Load Audio

The node is called "Load Audio" and it never loads audio. No waveform decoding, no PyAV, no resampling, nothing. What AudioLoader actually does is read a list of files out of ComfyUI's input/ folder, let you pick one from a dropdown, and hand its full path to the one node in this pack that does the real work: Qwen2_AQA.

That sounds trivial, and it is. But it's the difference between typing a path by hand and getting a dropdown. The companion node takes a PATH input, and ComfyUI's PATH sockets are invisible to the keyboard - you either wire them or you type an absolute filesystem path into a string box and hope. This node exists to save you from the hope part.

How it works

Under the hood it's four lines of Python. When the node is created, it scans ComfyUI/input/, keeps anything ending in wav, mp3, ogg, flac, aiff, or aif, and populates a single required input:

• audio - the dropdown of audio files sitting in your input directory. That's the whole input list.

When you pick one, it joins the filename to the input directory path and returns it on a single output:

• PATH - the absolute path to the file, which you wire into Qwen2_AQA's source_audio_path.

That's it. One input, one output, no options, no hidden behavior. It's a genuine utility node in the best sense - it does exactly one thing and does it without state.

Installing it

There's no separate install for this node; it ships inside the pack, so you install the pack:

• ComfyUI Manager - search for Qwen2 (the pack title is ComfyUI_Qwen2-Audio-7B-Instruct-Int4), install, restart.
• Manual - cd ComfyUI/custom_nodes && git clone https://github.com/IuvenisSapiens/ComfyUI_Qwen2-Audio-7B-Instruct-Int4, then pip install -r requirements.txt and restart. The requirements pull in torch, transformers>=4.45.0, bitsandbytes, librosa, and huggingface_hub - heavy, and shared with the main node. If the pack runs at all, this node runs.

The gotcha that'll trip you

The file list is built when the node is created, not when you run it. Drop a new mp3 into ComfyUI/input/ while the graph is open and it won't appear in an already-placed node's dropdown. Delete the node and re-add it (or reload ComfyUI) and it shows up. Annoying, but one second to fix once you know.

Also keep the file extensions honest. The scan filters by suffix, so a .mp3 renamed to .data is invisible to this node even though the audio decoder in Qwen2_AQA would happily read it. Leave the extension alone and you're fine.

The real workflow this feeds is: AudioLoader → Qwen2_AQA with a question like "tell me what you hear," then the text answer goes wherever you need it - a caption, a prompt, a metadata field. Given the whole pack gets you one honest node that actually talks about audio, this little file picker is the part that makes it pleasant to use.