Music Production Toolkit 3.1.3

Music Production Toolkit 3.1.3

A Swiss Army knife for local music production on ComfyUI.

A patch on top of 3.1.2: the model dropdown now says what is already downloaded and
offers six more chat models, one chat model is the default in every LLM node, the
restoration chain ships switched off, and two download problems were fixed.

The model dropdown says what is already there

The dropdown lists installed GGUFs and catalog candidates, and until now nothing
distinguished them — so choosing a model could start a several-gigabyte download without
saying so.

• Every entry is labelled. ✔ marks a file that is already in models/llm, ⬇ a
  catalog model that is not downloaded yet, with its download size and its star rating
  appended, and the widget line reads Model - 3 installed, 16 to download.
• The labels are display only. The value stored in a workflow stays the plain file
  name, so an existing workflow keeps loading after that download — nothing is rewritten
  by looking at a file. If the inventory cannot be read (Python side updated, ComfyUI not
  restarted yet), the dropdown falls back to plain names instead of guessing.

Six more chat models

Each one verified against its repository (pinned commit, exact byte size, publicly
readable) and rated for this toolkit's tasks, with the reason recorded in rating_note:

• gemma-4-26B_q4_0-it.gguf (13.45 GiB, ★★★★★) — Google's quantization-aware 26B
  mixture-of-experts with about 4B active parameters: the largest quality step that still
  fits a 16–24 GiB card, and faster than a dense model of that size.
• gemma-4-26B-A4B-it-UD-Q5_K_M.gguf (19.70 GiB, ★★★★☆) — the same model with more
  precision, for a 32 GiB card that can afford it.
• gemma-4-E4B_q4_0-it.gguf (4.80 GiB, ★★★★☆) — the small Gemma 4, same family as the
  12B QAT, for 8 GiB cards and quick drafts.
• The fast LFM2.5 line: LFM2.5-8B-A1B-Q4_K_M.gguf (4.80 GiB, ★★★☆☆, an 8B
  mixture-of-experts with ~1B active parameters per token), LFM2.5-2.6B-Q4_K_M.gguf
  (1.56 GiB, ★★★☆☆) and LFM2.5-1.2B-Instruct-Q4_K_M.gguf (0.68 GiB, ★★☆☆☆) for CPU runs.
  Their speed is a property; their quality on this toolkit's long structured prompts is
  not measured — the rating note says so.

All six are optional + no_auto_download like every other candidate: the model check
reports them, the model selected in the LLM node is the one that is fetched. The Gemma 4
26B/E4B need the gemma4 architecture in the llama.cpp build and LFM2.5 needs lfm2 /
lfm2moe; the build this release was tested against knows all three.

Qwen_Qwen3.5-9B-Q4_K_M.gguf is the default chat model

A 9B Q4_K_M at 6.2 GiB is the everyday size for this workload — it fits an 8–12 GiB card
as well as a larger one — and it is a catalog entry, so a first run that starts with the
default fetches the file itself. One constant feeds the chat node, and the central
settings node derives its fields from that node, so the same name is the default in every
LLM node. The bundled example workflow stores it in all four places.

The restoration chain ships off

FlashSR is the expensive part of the audio chain: 2.3 GB of weights on first use and
minutes per song. In the main workflow it was already off for YuE2 and on for MiniMax; in
the Audio Enhancement Lab it was on unconditionally.

• Both example workflows now ship with the chain off — the Lab's Refinement enabled
  switch is false, the PRE low-pass is bypassed (bypass=true) and the crossover stands
  at Original SRC only. Those two stages exist for FlashSR, so leaving them on while it
  is off was misleading — and at Original SRC only the D stage would have run and thrown
  the FlashSR signal away.
• Switching the chain on means switching its helpers on as well: B bypass=false
  (PRE 10 kHz is the starting point) and D mode=FlashSR only (Original + FlashSR air
  blends the original low band back in). Both workflow notes say so, as does
  the workflow documentation.
• The FlashSR stage lost its own auto_download switch. It was a second download
  toggle next to the model check node's flags, and with it OFF a missing weight set turned
  the stage into a silent no-op. A stage that runs now fetches the weights it needs —
  logged with progress, resumable, disk-space checked — exactly like the Whisper
  checkpoint, and a fetch that fails skips the stage with one warning line instead of
  ending the run. The Refinement gate stays the switch that decides whether the stage runs
  at all; the model check node stays the place where downloads are decided for a run.

