YuE2 Cover Studio · 1 plan

Covers have a control problem: "make it a bossa nova version" is a wish, not a parameter, and the model underneath - YuE2 - has no dial labelled how much of the original survives. So the toolkit invented one and did the unglamorous work of making it real. This node is step 1 of the Cover Studio: it turns a single 0–100 slider into a structured profile and writes the planning prompt for an LLM.

First thing to get straight: Interpretation Freedom is a toolkit abstraction, not a YuE2 parameter. Nothing is forwarded to the engine as a number. What happens instead is that the value selects what's preserved and what may be reworked, and the later stages realise that in operations YuE2 actually understands - a transposed key, a rewritten tempo field, a chord-free melody line, and explicit instructions.

What the slider actually buys

The bands, from the pack's own guide:

| Freedom | Band | Keeps | May change | | --- | --- | --- | --- | | 0–20 | Faithful cover | melody, structure, harmony, tempo, key, hooks | instruments, sound, vocal timbre | | 21–40 | Arrangement cover | melody, hook, structure, basic tempo | harmony, arrangement, instrumentation | | 41–60 | Reinterpretation | central melodies, hooks, rough structure | harmony, tempo, key, rhythm | | 61–80 | Creative reinterpretation | hook, key motifs, characteristic phrases | melody shape, section length, harmony | | 81–95 | Loose adaptation | selected motifs, basic dramaturgy | almost everything else | | 96–100 | Inspired recomposition | key, tempo, structure, character | compose rather than reproduce |

The weights interpolate between band boundaries, so every integer gives you a different profile. Default is 30.

Two things keep this from being decoration. Explicit settings always win: every preserve_* widget and melody_only is tri-state - auto takes the slider's value, yes pins the element, no releases it - and the degree widgets (melody_variation, rhythm_variation, harmony_freedom, structure_freedom) work the same way with none/low/moderate/high. And only real operations move: key_change transposes by ±12 semitones, tempo_change rewrites the Q: field by ±50%, and melody_only strips the quoted chord symbols.

One deliberate exception: melody_only follows the decode mode, not the slider. A full cover promises melody and harmony, so it keeps its chords at any freedom value; asking for melody_only = yes there is refused with an explanation instead of silently handing the engine a harmony score with no harmony in it.

The rest of the inputs

cover_source_json and cover_abc are required wiring - the cover source identity and the SheetSage2 score - and the node never edits that score, it passes the original through in its state.

target_style is the hint for how the material gets reworked. Empty means no hint, not "follow the source". Do not wire the song request node in here: it consumes the studio's rewritten score, so it's downstream, and the link is a cycle. That's what Style hint · template or text exists for.

cover_lyrics takes the Whisper report if you have one, so the studio can flag words that no longer fit the melodic material before a render is spent. It never changes the words.

Then the lyrics lock. lyrics_policy is auto (mode decides) - the normal case, where Song request and Whisper own the words - or keep source words / keep supplied words, which lock the Whisper transcription verbatim or the text in supplied_lyrics. Write supplied lyrics with the same section tags as the arrangement; nothing is rewritten or syllable-aligned for you. Impossible combinations are refused rather than guessed: an instrumental cover has no words to keep, and original lyrics already owns its words.

vocal_range (from auto up through Low male, Baritone, Tenor, Alto, Mezzo-soprano, Soprano) is a request to the model, not an enforced constraint - the report says as much. advanced_mode only widens the debug report; it never changes what gets applied.

The outputs: system_prompt and user_prompt go into an LLM Chat node for the plan call, and studio_json carries the profile, the analysed score, the overrides and the original ABC to step 2.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/jplenio/ComfyUI-MiniMax-Music-Production-Toolkit.git
cd ComfyUI-MiniMax-Music-Production-Toolkit
python -m pip install -r requirements.txt

Restart ComfyUI, refresh the browser, then reopen the bundled Music_Production_Toolkit.json - the studio path is optional and only exists in that graph. The whole cover chain needs SheetSage2 weights (about 1.39 GB) and a ComfyUI build with the native audio-encoder nodes; the model check node can fetch SheetSage2 when sheetsage2_models and auto_download are on.

Where people get stuck

Dependency cycle detected is the classic first-day error, and it means you wired the song request node into target_style. Use the style hint node instead.

The slider isn't doing what it says. Check whether an override is pinned: auto is the only value that lets the slider through, and the report lists which overrides were live. Also remember the freedom permissions apply to the source material only - a higher value never means less of your requested style.

A faithful cover that isn't. Freedom 0 is enforced rather than promised - a model score that rewrites a note, chord, bar or tempo gets rejected in step 3, so you really do get the source score back.

There's essentially no community footprint for this pack, so calibration comes from the project's demo gallery: 35 rendered tracks with the settings behind each one.