VRGDG_LoadAudioSplitUpload

The music-video pattern is scene-by-scene: you slice a song into chunks, generate a video per chunk, then stitch. VRGDG_LoadAudioSplitUpload is the slicing node - it takes an AUDIO tensor, splits it into up to 50 chunks, and hands you each chunk plus a meta dict describing the split. If you've used the pack's older splitter, this is the upload-friendly version: the audio comes in through the audio input rather than from a file path on disk, which means it plays nice with any audio source in your graph.

The mechanism is simple and honest: you tell it how many scenes (scene_count) and how long each chunk is (duration_1 through duration_N), and it cuts the waveform at those boundaries. It also has an infinite-talk mode (via using_infinite_talk) that pads chunks to a fixed internal length - useful for the pack's talking-head workflows where every scene must be exactly the same duration. The meta output and total_duration let downstream nodes (like VRGDG_CombinevideosV3) know how long each scene is, which is how the final stitch stays in sync with the music.

The inputs that matter

• audio - your AUDIO tensor.
• scene_count - how many chunks (1 to 50). This also drives the node's dynamic outputs: it shows exactly audio_1..audio_N.
• offset_seconds - skip this many seconds into the audio before the first cut (0 default).
• duration_1..duration_N - per-scene durations, default 3.88 seconds each. You set as many as you have scenes; each unused one just stays at default. For a fixed-timing talking-head pipeline, leave them equal. For music, set them to match the actual scene lengths (which you can get from your transcription/lyrics timestamps).
• using_infinite_talk - "false" or "true". True pads chunks so every scene is the same length.

Outputs: meta (DICT with the durations), total_duration (FLOAT), and audio_1..audio_N (AUDIO) - one per scene, ready to feed per-scene video generation.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/vrgamegirl19/comfyui-vrgamedevgirl

Or search vrgamedev in ComfyUI Manager. No extra dependencies beyond torch/librosa, which the pack requirements cover.

Common issues

• Chunks cut in the middle of words: the split is duration-based, not silence-based. If you need clean phrase boundaries, transcribe first and set each duration_N to match the lyric timing.
• audio_N outputs don't match scene_count: remember the outputs are dynamic - they adapt to scene_count on the node, but some graphs don't re-render the node's ports until you reconnect them. If you change scene_count, drag a new connection.
• Everything is 3.88 seconds even though your song isn't: that's the default. You have to set per-scene durations - the node can't guess your song's timing.

Honest take: it's a utility, not a marvel - you're still doing the timing work yourself. But it's the correct building block for the scene-by-scene music-video pattern, and the meta output is what makes the downstream combine node able to keep everything in sync. In the pack's ecosystem, it's the start of every music-video graph.