Fixed in 3.1.3

• A chat model picked from the catalog could not download at all. The catalog names
  the LLM folder once, on the group; the loader handed the raw per-file entry to the
  downloader, which refuses an entry without a target — so selecting a catalog model ended
  in LLM model download failed: entry has no target directory. The loader now hands over
  the effective folder (the group's value, models/llm as the fallback) and checks it
  before starting anything.
• The LLM check in the model check node is on by default again. It had been switched
  off as the guard against a multi-gigabyte download nobody asked for; since the catalog
  now marks every chat-model candidate optional and no_auto_download, the check
  reports which candidates are present without transferring anything.
• Two documentation claims about the FlashSR weights were wrong and are corrected: the
  stage skips when it cannot fetch them (it does not "fail fast"), and the weights come
  from laion/FlashSR_One-step_Versatile_Audio_Super-resolution since the earlier
  jakeoneijk/FlashSR_weights dataset stopped answering in June 2026.

Assets

• Music_Production_Toolkit_v3.1.3.json
• Music_Production_AudioEnhance_v3.1.3.json
• ComfyUI-MiniMax-Music-Production-Toolkit-v3.1.3.zip
• SHA256SUMS.txt

The previous release archives stay where they are.

Upgrading

Replace the toolkit folder, restart ComfyUI and refresh the browser. Re-import the two
bundled workflows if you want the new defaults (chain off, PRE bypassed, crossover
Original SRC only, the 9B chat model); an existing workflow keeps working unchanged.

One node input was removed: the FlashSR stage's auto_download widget. ComfyUI maps
widgets positionally, and the two remaining widgets (lowpass_input, output_sr) kept
their places, so workflows saved with the old node still load — the stale value and the
stale input entry are dropped the next time such a workflow is saved. If you had switched
that toggle off deliberately: the stage now fetches its weights whenever it runs, which is
what the Refinement gate is for.

Model files you already have stay valid. Everything new in the catalog is a catalog
entry — nothing is downloaded until you select or check it.

Music Production Toolkit 3.1.2 - Setup Help, Model Advisor

Music Production Toolkit 3.1.2

A Swiss Army knife for local music production on ComfyUI.

A patch on top of 3.1.1: the toolkit now looks at the PC it runs on and says which
models suit it, the language-model settings are entered once instead of three times,
long stages report their progress the way ComfyUI's own nodes do, and two broken
downloads were fixed.

Setup now picks models for your PC

The catalog always knew how to download a model; nothing knew whether that model
suits the machine in front of you.

