Multiple Paths Input

If you've looked at the "Multi-Image Query" or "Video Query" examples in this pack's README and wondered how a dozen images turn into one input socket, this node is the answer. MultiplePathsInput builds a batch of file paths from however many sources you throw at it, and that batch is what the pack's Qwen-VL query node actually reads when you ask it something like "tell me a story from these three photos" or "caption this video."

What it's for

The whole pack - it's built around Qwen3-VL (the repo folder still says Qwen2_5, a fossil from before the Qwen3 upgrade) - is Alibaba's Qwen line doing vision-language work: feed it text, an image, several images, or a video, and it talks back with a description or an answer. Qwen is the same lab that gave the community Qwen-Image and Wan, but VL is a different branch - this isn't generating pixels, it's reading them. For a single image you'd just wire one path in. For "here's a folder of reference shots, describe the person across all of them" or "sample this video and summarize it," you need several paths landing on one input, and that's a batching problem ComfyUI doesn't solve natively. MultiplePathsInput solves it with a dynamic-input pattern you'll recognize from other packs: a count widget plus an Update button that grows the node's sockets to match.

How it works

Set inputcount and click Update, and the node adds path_2, path_3, and so on up to whatever number you asked for - each one a PATH socket you feed from something like this pack's own ImageLoader (which outputs a PATH alongside the image). Everything you plug in gets collected into a single paths output, ready to hand to the query node.

The four optional inputs only matter when what you're pointing at is a video rather than a folder of stills: sample_fps controls how many frames per second get pulled, max_frames caps the total pulled, and the two booleans - use_total_frames and use_original_fps_as_sample_fps - decide whether the node spreads its sampling across the video's real length and native frame rate instead of taking your sample_fps literally. Both default to true, which is the sane starting point; if you're just batching photos, ignore this whole group, it's inert for you.

The inputs and outputs that matter

• inputcount (INT, default 1, 1–1000) - how many path_N sockets to expose. Set it, hit Update.
• path_1 (PATH, required) - your first source. Every additional path_N you add the same way.
• sample_fps / max_frames (INT, defaults 1 / 2) and use_total_frames / use_original_fps_as_sample_fps (BOOLEAN, both default true) - video-only frame sampling, optional.
• Output: paths (PATH) - wire this straight into the pack's Qwen-VL query node's path input.

How to install it

Through ComfyUI Manager: search Qwen3 and install. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/IuvenisSapiens/ComfyUI_Qwen2_5-VL-Instruct
pip install -r ComfyUI_Qwen2_5-VL-Instruct/requirements.txt

then restart ComfyUI. There's no model download tied to this specific node - the LLM weights the pack needs get pulled automatically into ComfyUI/models/prompt_generator/ the first time you actually run a query node, not when you load MultiplePathsInput.

Common issues & troubleshooting

You changed inputcount and nothing happened. This node's sockets are widget-driven, not automatic - you have to click Update after changing the number, same as the "add N inputs" pattern you'll have seen on other batching nodes across the ecosystem. Changing the number alone doesn't redraw the node.

PATH won't connect to anything else. PATH is a type this pack defines for itself, not a stock ComfyUI type. It only wires into this pack's own nodes (ImageLoader's PATH output going in, the query node's path input coming out) - you can't hand it a plain string or connect it to stock LoadImage.

You set max_frames to 1 and it errors. The schema floors max_frames at 2 - you can't sample a single frame through this input. If you genuinely want one frame from a video, that's a job for a still-image path instead.

Video captions look off no matter what you set sample_fps to. With use_total_frames and use_original_fps_as_sample_fps both on by default, the node is sampling based on the video's own length and rate, not literally honoring sample_fps. If you need exact control over frame timing, flip both booleans off first - leaving them on while also fighting sample_fps is fighting the node's own defaults.

Dependency conflicts on install. This pack ships a requirements.txt like most vision-model nodes, and custom-node dependency clashes are a known ComfyUI-wide headache - if pip install throws version errors, check what else in your custom_nodes folder is pinning the same libraries (transformers is the usual suspect) before assuming this pack is broken.