🎙️ VRGDG Load Audio Split HUMO Transcribe V3

This is the splitter the pack converges on. VRGDG_LoadAudioSplit_HUMO_TranscribeV3 is the latest and most complete member of the family: it transcribes with Whisper, splits the track into scenes, builds the per-scene instruction strings, manages the output folders, and drives the queue - all in one node. If you're starting a music-video workflow from scratch in 2026, this is the splitter you want, and the earlier variants are there mainly for people with existing graphs.

It's essentially LoadAudioSplit_HUMO_Transcribe (split + lyrics) fused with LoadAudioSplit_General's chunker machinery (sets, auto-queue, instructions, output folders). That fusion is the point: you no longer need two nodes doing two jobs - this one is the front end of the whole render pipeline.

What it does

The inputs that matter:

• audio (AUDIO) - the loaded track.
• trigger (*) - advance to the next scene/set, same pattern as the General splitter.
• scene_duration_seconds (FLOAT, default 4, range 1–5) - fixed scene length. Keep this modest; the 5-second cap is deliberate so scene jobs stay small enough to be resumable.
• folder_path (STRING, default video_output) - where per-scene output lands.
• enable_auto_queue (BOOLEAN, default true) - auto-advance through the queue. Off = step manually while debugging.
• language (enum, default english) - Whisper's transcription language.
• enable_lyrics (BOOLEAN, default true) - transcription switch.
• use_context_only (BOOLEAN, default false) - when on, generate scenes from your context_* inputs alone, ignoring the transcription. Handy when the lyrics are useless (instrumental, gibberish) but you still know what you want visually.
• overlap_lyric_seconds (FLOAT, default 0) - bleed between scenes for continuous motion.
• fallback_words (STRING, default thinking,walking,sitting) - caption fallback when a scene has no clean vocal content.
• context_1 .. context_16 (STRING, optional) - per-scene direction for the LLM.

Outputs: the usual meta, total_duration, lyrics_string, audio_1 .. audio_16, plus the chunker's set of scene bookkeeping - index, start_time, end_time, instructions, total_sets, groups_in_last_set, frames_per_scene, audio_meta, output_folder - and signal_out (*) to chain the next stage. instructions is the composed per-scene direction that feeds the LLM and the video side; groups_in_last_set tells you how the final batch is split.

Wiring the front end

Audio in, signal_out to the render branch, and the loop runs: the node produces the current scene's audio and instructions, downstream generates the video, and trigger advances. use_context_only is the escape hatch that rescues runs where transcription is fighting you - sometimes you know the visual story better than the model knows the song.

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:

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

If it's not working

• The queue stalls after scene one - check enable_auto_queue. Same trap as the General splitter; it defaults on, but if you (or a template) switched it off, nothing advances.
• Scene length surprises you - scene_duration_seconds is capped at 5 here. That's by design for resumability, not a bug; long scenes belong to a different workflow shape.
• The LLM ignores the song - if transcription is junk, try use_context_only with strong context_* text instead of fighting it.
• Whisper errors on import - the README's requirements don't name Whisper explicitly; make sure it's in your environment, then restart ComfyUI.