Save 4 hours! We auto-setup your workflow! Free!

Drop your workflow.json — we handle every dependency, custom node, and model. Just open the link and run.

Auto-Setup Workflow Json (Free) Now!
ComfyUI > Nodes > ComfyUI-MiniMax-Music-Production-Toolkit

ComfyUI Extension: ComfyUI-MiniMax-Music-Production-Toolkit

Repo Name

ComfyUI-MiniMax-Music-Production-Toolkit

Author
jplenio (Account age: 2279 days)
Nodes
View all nodes(1)
Latest Updated
2026-11-10
Github Stars
0.09K

How to Install ComfyUI-MiniMax-Music-Production-Toolkit

Install this extension via the ComfyUI Manager by searching for ComfyUI-MiniMax-Music-Production-Toolkit
  • 1. Click the Manager button in the main menu
  • 2. Select Custom Nodes Manager button
  • 3. Enter ComfyUI-MiniMax-Music-Production-Toolkit in the search bar
After installation, click the Restart button to restart ComfyUI. Then, manually refresh your browser to clear the cache and access the updated list of nodes.

Visit ComfyUI Online for ready-to-use ComfyUI environment

  • Free trial available
  • 16GB VRAM to 80GB VRAM GPU machines
  • 400+ preloaded models/nodes
  • Freedom to upload custom models/nodes
  • 200+ ready-to-run workflows
  • 100% private workspace with up to 200GB storage
  • Dedicated Support

Run ComfyUI Online

ComfyUI-MiniMax-Music-Production-Toolkit Description

ComfyUI-MiniMax-Music-Production-Toolkit generates songs and covers with YuE2 and MiniMax Music 3, combining lyrics, audio restoration, mastering, artwork, and tagged exports in one ComfyUI workflow.

ComfyUI-MiniMax-Music-Production-Toolkit Introduction

Music Production Toolkit for ComfyUI — YuE2, YuE2 Cover and MiniMax Music 3

ComfyUI-MiniMax-Music-Production-Toolkit is a connected workspace for creating, reinterpreting, enhancing, and exporting music in ComfyUI. The current documented release is 3.1.3.

You can start with a written idea, an existing song, or a recording that needs a better final sound. The toolkit brings together music generation, lyrics, cover arrangements, optional artwork, audio restoration, mastering, and release-ready exports.

For AI artists, this means fewer separate tools and less manual work between stages. You can:

  • Create a song from a description. Describe a mood, genre, voice, story, or arrangement, and use a language model to develop the musical brief and lyrics.
  • Make a new version of an existing song. Extract its musical structure and use YuE2 Cover to perform it in a different style.
  • Control how closely a cover follows the original. Start with one Interpretation Freedom slider, then use advanced controls if you need more direction.
  • Protect your lyrics. Keep transcribed source words or supply your own locked lyrics.
  • Enhance an existing recording. Repair certain audio problems, adjust tonal balance, and prepare a final master without generating a new song.
  • Generate matching artwork. Use an optional FLUX.2 branch to create a cover image from the song’s artwork prompt.
  • Export an organized release package. Save audio, artwork, tags, prompts, and a production record together.

The music-generation and audio-processing workflow can run locally on your own machine. The text-writing stage can also run locally, or you can choose an external cloud language model. Cloud mode sends prompts to the selected provider and may incur API charges.

Hear examples first

The demo gallery contains 35 generated tracks, including vocal songs, instrumentals, and covers made with YuE2, YuE2 Cover, and MiniMax Music 3. Search, filters, and the settings behind each example help you connect what you hear with how it was made.

How ComfyUI-MiniMax-Music-Production-Toolkit Works

Think of the toolkit as a small production team connected by a ComfyUI workflow:

StageRoleWhat it does
WRITESongwriter and plannerTurns your idea into a structured music description, lyrics, title, and artwork prompt.
GENERATEPerformerUses YuE2 or MiniMax Music 3 to produce the audio.
REFINERestoration assistantOptionally repairs or reshapes parts of the recording.
ILLUSTRATECover artistOptionally generates matching artwork with FLUX.2.
MASTERINGFinal sound engineerAdjusts tone, dynamics, loudness, and peaks.
DELIVERRelease organizerSaves audio, metadata, artwork, and production reports.

