Load Video Advanced

This is the friendly, point-and-click way to get a video into Qwen3-VL. Drop a clip in your ComfyUI input folder, pick it from the dropdown, and the node loads it and outputs a VIDEO object plus a PATH. That PATH is the payoff: it wires straight into the pack's captioning node, Qwen3 VQA, via its source_path input, so the vision model can watch the clip and answer questions about it. Think of it as Load Image, but for video.

It has a twin - Load Video Advanced (Path) - and they do the same job with the same outputs. The only difference is how you choose the file. This one gives you a dropdown of your input folder; the Path version takes a raw filesystem path you type in. If you just want to caption a video you dragged into ComfyUI, this is the node you want. If your file lives somewhere else on disk or your workflow is automated and computing paths on the fly, use the Path variant instead.

How it works and what it outputs

The single input is file - an enum, i.e. a dropdown, populated from the videos ComfyUI finds in your input directory. Pick one and the node decodes it for the rest of the graph. (If you introspect the node with an empty input folder you'll see zero choices; that's not a bug, there's just nothing there to list yet.)

Two outputs, and knowing which is which saves you a confused half hour:

• PATH - this is the one you connect to Qwen3 VQA's source_path. It's what turns "I loaded a video" into "the model actually captioned it." The VLM samples frames from the clip and responds to your text prompt.
• VIDEO - a normal ComfyUI video object for previewing or feeding other video nodes. Qwen3 VQA doesn't take a VIDEO input, so for captioning specifically it's the PATH you care about, not this.

There are no trim or frame-rate controls on the loader itself - how many pixels per frame the model actually looks at is set downstream on Qwen3 VQA (its min_pixels / max_pixels budget).

Install

Nothing to install for this node on its own - it comes with the pack. ComfyUI Manager → search Qwen3 → install → restart, or cd ComfyUI/custom_nodes && git clone https://github.com/IuvenisSapiens/ComfyUI_Qwen3-VL-Instruct, then pip install -r requirements.txt and restart. The Qwen3-VL weights fetch themselves on first run into ComfyUI/models/prompt_generator/; the loader needs no separate download.

Common issues

The classic one: the dropdown is empty or doesn't show your file. That's the input folder - the video has to actually be in ComfyUI's input directory, and if you just added it you'll usually need to refresh the node or reload the page for it to appear. This is also the exact moment people realize they wanted the Path variant instead, because their video is somewhere else entirely and they don't want to copy it in.

After that, the failures aren't the loader's - they're VRAM. A long clip is a lot of frames, and every frame is more work for the vision model downstream. On a mid-range card, pairing a long video with anything past the 4B model can freeze the whole workflow; 16GB users specifically report the 8B model struggling on longer clips. If it hangs after the video loads fine, the loader did its job - trim the clip or step down to the 4B Instruct model on the Qwen3 VQA side.