Extensions/MiniMax Music3 Semantic Studio

ComfyUI Extension

MiniMax Music3 Semantic Studio

Semantic Timeline, lyrics and generation authoring for MiniMax Music 3 in ComfyUI, with Capture / Freeze Audio workflow and a non-destructive Audio Editor with built-in DSP and optional Windows VST3 hosting.

By ukr8b3g-cmyk·Created about a month ago·Updated 23 days ago· 4

ukr8b3g-cmyk/MiniMax-Music3-Semantic-Studio

Nodes3

On cloudLocal install

Categoryaudio/minimax music, model/conditioning/minimax music

Stars4

Updated23 days ago

Nodes (3)

Capture / Freeze Audio

Keep the take you liked before the graph re-rolls it

audio/minimax music

Music3 Semantic Studio

How Semantic Studio compiles a song plan

model/conditioning/minimax music

Music3 Semantic Studio Audio Editor

There's a non-destructive DAW hidden in this ComfyUI node — and no, Save Edits doesn't save anything

audio/minimax music

Readme

MiniMax Music3 Semantic Studio

Music3 Semantic Studio is an external ComfyUI custom-node package for MiniMax Music 3 generation design and non-destructive post-generation audio editing.

Semantic Studio is a visual authoring/editor tool, not an automatic prompt generator or one-click song generator. It gives you a graphical workspace for designing, organizing and refining the instructions sent to MiniMax Music 3. You can build the song plan yourself or import Caption / Lyrics prepared with ChatGPT, another LLM or an external editor, then continue editing them visually in Semantic Studio.

Current status:

Semantic Studio — Timeline / Lyrics / Generation UI implemented
English / Japanese UI — ComfyUI locale-aware labels implemented; other locales fall back to English
Audio Editor — unified waveform editor, schema-2 automation, Browser Draft Preview, Edit / Mixer / Effects workspace, selection tools and non-destructive fades implemented
Built-in DSP — Gain, Compressor, Limiter, EQ / filters, Stereo Width, Reverb and Stereo Delay implemented
VST3 — optional Windows VST3 host, native plug-in UI and state capture available on demand

Neither Semantic Studio nor the Audio Editor patches ComfyUI core, MiniMax Music3 model code, KSampler, or VAE code.

Installation

cd ComfyUI/custom_nodes
git clone https://github.com/ukr8b3g-cmyk/MiniMax-Music3-Semantic-Studio.git

Restart ComfyUI after install/update. The core node package has no additional mandatory Python runtime dependency. The Windows VST3 host is optional and is installed only when a user explicitly requests it from the VST3 workspace.

Quick start — visual workflow

The normal V1 workflow is:

Import Prompt / design in Semantic Studio
                 ↓
       Queue in Capture mode
                 ↓
          generated AUDIO
                 ↓
       Capture / Freeze Audio
                 ↓
   switch to Frozen to keep the take
                 ↓
          Open Audio Editor
                 ↓
     Edit / Mixer / Effects / VST3
                 ↓
             Save Edits
                 ↓
              Queue
                 ↓
       authoritative edited AUDIO

1. Start from the Semantic Studio node

The compact graph node keeps the main generation controls close to the workflow. Music Seed (AR) controls the MiniMax Music3 autoregressive stage, Seed Behavior selects the normal ComfyUI seed behavior such as Randomize or Fixed, and Duration sets the AR generation ceiling. Import Prompt opens the structured prompt importer; Open Semantic Studio opens the full authoring interface.

The Music3 AR seed is separate from the later KSampler seed. Changing one does not replace the other.

2. Import a prompt, inspect it, then edit Lyrics

Import Prompt is intended for Caption / Lyrics text prepared in another LLM or editor. Paste the material, click Analyze, inspect the detected global settings, vocals and sections in Import Preview, then click Apply Import. The usual mode is Replace section structure; Merge detected fields is available for incremental updates.

The Lyrics workspace is split into three practical views:

Caption — the compiled semantic description sent through the Music3 text-conditioning path
Full Lyrics — the complete tagged lyrics document
Section Lyrics — lyrics grouped by Intro / Verse / Chorus / Outro and other timeline sections

Full Lyrics can be edited directly and then applied back to matching sections. Section Lyrics can also be edited independently when only one part of the song needs adjustment.

