MiniMax Music3 Semantic Studio for ComfyUI

MiniMax Music3 Semantic Studio adds a visual music-design environment and a lightweight non-destructive audio editor to ComfyUI.

This is not an automatic prompt generator or one-click song generator. Semantic Studio is an editor for visually designing, organizing, and refining the instructions you send to MiniMax Music 3. You can build them yourself or import a prompt created with ChatGPT or another LLM.

Instead of managing a long music prompt entirely as raw text, you can visually organize:

Genre, mood, BPM, key and meter
Song structure — Intro / Verse / Chorus / Bridge / Outro
Section timing and Energy
Vocal direction
Instruments for each section
Full Lyrics and Section Lyrics
MiniMax Music 3 generation settings

After generation, the same workflow continues into the included Audio Editor, where you can cut, arrange, fade, automate levels, mix and add effects.

GitHub:
https://github.com/ukr8b3g-cmyk/MiniMax-Music3-Semantic-Studio

Installation

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

Model Links

Diffusion Model

Low VRAM / Recommended

minimax_music3_dit_int8_convrot.safetensors

FP16

minimax_music3_dit_fp16.safetensors

Text Encoder

minimax_music3_text_encoder_pruned_int8_convrot.safetensors

VAE

minimax_music3_dav.safetensors

Model Storage Location

ComfyUI/

└── models/

├── diffusion_models/

│ ├── minimax_music3_dit_int8_convrot.safetensors

│ └── minimax_music3_dit_fp16.safetensors

│

├── text_encoders/

│ └── minimax_music3_text_encoder_pruned_int8_convrot.safetensors

│

└── vae/

└── minimax_music3_dav.safetensors

You only need one diffusion model.
For lower VRAM usage, use the INT8 ConvRot version.

What is Semantic Studio?

Semantic Studio is a visual front end for designing the semantic instructions sent to MiniMax Music 3.

Think of it as a song planner for ComfyUI.

Instead of writing everything as one large text prompt, you can build the musical idea from a graphical interface and see how the song is structured before generation.

For example, you can create:

Intro → Verse → Chorus → Verse → Chorus → Bridge → Final Chorus → Outro

Then give each section its own duration, energy, vocal direction and instruments.

The timeline is not a DAW timeline and the instrument lanes are not audio stems. They describe what you want MiniMax Music 3 to generate.

This makes complex music prompts much easier to understand and adjust.

Build the Song Visually

At the top level, you can define the overall musical direction:

Genre
Subgenre / influences
BPM
Key and scale
Meter
Mood
Vocal or instrumental mode
Main vocal character
Production direction

Then you can design the song section by section.

Shape the Energy

The Energy view lets you describe how the intensity of the song should develop over time.

You might start quietly, increase the energy through the verse, hit a strong chorus, drop down for a bridge and then build toward the final section.

This is generation guidance, not an audio-volume envelope. It tells MiniMax Music 3 about the intended musical intensity before the song is generated.

Click the ▼ arrow on any preset-based field to open the available options. You can quickly choose from predefined values such as vocal type, timbre, mood, instruments, and other supported settings.

Instruments by Section

Instrument lanes make it easy to describe which instruments should be present in different parts of the song.

For example:

Intro — Piano + Pad
Verse — Piano + Bass + Light Drums
Chorus — Guitar + Bass + Full Drums + Synth
Bridge — Piano + Strings
Final Chorus — Full arrangement

Again, these are semantic instructions rather than separated stems. MiniMax Music 3 decides how to realize the requested arrangement.

Lyrics Editing

Semantic Studio also provides dedicated Lyrics views.

You can work with:

Caption — the semantic music description
Full Lyrics — the complete tagged lyrics
Section Lyrics — lyrics organized by song section

This makes it much easier to work on a complete song without constantly editing a large prompt manually.

Individual sections can also be adjusted without rebuilding the entire lyric structure.

Already Have a Prompt? Import It.

You do not have to build everything manually.

If you already created a MiniMax Music 3 prompt with ChatGPT, another LLM, or your own prompt template, use Import Prompt.

Paste your Caption / Lyrics and click:

Analyze → Import Preview → Apply Import

Semantic Studio analyzes the text locally and tries to map the detected information into the graphical project:

Global music settings
Vocal information
Song sections
Lyrics
Other detected semantic information

You can inspect the result before applying it.

This does not require an LLM connection inside ComfyUI. Prompt Import is processed locally.

Generation Controls Are Included

You can also access the important MiniMax Music 3 autoregressive generation controls from the Studio.

Available controls include:

Music Seed (AR)
Seed Behavior
Music CFG (AR)
Music Top-K
Duration
Auto Sync with Timeline

The Studio edits the actual underlying ComfyUI node values rather than maintaining a separate copy.

The Music Seed (AR) is also separate from the later KSampler seed. Both stages can therefore be controlled independently.

Generate Normally in ComfyUI

Once the song design is ready, Queue the workflow normally.

A typical generation path is:

Semantic Studio
      ↓
MiniMax Music 3 generation
      ↓
VAE Decode Audio
      ↓
Capture / Freeze Audio
      ↓
Audio Editor
      ↓
Preview / Save Audio

This is where Capture / Freeze Audio becomes important.

Capture / Freeze Audio — Edit Without Regenerating the Song

This small node solves a practical problem.

Normally, after generating a song, changing something downstream and queuing the workflow again can cause the expensive upstream generation path to run again.

Capture / Freeze Audio lets you keep the generated take in memory.

Step 1 — Generate in Capture Mode

Set:

Capture / Freeze Audio → Capture

Then Queue the workflow.

MiniMax Music 3 generates the song normally.