Each stage has its own purpose. For example, changing your lyrics affects the song request, while changing the final loudness target affects the master—not the composition.

From an idea to a new song

  1. You describe the music or choose a prompt template.
  2. A language model develops the request into a more complete arrangement and lyrics.
  3. YuE2 or MiniMax Music 3 generates the song.
  4. Optional processing improves or changes the sound.
  5. The workflow exports the result and records the settings used.

You can also disable the language model and enter the caption, lyrics, title, and artwork prompt manually.

From an existing recording to a cover

A cover uses an additional preparation process:

  1. Read the music. SheetSage2 transcribes the recording into a score containing melody, chords, phrases, and section boundaries.
  2. Plan the changes. Cover Studio determines what should stay and what may change, using Interpretation Freedom and any advanced settings.
  3. Rewrite the score. The toolkit works with ABC, a text-based form of musical notation. You do not need to write this notation yourself.
  4. Validate the result. Model-generated notation is checked against the supported format and arrangement plan. If it fails, the toolkit uses a validated deterministic result instead.
  5. Perform the cover. YuE2 renders the prepared score in your requested style.
  6. Optionally check for vocals. For an instrumental, the toolkit can transcribe the generated audio and retry if it detects too many words.

This is closer to giving a new performer a score than applying an audio effect to the original recording. It does not simply preserve a loop and change the surrounding sound.

Important: A valid score guides generation, but it cannot guarantee an exact audio performance. Listen to the finished result, especially when lyric accuracy or the absence of vocals matters.

ComfyUI-MiniMax-Music-Production-Toolkit Workflows

Two bundled workflows cover the main creative tasks.

Main production workflow

Music_Production_Toolkit.json supports:

  • New songs with YuE2 or MiniMax Music 3.
  • Covers with YuE2 Cover.
  • Optional Cover Studio controls.
  • Artwork generation.
  • Restoration, refinement, and mastering.
  • Metadata, exports, and production records.

Open the JSON in ComfyUI or drag it onto the canvas. The grouped notes explain where to begin.

Main workflow with labelled production groups

The main workflow is organized as CHOOSE → START → WRITE → GENERATE → REFINE → ILLUSTRATE → MASTERING → DELIVER.

Create a new song

  1. In CHOOSE, select YuE2 or MiniMax Music 3.
  2. Set your output folder, artist, and album.
  3. Choose a template or describe your idea.
  4. Adjust the genre, tempo, language, voice, and requested length as needed.
  5. Review model choices and memory requirements.
  6. Queue the workflow.
  7. Listen before deciding which refinement or mastering stages to keep.

For example, a brief such as “a restrained electronic ballad with a soft lead vocal, sparse percussion, and a gradually expanding chorus” gives the text model clearer direction than a broad request such as “make a good song.”

Create a cover

  1. Select YuE2 Cover in CHOOSE.
  2. Select a recording in Cover song · source audio.
  3. Choose full conditioning for melody and harmony, or melody conditioning for more freedom in the accompaniment.
  4. Describe the new style in WRITE.
  5. Choose the cover lyrics mode.
  6. Leave the optional controls on automatic settings, or adjust Interpretation Freedom.
  7. Queue the workflow and review the result.

The source score is supplied to the language model automatically. You do not need to prepare sheet music.

Cover titles come from the source filename. For example, My Song.wav becomes My Song-cover throughout the export pipeline. Manual or language-model title suggestions do not replace this cover title.

Audio-enhancement workflow

Music_Production_AudioEnhance.json is for recordings you already have.

It provides the restoration and mastering stages of the main workflow without requiring new music or artwork generation. It also copies the source file’s tags and embedded cover art onto both exports.

Audio-enhancement workflow

Use this workflow when you want to:

  • Improve an existing recording.
  • Compare EQ or mastering settings without regenerating music.
  • Preserve existing release metadata and artwork.
  • Focus on audio finishing rather than composition.

For a fuller walkthrough, see the complete workflow guide.

ComfyUI-MiniMax-Music-Production-Toolkit Features

Prompt templates and creative direction

The prompt library provides starting points for genres and musical styles. Templates can fill in controls, but you can edit them and save your own versions.