3. Shape the song structure, Energy, Instruments and AR generation

The Timeline is a semantic song plan rather than a stem editor. Click a section to edit it in the Section Inspector. Duration, section type, vocal direction and other section fields can be adjusted there. Energy can be edited numerically and also manipulated from the timeline graph; it describes the intended musical intensity for generation, not the amplitude of already-rendered audio.

The Instruments lanes are per-section semantic guidance. Turning an instrument lane off for a section means that instrument is no longer requested for that section; it does not delete an audio stem. These lanes describe what MiniMax Music3 should aim to generate.

The Generation tab edits the same underlying ComfyUI node widgets used at execution time:

Music Seed (AR) — autoregressive music/token randomness
Seed Behavior (AR) — Fixed / Randomize / Increment / Decrement behavior
Music CFG (AR) — autoregressive token guidance
Music Top-K — AR token candidate restriction
Duration Limit — maximum AR generation duration
Auto Sync with Timeline — keeps the duration ceiling aligned with the semantic timeline total

These controls are separate from KSampler Seed and KSampler CFG later in the graph.

4. Edit the generated audio

Connect the generated/decoded AUDIO to Capture / Freeze Audio, then connect its AUDIO output to Music3 Semantic Studio Audio Editor. Start in Capture mode and Queue once to generate and snapshot the take. When you want to keep that take for editing, switch to Frozen before re-queuing; Frozen reuses the captured in-memory AUDIO so the upstream Music3 generation does not run again. Then click Open Audio Editor.

The right-side workspaces have separate roles:

Edit — source range, timeline position, clip gain/pan, fades, reverse and clip mute
Mixer — input gain/pan plus output gain, channel mode and normalization
Effects — built-in non-destructive DSP rack
VST3 — optional third-party Windows VST3 effects

The built-in Effects rack includes Gain / Amplify, Compressor, Limiter, EQ (3-Band), High-Pass Filter, Low-Pass Filter, Stereo Width, Reverb and Stereo Delay. Effects can be enabled/bypassed, reset, removed and reordered.

Envelope means gain automation over time. Use the Envelope tool to add and move points on the waveform so the audio becomes louder or quieter across chosen parts of the timeline. This is post-generation audio level automation and is different from the Semantic Studio Energy guidance used before generation.

For a conventional fade workflow, drag a range on the waveform, right-click the selection, then choose Fade In or Fade Out. The selected range is split non-destructively and the fade spans exactly that selection. The Edit workspace still exposes numerical fade duration/curve controls for precise adjustment.

What happens after editing?

Audio Editor changes are non-destructive. The connected source AUDIO remains the source of truth; in the V1 template this is the AUDIO held by Capture / Freeze Audio.

Browser Draft · Current Edits gives immediate preview feedback for supported built-in edits/effects.
Save Edits stores the current edit state back into the Audio Editor node.
Keep Capture / Freeze Audio in Frozen mode and Queue the workflow again.
The Python/PyTorch backend applies the saved edit state to the frozen source AUDIO and outputs the authoritative edited AUDIO without re-running upstream Music3 generation.
A downstream Preview/Save Audio node receives that edited output.

So Save Edits does not permanently rewrite the source file. The final result is created when the Audio Editor node is queued again. The Frozen snapshot lives in process memory and is cleared when ComfyUI restarts; after a restart, use Capture and Queue once to create a new snapshot.

5. Optional VST3 effects and native plug-in UI

VST3 support is for users who already work with third-party audio plug-ins. The example above shows MuseFX Chorus and MuseFX Compress; MuseFX is only an example and is not bundled with this repository. VST3 plug-ins themselves must be installed by the user in the normal Windows VST3 locations.

The VST3 host is also optional. Automatic package installation is intentionally not exposed through a ComfyUI HTTP route. On Windows, when the Audio Editor detects that the host is missing, the VST3 workspace shows a fixed PowerShell command for the same Python environment currently running ComfyUI. Copy that command, run it in a local terminal, and restart ComfyUI. Users who never open/use VST3 do not need to install it.

VST3 discovery and native editor controls are available only when ComfyUI itself is bound to a loopback address and the UI is opened locally. They are disabled when ComfyUI listens on a LAN or public interface.

After the host reports Ready:

Click + Add VST3 and choose an installed effect.
Click Open UI to launch the plug-in's original native Windows interface.
Change the plug-in settings.
Close the native UI; the editor captures the plug-in state.
Click Save Edits and Queue the workflow to apply the VST3 processing to authoritative AUDIO.

Browser Draft does not replace the authoritative VST3 render. Queue rendering remains the source of truth for third-party VST3 processing.

