Load Image Advanced

Displayed in the node picker as "Load Image Advanced," this is basically stock ComfyUI's LoadImage with one extra socket bolted on: a PATH output. That sounds minor until you realize it's the whole reason it exists - this pack's Qwen-VL query nodes want a file path, not a raw image tensor, and stock LoadImage doesn't give you one.

What it's for

This pack - built on Alibaba's Qwen3-VL vision-language model (the repo folder name still says Qwen2_5, left over from before the Qwen3 upgrade) - reads images and video and talks back with a caption or an answer to whatever you asked. To do that on a single image, the query node needs to know where the file lives, and ImageLoader is the node that gets you there without giving up the normal stuff you'd want from a loader: a preview, and a mask if the file has one. Reach for it any time you're building a single-image query workflow, or feeding one photo into a batch via MultiplePathsInput alongside others.

How it works

There's no vision model running here - this is pure I/O. Pick a file from the dropdown (populated from whatever's sitting in ComfyUI's input/ folder, same as stock LoadImage) or drop an image onto the node, and it decodes the image, pulls any alpha channel into a mask, and - the part that's actually new - reports the file's own path as a third output. The captioning happens downstream, in whichever Qwen-VL node you wire this into.

The inputs and outputs that matter

• image (required) - an enum dropdown of files already in your ComfyUI input folder. Nothing to configure beyond picking or uploading the file.
• IMAGE - the decoded image tensor, exactly like stock LoadImage. Wire it anywhere you'd normally use an image (a preview node, an image-based query).
• MASK - the alpha channel if the file has one, for standard ComfyUI masking flows. Blank/full-white on a plain JPEG or PNG without transparency - that's not a bug, it's how ComfyUI image loading has always behaved.
• PATH - the file's own path, this pack's real reason for existing. Feed it straight into a Qwen-VL query node, or into MultiplePathsInput's path_1 (or any later path_N) to batch it with other images.

How to install it

Via ComfyUI Manager: search Qwen3, install, restart. 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. ImageLoader itself downloads nothing - the pack's Qwen-VL weights land automatically in ComfyUI/models/prompt_generator/ the first time you actually run a query, not when you load this node.

Common issues & troubleshooting

Your file isn't in the dropdown. Same limitation as stock LoadImage: the enum only lists files already sitting in ComfyUI's input/ folder. It's not a file browser for your whole disk - drag the image onto the node or use the upload button and it'll land in the right place and show up.

MASK output looks empty. If the source file has no alpha channel - most JPEGs and plenty of PNGs - you'll get a blank mask. That's expected; only feed downstream masking nodes with images that actually carry transparency, or generate a mask elsewhere in your graph.

PATH won't connect to a node you'd expect. PATH is a type this pack invents for itself, not a plain string. It only wires into this pack's own nodes - the query node's path input, or MultiplePathsInput's path_N sockets - not into stock nodes that want a STRING or FILEPATH.

Missing "Display Text" node when you load the pack's example workflows. The README flags this directly: their sample workflows expect a Display Text node that ships in a sibling pack, ComfyUI_MiniCPM-V-4_5, not this one. It's unrelated to ImageLoader itself, but it's the error you'll hit if you drag in one of their example graphs and expect it to run as-is.