Load AudioX Audio

Straight to the point: in the version of this pack that's currently shipping, Load AudioX Audio ignores the file you give it and always outputs silence. The node's load_audio method reads your audio_path input and then immediately overwrites it with audio_path = None, which makes the underlying loader return a tensor of zeros at your chosen sample rate and length. The input field is a decoy. This isn't a subtle incompatibility - it's a hardcoded line in nodes.py.

Which is a shame, because the node is supposed to be the audio-condition half of AudioX. The model behind this pack takes text, video, and audio conditions simultaneously (it's HKUST's "anything-to-audio" diffusion transformer), and one of its headline tricks is audio-conditioned generation - style transfer, re-dubbing, audio-to-audio. This node is how that condition enters the graph, and right now it can't.

What it does when it works (and what to change)

The design is straightforward. It loads a file with torchaudio.load, crops from seconds_start for seconds_total seconds, pads if the file is short, and resamples implicitly by slicing at your target sample rate. The output is a TENSOR you feed into Condition's audio_tensor input.

Inputs:

• audio_path (STRING) - the file to load. Broken at the moment, see below.
• sample_rate (SampleRate) - from LoadAudioXModel's sample_rate output (44100 for the stock model).
• seconds_start (INT, default 0) and seconds_total (INT, default 10) - the crop window.

The fix is one line. Open ComfyUI/custom_nodes/ComfyUI-AudioX/nodes.py, find LoadAudioXAudio.load_audio, and delete the audio_path = None line. Restart ComfyUI and the node loads real audio. If you're not comfortable editing source, you can still use the pack - the chain just treats the audio condition as silence, which is not the same as no condition at all and will quietly bias your output.

Installing AudioX

Install via ComfyUI Manager (search "ComfyUI-AudioX") or cd ComfyUI/custom_nodes && git clone https://github.com/Yuan-ManX/ComfyUI-AudioX.git, then pip install -r requirements.txt and conda install -c conda-forge ffmpeg libsndfile, then restart. The dependency list is heavy - the pack vendors AudioX's full stack (pinned pandas==2.0.2, descript-audio-codec, laion-clap, wandb) - so budget for a real install, and give a managed ComfyUI environment its own venv if pip fights back. The model downloads itself from Hugging Face on first LoadAudioXModel run.

Why this matters in the workflow

Condition requires a video tensor, a text prompt, and an audio tensor - all three, no skips. Because this loader is broken, every stock workflow feeds the model a silent audio condition. For video-to-audio (the use case most people are here for) that's tolerable; the model leans on the video and text. But if you came for audio-to-audio work - say, restyling an existing track - you need the one-line patch before anything you do will make sense. Worth knowing before you sink an hour into tuning steps and cfg_scale on a pipeline whose audio input has been zeros the whole time.