Use the free-description field for details such as:

  • Atmosphere and emotional direction.
  • Lead and supporting instruments.
  • Vocal character and language.
  • Story or lyrical theme.
  • Arrangement development.
  • Production style.

Choose custom for fields you want to leave unspecified. Separate system prompts control how the language model develops your brief.

If you prefer complete manual control, disable the language model and supply the parser fields yourself.

Explore the prompt library.

Interpretation Freedom for covers

Interpretation Freedom runs from 0 to 100. It determines how strongly the cover should preserve the source composition.

SettingIntended behavior
0A faithful new performance using the source score. Changes to notes, chords, bars, or tempo in a model response are rejected.
Between 0 and 100Progressively more room to reinterpret the composition within the generated plan and your explicit settings.
100A freer recomposition using the source as a reference, with the default profile retaining key, tempo, structure, and character.

This slider is a toolkit control, not a native YuE2 parameter. The toolkit translates it into a concrete profile and score operations rather than forwarding the number to the music engine.

Two constraints remain important:

  • Your requested style and lyrics keep priority. More freedom does not mean ignoring your template.
  • A full cover retains harmony at every freedom level. Choose melody conditioning when you want more freedom in the accompaniment.

Advanced cover arrangement controls

Advanced mode lets you decide element by element what should remain recognizable:

  • Main melody.
  • Chorus hook.
  • Song structure.
  • Harmony.
  • Tempo.
  • Key.

You can also control variation in melody, rhythm, harmony, and structure; specify a key shift in semitones; change tempo; and set the vocal range.

For example, you can keep the chorus hook recognizable while allowing a new rhythmic treatment, or transpose the song for a different singer.

Explicit settings override the slider-derived profile. See YuE2 Cover Studio.

Cover lyrics modes and lyric locking

The Cover lyrics setting offers three approaches:

ModeWhat happensWhen to use it
InstrumentalVocal notes become rests, and their melody moves into the instrumental part.You want instruments to carry the tune instead of a singer.
New lyricsWhisper supplies timed source phrases, and the language model writes new words with guidance from the phrasing.You want a new lyrical theme while retaining useful musical phrasing.
Original lyricsWhisper transcribes the source, and the toolkit restores those words into measured score sections.You want to keep the words recognized from the recording.

For new lyrics, approximate syllable density, stress, and breathing points guide the rewrite. Score note counts are guidance, not exact sung-syllable counts.

For original lyrics, the parser checks every word in order, including repeated words. It repairs language-model rewrites automatically and stops if the final lossless-word check fails.

Lyrics policy adds control over whether the mode decides, the source transcription is locked, or your supplied lyrics are locked. The freedom slider does not rewrite locked words.

However, preserving a transcription exactly is not the same as proving that transcription is correct. Whisper can mishear or invent lyrics, so inspect it before relying on it.

See cover lyrics modes and the cover review and correction plan.

Lead instrument for instrumental covers

When you choose an instrumental cover, the former vocal melody is assigned to the lead instrument you name.

For example, choosing a piano lead requests that the former vocal tune be played instrumentally rather than sung. This mode overrides vocal requests in the selected template.

Where the transferred melody overlaps existing instrumental material in the relevant score blocks, that material is replaced and the change is reported.

Optional instrumental vocal check

An instrumental score can still produce voice-like sounds or words. The optional vocal check evaluates the rendered audio rather than trusting the request alone.

It:

  1. Transcribes the generated take.
  2. Counts recognized words.
  3. Compares the count with your word tolerance.
  4. Generates another take only when needed, within your retry limit.
  5. Keeps the take with the fewest recognized words if none meets the tolerance.

A clean first take needs only one generation. Transcription still adds processing time, so the check is opt-in.

This is a practical screening tool—not a guarantee that every voice-like sound will be detected correctly.

Flexible language-model connections

The language-model node offers three modes:

ModeWhere text is generatedPractical consideration
In ComfyUI (GGUF)Inside ComfyUI using a compatible local modelUses your computer’s memory.
Local app / serverIn another local application or serverThat application manages its own model memory.
Cloud serviceAt an external providerSends prompts externally and may incur charges.