Capture / Freeze Audio stores a snapshot of the resulting AUDIO in CPU memory.

Music Generation
      ↓
generated AUDIO
      ↓
Capture / Freeze Audio
   [ Capture ]
      ↓
Audio Editor

Step 2 — Switch to Frozen

Once you have a take you want to edit, change:

Capture → Frozen

Now the captured audio becomes your working source.

When the workflow is queued again in Frozen mode, the node reuses the stored audio instead of requesting another Music3 generation.

This is especially useful when you want to spend time editing one good take.

Step 3 — Open Audio Editor

Click Open Audio Editor.

You can now work on the captured song while keeping the original generated audio untouched.

Important

The frozen audio is stored in CPU RAM for the current ComfyUI session.

It is not permanently saved inside the node.

If you restart ComfyUI, the frozen snapshot is cleared. Switch back to Capture and Queue a generation again to create a new snapshot.

A Lightweight Audio Editor Inside ComfyUI

The included Audio Editor is designed for practical post-generation work without turning the custom node into a full DAW.

You get one familiar waveform workspace where you can quickly clean up and finish a generated track.

Basic editing includes:

Select ranges
Cut / Copy / Paste
Split
Duplicate
Delete
Silence
Reverse
Clip mute
Fade In / Fade Out
Crossfade helper
Loop a selected range
Undo / Redo

For example, simply select the end of a song, right-click and choose Fade Out.

Volume Automation with Envelope

Switch to the Envelope tool to draw gain changes directly over time.

This is useful for:

Lowering an overly loud section
Creating gradual volume changes
Controlling an intro or outro
Making small level corrections

Unlike Semantic Studio's Energy curve, this changes the actual generated audio.

Mixer

The Mixer provides quick controls for the audio and final output.

You can adjust level, pan and final output behavior without adding extra processing nodes throughout the ComfyUI graph.

Built-in Effects

The Audio Editor also includes a built-in effects rack.

Available effects include:

Gain / Amplify
Compressor
Limiter
3-Band EQ
High-Pass Filter
Low-Pass Filter
Stereo Width
Reverb
Stereo Delay

Effects can be enabled, bypassed, reordered, reset or removed.

Reverb provides controls such as Room Size, Pre-delay, Reverberance, Damping, Tone and Wet/Dry.

Stereo Delay includes Delay Time, Feedback, Wet/Dry and optional Ping-Pong behavior.

The built-in effects work with both the immediate browser preview and the authoritative ComfyUI render.

Optional VST3 Support on Windows

If you want to go further, the editor can optionally host Windows VST3 effects.

VST3 support is completely optional. You do not need it to use Semantic Studio or the built-in Audio Editor.

If enabled, installed VST3 effects can be added from the editor.

You can even open the plug-in's native interface.

Change the plug-in parameters in its normal interface, close it, save your edits and Queue the workflow for the final render.

Third-party VST3 plug-ins are not bundled with this project.

Draft Preview vs Final Audio

While editing, Draft · Current Edits gives fast feedback without requiring a full Queue after every small adjustment.

This makes normal editing much more comfortable.

When you are satisfied:

Click Save Edits
Keep Capture / Freeze Audio in Frozen
Queue the workflow
The Python/PyTorch backend renders the edits
Preview or save the resulting AUDIO

Frozen AUDIO
     ↓
Audio Editor
     ↓
Edit / Mixer / Effects / VST3
     ↓
Save Edits
     ↓
Queue
     ↓
Authoritative Edited AUDIO

The browser preview is for fast editing feedback. The queued Python/PyTorch result is the authoritative final output.

Save & Queue

Click Save Edits, then Queue while Frozen.

The Music3 generation does not need to run again for each edit.

Who Is This For?

This custom node is especially useful if you:

Generate full songs with MiniMax Music 3
Find long music prompts difficult to manage
Want to see the song structure visually
Frequently experiment with Lyrics and section structure
Want to keep a good generated take while editing it
Want basic audio editing without leaving ComfyUI
Want built-in DSP without building a large post-processing graph
Occasionally want to use existing VST3 effects

You can use only the Semantic Studio, only the Audio Editor, or the complete workflow.

Installation

The project is available on GitHub:

MiniMax Music3 Semantic Studio
https://github.com/ukr8b3g-cmyk/MiniMax-Music3-Semantic-Studio

Manual installation:

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

Restart ComfyUI after installation.

The core package does not require additional mandatory Python dependencies.

VST3 hosting is optional and can be installed separately only if you want to use it.

Workflow

A ready-to-use example workflow will be provided with this Civitai release.

It is recommended to start with the included workflow because the intended connection is:

Music3 Generation
       ↓
VAE Decode Audio
       ↓
Capture / Freeze Audio
       ↓
Music3 Semantic Studio Audio Editor
       ↓
Preview / Save Audio

Then open Semantic Studio to design the music and use Audio Editor after generation.

Notes

Semantic Studio provides semantic guidance to MiniMax Music 3.

BPM, key, exact section timing, energy, vocal treatment and instrumentation should therefore be understood as generation targets rather than strict MIDI/DAW-style guarantees.

The Audio Editor is different: its edits and supported DSP are applied to the actual generated audio.

This separation is intentional:

Semantic Studio = design what you want Music3 to generate.

Audio Editor = edit what Music3 actually generated.

Links

GitHub / Documentation / Latest Version
https://github.com/ukr8b3g-cmyk/MiniMax-Music3-Semantic-Studio

For detailed specifications, controls, DSP behavior, VST3 information and development notes, see the full GitHub README and documentation.

Files with SHA256: d794384d2dd952cedc691acb355ea1eb3b061534e9b27716f73c38bd4010f971