VRGDG_LoadAudioSplitDynamic

Most audio splitters work on a grid: "every scene is N seconds." VRGDG_LoadAudioSplitDynamic is the version that lets you play director instead. You pick how many scenes you want, then set each scene's duration individually - scene 1 is 8 seconds, scene 2 is 12, scene 3 is 5 - because real songs don't have uniform verses, and you shouldn't have to pretend they do.

This is one of the pack's audio-to-scene nodes, and it's built around the idea that each scene clip feeds an audio-driven video model. The workflow family here covers two targets: HUMO (ByteDance's human-centric video model, which drives motion from the audio) and talking-head style pipelines. The using_infinite_talk dropdown exists precisely because the splitter was written with one of those in mind first - its own tooltip says "If using HUMO, change this to false." That tooltip is the author telling you which mode is which.

What it does

The inputs that matter:

• path (STRING, default ./audio.mp3) - where the audio lives. This node loads from a path rather than taking an AUDIO input, which is the key difference from the sibling splitters.
• offset_seconds (FLOAT, default 0) - skip the first N seconds before splitting. Handy for trimming intros with no vocals.
• scene_count (INT, 1–50, default 1) - how many scene clips you want out.
• using_infinite_talk (enum false/true) - set to false if you're on the HUMO workflow, per the tooltip.
• duration_1 .. duration_50 (FLOAT, default 3 each) - the per-scene lengths, in seconds. These are your creative control: set scene_count to however many scenes you have, then dial each duration to match the music.

Outputs:

• meta (DICT) - the scene metadata (counts, offsets), ready for VRGDG_Json2String if you want to hand it to an LLM.
• total_duration (FLOAT) - how long the summed scene clips come to; a sanity check that your durations roughly match the track.
• audio_1 .. audio_50 (AUDIO) - the individual scene clips. Wire each into whatever generates video from a scene.

How to use it like a human

Load the track, set scene_count, and start with the verse-chorus-verse structure you actually hear. Intro 4s, verse 8s, chorus 12s, verse 8s, chorus 12s, outro 6s - that's a six-scene split with six different durations. The defaults are all 3 seconds, which is a fine placeholder, but the whole point of this node is that you replace them with real timings. Check total_duration against the track length after you're done.

Installing it

Part of the comfyui-vrgamedevgirl pack. ComfyUI Manager: search "vrgamedev", install, restart. Or:

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

then install the README's requirements - this node is why librosa is on the list:

pip install -r custom_nodes/comfyui-vrgamedevgirl/requirements.txt

If it's not working

• Sum of durations doesn't match the track - that's on you, not the node; the last scene will just be shorter (or the split overlaps the track end). Verify against total_duration.
• Wrong target workflow - if your clips come out mistimed for the video side, re-check using_infinite_talk. The tooltip is explicit: HUMO wants false.
• Path issues - it's a raw path string, so relative paths resolve against ComfyUI's working directory. If the node can't find the file, give it an absolute path.