Supported local options include LM Studio, Ollama, llama.cpp, Unsloth Studio, and vLLM. Cloud options include OpenAI, Claude, Gemini, DeepSeek, Qwen, MiniMax, OpenRouter, and Groq, with manual entry for other OpenAI-compatible endpoints.

Helpful controls include:

  • Set API key: keeps the secret out of your workflow.
  • Find models: helps select an available server model.
  • LLM settings · central: shares provider, model, and sampler settings across the main workflow’s three LLM calls.

Each queued execution generates fresh text. Cloud requests may therefore incur charges each time.

For YuE2 Cover, cloud prompts include the source filename and ABC transcription. ComfyUI’s LLM unload node cannot unload a model managed by another application.

See the language-model connection guide.

Optional artwork

The FLUX.2 artwork branch creates a cover image from the song’s image prompt.

In CHOOSE, the control named Cover means cover artwork and defaults to ON. It is separate from the YuE2 Cover song mode.

  • Turn Cover OFF to skip artwork generation and FLUX downloads.
  • Lower artwork resolution if memory is limited.
  • Keep it enabled when you want an integrated visual identity for the release.

See the artwork workflow guide.

Audio restoration and refinement

The audio chain includes tools for:

  • De-clicking and de-clipping.
  • Band-limited restoration.
  • FlashSR-based refinement.
  • High-frequency blending and repair.
  • Optional artifact reduction.
  • EQ, sample-rate conversion, and mastering.

In version 3.1.3, the restoration chain is disabled by default in both example workflows. The PRE low-pass is bypassed, and the crossover starts at Original SRC only. When enabling restoration, review its helper stages as well.

FlashSR requires about 2.3 GB of weights on first use and can add minutes per song. If its weights cannot be downloaded, the stage logs the issue and passes the audio through rather than ending the run.

Restoration is not automatically an improvement. Compare processed and unprocessed versions at similar listening loudness.

Experimental audio artifact reduction

The main workflow includes experimental processing for brief whistles and metallic spectral spikes.

  • Controlled by artifact_reduction_enabled in CHOOSE.
  • Enabled by default with Balanced sensitivity.
  • Runs independently between Refinement and Mastering.
  • Includes analysis-only mode and a removed-audio output for evaluation.

Listen to what is removed: the processor cannot always distinguish unwanted artifacts from intentional musical content.

See artifact-reduction usage and limitations.

Mastering and EQ

Mastering prepares the final tonal balance, dynamics, loudness, and peak level.

Auto-EQ adjusts the broad tonal balance. The main workflow uses Warm - gentle by default when Mastering runs:

  • Warm tilt.
  • 35% strength.
  • Maximum 2 dB adjustment.
  • Four bands.
  • 40–16000 Hz range.

No reference track is needed for this preset. Reference presets can use a connected recording for tonal guidance; without that recording, Reference mode warns and skips correction.

The manual 8-band EQ remains editable whether Auto-EQ is enabled or not. Its own bypass control disables only manual EQ.

Available choices include 10 Auto-EQ presets, 24 manual EQ presets, and Custom settings. Manual EQ starts Flat. For harsh YuE2 high frequencies, try YuE2 - Smooth highs or its stronger variant.

The compressor begins with a gentle 1.5:1 ratio, and the master targets −14 LUFS / −1 dBTP:

  • LUFS describes perceived loudness.
  • dBTP describes the true-peak ceiling used to control signal peaks.

Turning compressor_enabled off retains loudness targeting and limiting without compression. Full mastering bypass disables dynamics and loudness processing.

The output sample rate defaults to 44.1 kHz. For 48 kHz, use the Output rate node and leave the master’s rate at keep.

The main workflow’s central Mastering switch skips the entire mastering area, including sample-rate conversion. This differs from bypassing individual processors within that area.

Peak and gain-reduction safety limits take priority over an unattainable loudness target. Check the report when the result does not reach the requested level.

Resources:

Exports and production records

The main production workflow saves:

  • Source FLAC.
  • Mastered FLAC.
  • MP3.
  • Cover JPG when artwork is generated.
  • Standard audio tags, including title, artist, album, track, genre, and embedded artwork.
  • A prompt report.
  • A central production JSON recording the run.