• New node Model advisor · what suits this machine (in the shipped workflow next
  to the model check). It reads the detected hardware — CPU cores, RAM, the class of
  the accelerator — and reports, per task, which file fits, how large it is and how it
  is rated. The verdict is measured against the free budget with a stated margin,
  so "the weights fit" and "the weights plus context fit" stay two different answers,
  and an unreadable device is reported as unknown instead of being guessed.
• Every task is checked as a combination. A diffusion model and its text encoder
  load in the same run, so on a 12 GiB card the full-precision FLUX.2 pair is reported
  as too large together with the fp8/fp4 set (7.7 GiB) that does fit.
• Star ratings for the job at hand. Every catalog entry carries a 1–5 star rating
  with its reason — suitability for this toolkit's tasks (structured caption, lyrics
  and score output, generation quality, transcription), not a benchmark of the model in
  general. The stars are also printed in the model check's report and documented in
  installation.
• Smaller alternatives that actually exist, each verified against its repository
  (pinned commit, exact byte size) and loadable with ComfyUI's own loaders — no extra
  custom node:
  • language models: 2B, 4B, 9B-distill, 12B and a plain llama-architecture 8B for
    older llama.cpp builds, on top of the existing 9B/12B/27B candidates;
  • MiniMax Music 3: the int8 diffusion model (2.3 GiB) and the pruned int8 text
    encoder — the pair that makes a 12–16 GiB card workable;
  • YuE2: the 3B int8 checkpoint (3.7 GiB);
  • FLUX.2 klein: fp8 diffusion plus the fp4 text encoder;
  • Whisper: large-v3-turbo (1.5 GiB) and a turbo int8 (0.8 GiB) beside the
    reference large-v3.
• A model you selected downloads itself. The Whisper node now fetches the
  checkpoint folder you picked (the group checkbox keeps fetching only the default), and
  the LLM node fetches the GGUF you selected — the download happens where the decision
  was made, not by a checkbox that would pull in several alternatives.
• What to expect from your PC is now a table in the README: per configuration
  (CPU only, 6–8, 10–12, 16, 24 and 32+ GiB) it names the fitting music model, language
  model, Whisper checkpoint and artwork pair, the speed band per stage and the rating of
  each choice — marked as a starting point, not a measurement.

Enter the language-model settings once

• LLM settings · central holds provider, model, context and sampler values once
  and hands them to every LLM chat node in the workflow; its values win field by field,
  so a half-filled central node changes nothing it was not asked to change, and an
  unreadable payload is ignored with a log line instead of stopping a run. Per-call
  settings (enabled, the prompt texts, reset_session) stay on the chat node.
• A leftover model name no longer blocks a run. ComfyUI validates every node's
  widgets before execution and hands a linked input over as a placeholder (never as the
  payload), so a connected settings node is treated as the decision — the model is
  checked when it is loaded.

Runs are easier to follow

• The LLM and FlashSR stages draw the bar ComfyUI's own nodes draw (tqdm, the one
  YuE2 uses): one line that updates in place with count, percentage, elapsed time,
  remaining time and rate — LLM streaming: 8%|# | 1958/24576 [01:15<14:24, 26.1token/s].
  No more one log line per step; the log keeps the start line and one summary per stage.
• A refinement stage that cannot run switches itself off. If FlashSR's weights are
  neither installed nor downloadable, the stage logs why and passes the audio through
  unchanged instead of ending a run that can still produce the song.
• A damaged source file is refused before the run starts, with its name in the
  message. The cover transcription node decodes the file once and reports
  the source audio 'x.mp3' cannot be decoded past 2:42 (…) instead of letting
  ComfyUI's loader fail 1.5 seconds in with a traceback that never named the file.

Fixed in 3.1.2

• The FlashSR weights pointed at a repository that stopped answering. Every request
  to jakeoneijk/FlashSR_weights returns HTTP 401 repo-wide, so a fresh install could
  not download the three weights at all. The catalog now reads the identical files —
  same byte sizes — from the authors' repository.
• A chat model could be listed and still refuse to download itself. The loader asked
  the catalog entry for a url field, which the current entries do not have (they name
  a repository, a pinned commit and a remote filename), so every catalog model answered
  "no download URL is configured".
• Whisper alternatives were unreachable although they were in the catalog: the group
  checkbox fetches only the default checkpoint by design, and nothing fetched the selected
  one. The lyrics node does now, and a checkpoint folder that exists without model.bin
  is reported as an incomplete download instead of being handed to the engine.
• Layout: the Model advisor sits inside 01 · START / Files & models, and one node
  that had been overlapping its neighbour was nudged back into place.

Assets

• Music_Production_Toolkit_v3.1.2.json
• Music_Production_AudioEnhance_v3.1.2.json
• ComfyUI-MiniMax-Music-Production-Toolkit-v3.1.2.zip
• SHA256SUMS.txt

The previous release archives stay where they are.