Semantic Studio — generation design

Node ID: MiniMaxMusic3SemanticStudio
Display name: Music3 Semantic Studio
Category: model/conditioning/minimax music
Outputs: CONDITIONING, seconds

Load CLIP
   |
   v
Music3 Semantic Studio ---------------------> KSampler positive
   |
   +---- seconds ----> Empty MiniMax Music3 Latent Audio ----> KSampler latent_image

Load Diffusion Model -----------------------------------------> KSampler model
Conditioning Zero Out ----------------------------------------> KSampler negative

Click Open Semantic Studio to open the authoring UI. Semantic Studio and the Audio Editor open maximized by default and can be restored to the remembered normal size.

Semantic Studio is semantic: BPM, key, exact section timing, energy, vocal treatment and instrumentation are generation targets rather than strict symbolic guarantees.

Timeline / Lyrics / Generation

The main views are explicit horizontal tabs:

Timeline — song design, structure, energy, vocal style and instrument guidance
Lyrics — Caption, complete tagged Lyrics and per-section Lyrics editing
Generation — MiniMax Music3 autoregressive generation controls

The Timeline header exposes Genre, BPM, Key, Scale / Mode, Effective Key (for example D minor), Meter and Vocal / Instrumental mode. Key and Scale remain stored separately; Effective Key is display-only.

Main Vocal contains the song-wide lead/voice type, timbre, delivery, harmony and vocal-effects wording. More Settings contains title, subgenres/influences, mood/direction and production profile. Preset-backed expressive fields remain editable and searchable; imported/custom wording is not locked to the local catalog.

The complete Song Timeline is an accordion that defaults open. Timeline order remains:

Structure
Energy
Lyrics summary
Vocal Style
Instruments

Structure sections can be added after the current selection, resized and reordered by drag/drop or Inspector arrows. Instrument lanes are semantic section.instruments[] guidance rather than stems or audio analysis. Each section owns its own instrument membership.

Undo / Redo is available for structured project editing:

Ctrl/Cmd+Z             Undo
Ctrl/Cmd+Shift+Z       Redo
Ctrl/Cmd+Y             Redo

Lyrics workspace

The Lyrics workspace contains:

Caption — authoritative compiler Caption; Edit creates a temporary Draft that must pass Analyze -> Import Preview -> Apply.
Full Lyrics — editable tagged Lyrics; Apply to Sections updates matching section Lyrics while preserving semantic fields.
Section Lyrics — compact per-section accordion.

Generation controls

Generation edits the same existing ComfyUI node widgets; it does not duplicate these values into project_json:

seed -> Music Seed (AR)
max_duration -> Duration Limit
cfg_scale -> Music CFG (AR)
top_k -> Music Top-K
linked ComfyUI value-control widget -> Seed Behavior (AR)

These are MiniMax Music3 autoregressive-stage controls and are separate from KSampler controls later in the graph:

Music Seed (AR)  -> autoregressive music/token randomness
KSampler Seed    -> diffusion latent noise

Music CFG (AR)   -> autoregressive token guidance
KSampler CFG     -> diffusion guidance

Neither pair overrides the other; the stages are separate and both are used.

Generation shows Timeline Total beside Duration Limit. Auto Sync with Timeline defaults on and keeps the AR duration ceiling synchronized with the semantic section total. It can be disabled for an independent manual ceiling. MiniMax Music3 may end the song earlier than the ceiling.

See docs/PHASE_A_SEMANTIC_UI.md.

Prompt Import

External LLM output can be pasted into Import Prompt and processed locally:

Import Prompt
   -> Analyze
   -> Import Preview
   -> Replace / Merge
   -> Semantic Studio fields

The normal external-import default is Replace section structure. Merge detected fields remains available for incremental edits. Prompt Import is deterministic and does not require an LLM connection at runtime.

See docs/PROMPT_IMPORT.md.

Audio Editor — unified non-destructive editing

Node ID: MiniMaxMusic3SemanticStudioAudioEditor
Display name: Music3 Semantic Studio Audio Editor
Category: audio/minimax music
Input: audio: AUDIO
Output: AUDIO

The public V1.0 Audio Editor uses one connected AUDIO input.

In the V1 template, place Capture / Freeze Audio between audio decode and the Audio Editor:

KSampler
   |
   v
VAE Decode Audio
   |
   v
Capture / Freeze Audio
   |
   v
Music3 Semantic Studio Audio Editor
   |