Files follow the Album - Title naming convention. Run records include date, time, and source file, helping you track which inputs and settings produced a result.

ComfyUI-MiniMax-Music-Production-Toolkit Models

Different models handle different jobs. You do not need every model for every workflow.

Music-generation models

Model or modePurposeWhen to choose it
YuE2Generates new songs from prepared text and lyrics.You want an original song using the YuE2 path.
YuE2 CoverUses YuE2 to perform a prepared score derived from source audio.You want a new version of an existing track.
MiniMax Music 3Generates new songs from the production brief.You want an alternative music-generation engine.

YuE2 Cover is a workflow mode using YuE2, not a separate text-writing model.

Documented weight options include:

  • YuE2: yue2_3b_int8 at about 3.7 GiB, and a bf16 option at about 7.3 GiB.
  • MiniMax Music 3: minimax_music3_dit_int8, minimax_music3_dit_fp16, and an optional fp32 DiT variant. Its pruned int8 text encoder is about 8.6 GiB, making memory planning important even with a smaller DiT.

Lower-precision variants reduce memory requirements. Their practical speed and output differences depend on the backend and hardware; the supplied context does not provide measured comparisons.

Score, transcription, artwork, and refinement models

ModelJobImportant detail
SheetSage2Extracts a musical score from audio.Used for YuE2 Cover; the model checker can download it when that mode is selected.
Whisper large-v3Transcribes lyrics and supports optional vocal checking.About 2.9 GiB; recommended in the supplied guidance for sung, mixed, and multilingual material.
Whisper large-v3-turboFaster, smaller transcription option.About 1.5 GiB, with a possible quality trade-off on dense mixes.
Whisper large-v3-turbo-int8Lower-memory transcription option.About 0.8 GiB; its quantization quality cost has not been measured by the project.
FLUX.2 kleinGenerates artwork.Documented options include bf16 or a lower-memory fp8 diffusion / fp4 encoder combination.
FlashSROptional audio refinement and high-frequency restoration.Downloads needed weights when the stage runs; failure to fetch them skips the stage.

In the bundled cover workflow, Whisper runs for new lyrics and original lyrics, not automatically for instrumental covers or ordinary song generation. The optional instrumental vocal check also uses transcription.

Local language models

The default integrated model is:

Qwen_Qwen3.5-9B-Q4_K_M.gguf — approximately 6.2 GiB.

The documented catalog and hardware guidance also mention:

Model family or variantIntended role
Qwen3.5-2B / Qwen3.5-4BLower-memory text options; complex cover rewriting may require more supervision.
Qwen3.5-9B-Q4_K_MDefault middle-sized option for structured text tasks.
Qwen3.8-9BAnother middle-sized option in the hardware guidance.
Qwen3.8-27B: UD-IQ3_XXS, UD-IQ4_XS, UD-Q4_K_MLarger options for quality comparisons, with increasing memory requirements across these variants.
Gemma 4 E4BSmaller alternative listed at about 4.8 GiB.
Gemma 4 12B IT QATMid-sized option listed at about 7.0 GiB.
Gemma 4 26B QAT / 26B-A4B UD-Q5_K_MLarger mixture-of-experts options for machines with more memory.
LFM2.5 lineSpeed-oriented options, including an 8B mixture-of-experts entry and smaller dense models.
LFM2.5-2.6BCompact option listed at about 1.6 GiB, including CPU-oriented use.

A smaller model saves memory, but the cover task is demanding: it may need to read a score, understand a plan, and return a complete valid rewrite. A model that handles ordinary song descriptions adequately may still struggle with covers.

Choosing models for your computer

Use the Model advisor node to see detected memory, per-task recommendations, and the memory margin assumed.

The language-model dropdown marks:

  • ✔: a file already present in models/llm.
  • ⬇: a catalog model available for first-use download, with size and rating.

The stored model value remains the plain filename, so saved workflows remain compatible after download.

Practical ways to reduce memory use include:

  • Choose a smaller language model.
  • Use an external local server or cloud model for text.
  • Lower artwork resolution or turn artwork off.
  • Use the enhancement-only workflow when you do not need generation.
  • Review context and output budgets carefully rather than reducing them indiscriminately.

