Vibe Music Engine

This is the node the pack is named after, and it's genuinely clever. Vibe Music Engine takes an AUDIO input - a song, a voiceover, anything with a vocal - and turns it into a complete plan for a lyric/music video: the transcribed lyrics, frame-accurate per-frame text you can use to prompt image generation, shot boundary timings, and written-out EDL and SRT files. It's Whisper-based speech-to-text bolted onto a music-video production pipeline, and it's the most interesting thing in lazniak's suite by a wide margin.

Under the hood: it dumps your audio to a temp WAV, runs OpenAI's Whisper with word-level timestamps, groups the words into lyric lines (words_per_line, default 3), converts timestamps to frames at your fps, enforces shot lengths between min_shot_length_frames (default 25) and max_shot_length_frames (default 125) so cuts feel deliberate instead of frantic, and writes the results to ComfyUI/output/<prefix>/ as an EDL (CMX3600 format, readable by NLEs) and SRT subtitles. There's also optional numba-accelerated beat detection that can align your cuts to the music. That's a lot of pipeline for one node.

Inputs worth caring about

• audio (AUDIO) - wire this from any node that outputs ComfyUI's AUDIO type. The tooltip's advice is the actual requirement: clear vocal content transcribes dramatically better.
• choose_model - Whisper size, default base. This is where the whole thing can bite you: the model downloads on first run, and it's not tiny. base ≈ 142 MB, medium ≈ 1.5 GB, large-v3 ≈ 3 GB+. The .en variants are faster if you only work in English. It loads for the run and frees VRAM after, so it won't sit in memory, but budget the download and the run time.
• source_language (default auto) / target_language (default source) - auto-detect, or force a language; set a different target to get a translated transcription.
• fps (default 25) - the master clock. Every frame number and EDL timecode derives from this, so match it to your video workflow or nothing lines up.
• words_per_line, min_shot_length_frames, max_shot_length_frames - the editorial controls. Short lines = more cuts; min shot length stops seizure-inducing staccato.
• filename_prefix (default sc) and files_in_edl (default mp4) - name the output clips and tell the EDL what files to reference.
• beat_detection, beat_sensitivity, beat_brutality - the music-sync layer. Sensitivity finds beats (higher = more, plus false positives), brutality is how hard cuts snap to them.
• save_edl / save_srt (overwrite/no/iterate) and srt_mode (raw/clean/word/original/all) - raw keeps system markers, clean is dialogue only, word is karaoke-style word-by-word timing.

What comes out

Eleven outputs. The ones you'll actually use: Lyrics (the full transcription as a string), text_lines plus frame_numbers/start_frame_numbers/end_frame_numbers (lists that let you drive per-shot image generation - feed each line's frame range into an image-gen pass with the lyric as its prompt), Frame_long (total frames), num_of_elements (how many shots/lines), and the Start_ms/Duration_ms/durations lists for finer timing work. frame_based_text is the assembled per-frame text stream.

Install & first-run reality

Same pack: Manager search "videoclipgenerator", or

cd ComfyUI/custom_nodes
git clone https://github.com/lazniak/videoclipgenerator
cd videoclipgenerator
pip install -r requirements.txt   # pulls openai-whisper + numba

Restart. First real run downloads the Whisper model - give it time and bandwidth. Troubleshooting-wise: missing whisper or numba imports mean the requirements didn't fully install (re-run the pip install against the ComfyUI Python, not your system one); a garbled transcription usually means the audio's too quiet or buried in music - the tooltip's "clear vocal content" isn't a suggestion; and the buy_me_a_coffee checkbox opens a browser tab, as it does everywhere in this pack. Expect the README's grander claims (99-language dialect handling, cross-lingual pivot translation) to be flattered marketing - the real code is Whisper + careful bookkeeping, and that's plenty.