VRGDG Load Audio Split (General)

This is the node that makes a full music video actually finish on a consumer GPU. VRGDG_LoadAudioSplit_General takes a loaded audio track, chops it into scenes of a fixed length, and - here's the important part - manages them as a queue of work batches so you're never trying to render a whole song in one shot. Instead of one giant video job that OOMs your card, you get a series of per-scene jobs with state tracked between them.

That's the "set" system you'll see scattered across this pack's outputs: the pipeline runs in sets, writes intermediate results to disk, and lets you resume. Long-form video generation in ComfyUI is a marathon, and this node is the pace car.

What it does

The inputs that matter:

• audio (AUDIO) - the loaded track (wire it from a loader like VRGDG_LoadAudioWithPath).
• trigger (*) - an any-type kick to advance to the next scene/set.
• scene_duration_seconds (FLOAT) - the fixed scene length. This is the uniform-grid splitter, so every scene is this long (unless you override below).
• fps (INT, default 24) - used to compute frame counts for the LTX side.
• folder_path (STRING, default VRGDG_Video) - where per-scene outputs go.
• enable_auto_queue (BOOLEAN, default true) - auto-advance through the sets. Turn it off to step through manually while you debug.
• override_chunk_index (INT, default -1) - start from a specific chunk instead of the beginning; your resume mechanism.
• overwrite_mode (enum overwrite/backup) - whether rerunning a scene replaces it or keeps a copy.
• use_humo_alignment (BOOLEAN, default false) - flip on when the target workflow is HUMO-aligned scene generation.
• List_of_Scene_durations (FLOAT, default 0) and manual_total_sets (INT, default 0) - escape hatches for non-uniform scene lengths and manually capping how many sets run.

Outputs: meta (DICT) and audio_meta (DICT) carry the scene state; index, total_sets, start_time, end_time, instructions, frames_for_ltx, frames_per_scene, and preroll_frames describe the current scene and how to render it; audio (AUDIO) passes the current scene's audio through; signal_out (*) chains to the next stage; output_folder and overwrite_mode report where things are landing.

How it fits the pipeline

The flow is: splitter produces the current scene's slice, downstream nodes generate the video for that slice, and the signal_out/trigger loop advances to the next scene. The instructions string is where the per-scene direction (lyrics, motion notes) gets composed for the LLM. Because every scene's metadata and audio come out as proper typed outputs, you can build one self-contained "render one scene" subgraph and let the queue call it N times.

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 - librosa is here for the audio work:

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

If it's not working

• It stops halfway - check enable_auto_queue. Off means it waits for a manual nudge on every scene. On is the setting you want for unattended runs.
• A rerun clobbered good output - switch overwrite_mode to backup. The default is overwrite, which is exactly what you want during iteration and exactly wrong when you're happy with a scene.
• It restarts from scratch - use override_chunk_index to resume from the scene that failed instead of replaying the whole queue.
• Scenes mistimed for the video model - if you're on the HUMO path, use_humo_alignment is the switch you forgot to flip.