Model file size is not a guarantee that the complete workflow will fit. Other models, context buffers, audio buffers, and backend requirements also use memory.

The author’s hardware table is guidance, not a benchmark. The project’s measurement matrix is still marked untested, so generation-time estimates should not be treated as measured performance promises.

What’s New with ComfyUI-MiniMax-Music-Production-Toolkit

Version 3.1.3

The documented release adds clearer model selection and less expensive default processing.

  • Installed and downloadable models are visibly distinguished. The dropdown shows status, download size, and rating, helping you avoid unexpected downloads.
  • The model slot summarizes availability. For example, it can show how many entries are installed and how many remain available to download.
  • Six additional chat models expand the catalog. These include Gemma 4 26B, Gemma 4 E4B, and the speed-oriented LFM2.5 line.
  • The default LLM is now Qwen3.5 9B Q4_K_M. It is approximately 6.2 GiB and downloads on first use when absent.
  • Restoration starts disabled in both example workflows. This avoids automatically paying FlashSR’s download and processing cost.
  • FlashSR no longer has a separate auto_download switch. An enabled stage fetches its requirements; a failed fetch skips the stage instead of stopping production.
  • Catalog chat-model downloads were fixed. This addresses the entry has no target directory error.
  • LLM checking is enabled again by default in the model-check node.

See the 3.1.3 release notes.

Other highlighted capabilities in the current documentation

The supplied README also highlights:

  • Cover Studio with Interpretation Freedom and validated score rewriting.
  • Three cover lyrics modes and locked-lyrics handling.
  • Optional instrumental vocal checking with best-take selection.
  • Consolidation into two workflows.
  • Run records with date, time, and source file.
  • Enhancement exports that preserve source tags and embedded artwork.

The supplied context does not assign each of these changes to an individual earlier release. For version-specific history, consult:

3.1.2 · 3.1.1 · 3.1.0 · 3.0.1 · 3.0.0 · 2.x · 1.0.x · Changelog

Troubleshooting ComfyUI-MiniMax-Music-Production-Toolkit

The workflow runs out of memory

  1. Check the Model advisor recommendations.
  2. Turn artwork Cover OFF in CHOOSE.
  3. Select a smaller local language model or move text generation to another server or a cloud provider.
  4. Review generation-model variants and offloading requirements.
  5. Reduce text budgets only if they still leave enough room for the task.

For covers, insufficient output space can truncate the rewritten score. A budget suitable for a short caption may be too small for notation.

A cover rewrite fails or falls back to a simpler result

  1. Inspect the report for notation-validation or incomplete-output messages.
  2. Use a short, concrete style description.
  3. Check that the context and output budgets can hold the score.
  4. Try advanced settings that simplify the requested changes.
  5. If failures continue, try a stronger language model or cloud text provider.

The fallback protects the music engine from unchecked notation. It does not guarantee that a weak model can deliver the creative rewrite you requested.

An instrumental still contains words or vocal sounds

  1. Confirm that Cover lyrics is set to Instrumental.
  2. Choose the intended lead instrument.
  3. Enable the optional instrumental vocal check.
  4. Set a word tolerance and retry cap.
  5. Review the recognized words and listen to the retained take.

If no take meets the tolerance, the workflow keeps the one with the fewest recognized words—not necessarily a completely vocal-free result.

Original lyrics are incorrect or incomplete

  1. Inspect the Whisper transcript before trusting the lyric lock.
  2. Check for missing phrases, incorrect repetitions, or words that do not match the recording.
  3. Use the supplied-lyrics policy when you need to lock your own text.
  4. Review the cover correction plan.

The workflow can preserve a transcription exactly, but it cannot make an incorrect transcription become accurate. Obvious opening-only fragments can stop the run before it generates an almost wordless cover.

Model downloads are unexpected or fail

  1. Check whether the dropdown shows ✔ or ⬇.
  2. Review missing files in the model-check node.
  3. Confirm whether the model requires accepted license terms or authentication.
  4. For entry has no target directory, consult the 3.1.3 fix and release notes.