Upgrading

Replace the toolkit folder, restart ComfyUI and refresh the browser. Re-import the two
bundled workflows to get the new Model advisor node and the central LLM settings node;
everything else loads unchanged — no node input was renamed, reordered or removed, and
an existing workflow keeps working without the new nodes (the settings node is optional
everywhere it is wired).

Model files you already have stay valid; the new entries are catalog entries, so
nothing is downloaded until you select or check it.

Music Production Toolkit 3.1.1

Music Production Toolkit 3.1.1

A Swiss Army knife for local music production on ComfyUI.

A patch on top of 3.1.0: one cover stopper fixed, every run now logged with its date,
time and source file, and fresh branding for a toolkit that is no longer MiniMax only.

The cover feature is the highlight

Turn an existing track into your own version of it — with as much or as little
control as you want. SheetSage2 reads the actual musical score out of your
recording, and everything after that works on real notes.

• One slider, 0 to 100. From a faithful re-recording to a free recomposition that
  keeps only the character. At 0 the source score is enforced: a model answer that
  rewrites a note, a chord, a bar or the tempo is rejected, and a full cover keeps
  its harmony at every freedom level.
• Advanced mode, if you want it. Element by element: main melody, chorus hook,
  structure, harmony, tempo and key, plus how far melody, rhythm, harmony and
  structure may vary, a key shift in semitones, a tempo change and the vocal range of
  the new version.
• The words never follow the slider. Lock the source transcription or your own
  lyrics and they stay exactly as they are.
• Instrumental that stays instrumental. The notes that carried the vocals move
  into the native instrumental part and are played by the lead instrument you
  name, while the vocal line becomes rests. If you want proof, the optional vocal
  check transcribes the render, counts the words and re-renders within your limit —
  keeping the least vocal take when none is clean.
• The style you chose always wins. More freedom means that less of the original
  audio survives, never that less of your template is delivered.
• Nothing unchecked reaches the engine. Every model answer is validated against
  the native ABC dialect and the plan first; an answer that breaks the contract is
  replaced by the validated deterministic result.

The minimum is a file, a style and a queue. Everything else is optional and off the
critical path.

Fixed in 3.1.1

original lyrics covers stopped before the first note. The bundled workflow
carried the toolkit's own "field not set" placeholder in the Whisper node's language
field — a value that node never offers — and the run aborted with

YuE2 Cover: unsupported Whisper source language 'custom'

before a single frame was decoded. The shipped workflow now stores auto, and a
workflow test refuses any stored value the node does not offer. A misspelled language
is still rejected — transcribing in the wrong language silently would be worse — but a
saved workflow from an earlier version that still carries the old placeholder now
auto-detects instead of stopping.

Your runs are easier to follow now

• Every log line carries the date and time (2026-09-18 14:22:31 Saved artwork: …),
  so a ComfyUI log reads as a timeline: when a run started, how long a stage took,
  which result belongs to which attempt.
• A cover run names the audio file it was made from — the file, its full path, its
  size and the three cover choices, once when the file is selected and again for every
  generated song. The exports only carry the derived <name>-cover title, so the log
  is what tells you what a run was built on later.
• Documented, not mysterious: the _ProactorBasePipeTransport error a long Windows
  run can print is CPython reporting a dropped client connection — usually the browser
  tab that queued the prompt. The run and its files are unaffected. See
  troubleshooting.

Branding for more than MiniMax

The toolkit has grown past its name: YuE2, YuE2 Cover and MiniMax Music 3 live side by
side, and the new banner and icon say so. The README now pictures both workflows where
they are described — the main workflow end to end, and the audio-enhancement workflow
that carries every restoration and mastering stage of the main one.

Assets

• Music_Production_Toolkit_v3.1.1.json
• Music_Production_AudioEnhance_v3.1.1.json
• ComfyUI-MiniMax-Music-Production-Toolkit-v3.1.1.zip
• SHA256SUMS.txt

The previous release archives stay where they are.

