Load Video Advanced (Path)

This is the plumbing node that lets Qwen3-VL watch a video. You give it a path to a video file, it loads that file, and it hands the rest of your workflow two things: a VIDEO object and a PATH. The whole reason it exists is that the pack's captioning node, Qwen3 VQA, has an optional source_path input - and this loader's PATH output is exactly what plugs in there. Loader in, question node next, text description out.

The "(Path)" in the name is the important bit. This is the variant that takes a raw filesystem path as a string, versus its sibling Load Video Advanced, which gives you a dropdown of whatever's sitting in your ComfyUI input folder. Same outputs, same downstream wiring - the only difference is how you tell it which file to load.

When you'd reach for the Path version

Use this one when the video isn't in your input folder, or when you don't want to move it there. A clip sitting on another drive, a render your pipeline just wrote out somewhere, a batch job pointing at a stack of files by absolute path - those are all cases where typing the path beats copying the video into ComfyUI's input directory first. If you're building anything automated, the path string is the thing you can compute or template; a dropdown isn't. For the casual "I dragged a video in and want to caption it" case, the dropdown sibling is friendlier, and there's no shame in using that instead.

How it works and what it outputs

There's really one input worth naming: file, a string, which is the full path to your video. Point it at an .mp4 (or the usual formats) somewhere the ComfyUI process can actually read, and it decodes the clip so a downstream node can use it.

The two outputs:

• PATH - wire this into Qwen3 VQA's source_path. This is the connection that makes video captioning work. The vision model samples frames from the file and answers your prompt about them ("summarize this clip," "describe what happens").
• VIDEO - a standard ComfyUI video object, for previewing or handing to other video nodes if your graph does more than captioning.

That's the entire node. No trimming, no frame-rate knobs here - the actual frame sampling and how much detail the model sees is governed on the Qwen3 VQA side (its min_pixels / max_pixels budget, applied per frame).

Install

It ships with the pack, so there's nothing to install for this node specifically. ComfyUI Manager → search Qwen3 → install → restart, or cd ComfyUI/custom_nodes && git clone https://github.com/IuvenisSapiens/ComfyUI_Qwen3-VL-Instruct followed by pip install -r requirements.txt and a restart. The Qwen3-VL model weights download themselves on first run into ComfyUI/models/prompt_generator/; the loader itself needs nothing extra.

Common issues

Path problems are the whole failure mode here. If the node errors, 90% of the time the string is wrong - a typo, a relative path that resolves somewhere you didn't expect, or a file the ComfyUI process doesn't have permission to read. Use the full absolute path, and on Windows either forward slashes or escaped backslashes.

The other thing that bites people isn't the loader at all - it's what happens after. Long videos mean lots of frames, and every frame is more visual tokens for the model to chew on. Pair a long clip with anything bigger than the 4B model and a mid-range card and the workflow can freeze solid; folks on 16GB report the 8B struggling specifically on long video. If it hangs after the loader, that's a VRAM ceiling downstream, not a broken path - shorten the clip or drop back to the 4B Instruct model.