FlashSR is an exception to a hard failure: if its required weights cannot be fetched, the audio continues unchanged through that stage.

Audio export fails with a blank AssertionError

  1. Review the decoding and export messages.
  2. Check whether decoding attempted its smaller-tile retry.
  3. Consult the specific audio-export troubleshooting entry.

The decoder can retry once with smaller tiles without regenerating the music. Audio savers reject invalid samples rather than deliberately writing a broken file.

The sound becomes harsh or overprocessed

  1. Compare against a bypassed version at similar loudness.
  2. Disable restoration and artifact reduction separately to identify the source.
  3. For harsh YuE2 highs, try YuE2 - Smooth highs.
  4. Reduce Auto-EQ strength or bypass it.
  5. Listen to the artifact reducer’s removed-audio output before keeping its changes.

De-clipping cannot recover information that is completely lost, and high-frequency restoration may generate content that needs further adjustment.

Reference Auto-EQ does not make a correction

  1. Confirm that a Reference preset is selected.
  2. Connect the reference audio.
  3. Check the warning messages.

Without reference audio, Reference mode skips correction. Alternatively, use a non-reference preset such as Warm - gentle.

The final sample rate does not change

  1. Set the requested rate in the Output rate node.
  2. Leave the master’s rate at keep.
  3. Check whether the central Mastering switch is off.

The central switch skips the whole mastering area, including rate conversion. Individual processor bypasses inside that area behave differently.

A saved workflow does not show the current controls

  1. Keep a copy of your personalized workflow.
  2. Open the current bundled workflow JSON.
  3. Compare the grouped controls and wiring before transferring your settings.

Updates do not automatically replace saved personal workflows. The 3.1 series consolidated the examples into the two workflow files linked above.

Frequently asked questions

Do I need to use cloud services?
No. Local text, music, and artwork options are supported. Cloud language models are optional.

Is selecting MiniMax Music 3 the same as using a cloud MiniMax API?
No. The documented music-generation path uses local model weights. MiniMax is also listed separately as an optional cloud provider for the text-writing stage.

Do I need a language model for audio enhancement?
No. The enhancement workflow does not need one.

Can I work without generated artwork?
Yes. Turn Cover OFF in CHOOSE. This does not disable the YuE2 Cover music mode.

Does Interpretation Freedom change my locked lyrics?
No. Lyrics policy is independent of the slider.

Does freedom 0 guarantee an identical recording?
No. It enforces the source score rather than an identical audio performance. The generated sound is still a new rendering.

Can I use only a CPU?
The documentation lists CPU-oriented text and transcription options, and mastering processors do not require GPU memory. Music and artwork generation are described as impractical for interactive CPU-only use.

Why does the master miss the loudness target?
Peak and gain-reduction limits take priority when the requested loudness cannot be reached safely. Read the production report.

For additional help, see Troubleshooting.

Learn More about ComfyUI-MiniMax-Music-Production-Toolkit

Start with creative examples

Explore covers and lyrics

Improve the finished sound

Customize the supporting tools

ComfyUI-MiniMax-Music-Production-Toolkit Limitations and License

Always listen before publishing. Musical-score validation, transcription checks, restoration, and mastering provide useful safeguards, but none can guarantee a perfect artistic or technical result.

In particular:

  • Transcription can mishear or hallucinate words.
  • A score cannot force an exact generated performance.
  • Instrumental requests can still produce voice-like material.
  • Restoration cannot recover all missing information.
  • EQ and mastering cannot repair every arrangement or mix problem.
  • Encoded exports should be checked when final loudness and peaks matter.

This is an independent community project. The toolkit is licensed under MIT, but model weights are not included and remain subject to their own licenses and terms.

See LICENSE and NOTICE.md.

ComfyUI-MiniMax-Music-Production-Toolkit Related Nodes

RunComfy
Copyright 2025 RunComfy. All Rights Reserved.

RunComfy is the premier ComfyUI platform, offering ComfyUI online environment and services, along with ComfyUI workflows featuring stunning visuals. RunComfy also provides AI Models, enabling artists to harness the latest AI tools to create incredible art.

ComfyUI-MiniMax-Music-Production-Toolkit detailed guide | ComfyUI