranoia - openvino-genai is a heavy, opinionated dependency and it plays badly with crowded ComfyUI envs. There's no requirements.txt in the repo, so this manual install is the real install. Ignore the clone URL in the README: it points at the older Z-Image turbo repo.

Common issues

• Nothing happens on first run - the model download is slow and silent. Give it time, then check models/openvino/ for the folder.
• The node errors on load - the folder you pointed model_path at must contain the OpenVINO IR files like openvino_language_model.xml. Point it at the auto-downloaded folder and it just works.
• It's slow - it's a 7B model on a CPU. It's OpenVINO-fast, not fast. Roughly the point of the pack is that this is usable where GGUF wasn't.
• You have an NVIDIA GPU - this pack isn't for you. OpenVINO runs on NVIDIA but you gain nothing; a standard Qwen2.5-VL GGUF node will serve you better.

One warning that applies to all five nodes in this pack: the two description nodes swallow errors and return the string "❌ 模型推理失败: …" instead of failing loudly. If your text output suddenly looks like an error message, that's the node telling you the model couldn't run - check the console for the real traceback.