Upgrading

Replace the toolkit folder, restart ComfyUI and refresh the browser. Re-import the two
bundled workflows if you keep your own copies: 3.1.1 corrects one value in the main
workflow (the Whisper node's language field now reads auto). Everything else loads
unchanged — no node input was renamed, reordered or removed.

Music Production Toolkit 3.1.0

MiniMax Music Production Toolkit 3.1.0

A Swiss Army knife for local music production on ComfyUI.

This release adds the Cover Studio, an honest instrumental vocal check, and
consolidates the example workflows into the two that are actually needed.

Highlights

Cover Studio: one slider for how far a cover may go

A new Interpretation Freedom slider runs from 0 — a faithful cover that keeps
the melody, structure, harmony, tempo and key and only changes the sound — to 100
— a creative recomposition that keeps only the character of the original.

It is a toolkit idea, not a YuE2 setting, and it is honest about what it does:

• The style never weakens. More freedom means that less of the original audio
  survives, never that less of the template you chose is delivered.
• At 0 it really is faithful. A model that rewrites a note, a chord, a bar or
  the tempo is rejected and the untouched score is used.
• The style you asked for always arrives. The requested style travels into the
  rework on its own path, so the rework can never quietly replace it.
• The words never follow the slider. Keep the source transcription or your own
  lyrics exactly as they are, whatever the slider says.
• The result is checked before the engine sees it. The score is validated
  against the native ABC dialect; if the model's answer does not hold up, the
  validated deterministic result is used instead.

Instrumental covers that stay instrumental

YuE2 can add vocal-like material to an instrumental even though the score contains
no vocal notes at all. Since that can only be heard and not promised, the toolkit
now listens: the optional instrumental vocal check transcribes the freshly
generated audio, and a take that still contains words is re-rendered — as often as
you allow, up to ten times.

If no take reaches your tolerance, the take with the fewest recognised words
is used, so a long run always ends with the best audio it produced rather than
erroring out. Every take is written to a temporary file while the run is in
progress and the words Whisper heard are logged, so you can see whether it was
humming or a full chorus — and the candidates that lost are removed.

You set how many words are still acceptable, and the check runs on the raw render,
before any EQ, mastering or encoding, so the retries stay as cheap as they can be.

Two workflows instead of three

• Music_Production_Toolkit.json — the main workflow. YuE2, YuE2 Cover and
  MiniMax Music 3, the Cover Studio, and the full mastering and release chain.
• Music_Production_AudioEnhance.json — bring a finished recording and enhance
  it. It now carries every restoration and mastering stage of the main workflow,
  including the experimental artifact reduction and the bypass gates.

The AudioEnhance workflow also copies your original file's tags and cover art
onto the new export, so an enhanced version still looks like the file it came
from.

What the toolkit does

One toolkit instead of a shelf of scripts. Everything below is part of the two
shipped workflows, and everything runs on your own machine.

• Write a song from a description. Pick a template or write your own brief; a
  local or cloud language model turns it into a complete song request with
  arrangement, synchronized lyrics, a title and an album-cover idea.
• Cover an existing track. Read the score from your own audio and render a new
  version in the style you describe - instrumental, with new words, or with the
  words that are already there.
• Direct how far the cover may go. One slider, from a faithful rendition to a
  free recomposition, plus an advanced mode that decides element by element what
  stays.
• Repair and restore a recording. De-click and de-clip, band-limited
  restoration, high-frequency repair, optional artifact reduction, automatic and
  manual EQ, sample-rate conversion and mastering to a target loudness.
• Compare honestly. Every processing stage can be switched off, and the run
  records what ran, what was skipped and with which settings.
• Render artwork. A FLUX.2 cover built from the song's own image prompt.
• Tag and release. Title, artist, album, track, genre and embedded cover art in
  FLAC and MP3, plus a production record and a prompt report for every run.
• Bring your own models. YuE2 and MiniMax Music 3 for songs, SheetSage2 for
  scores, faster-whisper for lyrics, any local or cloud language model for the
  text, and FLUX.2 for the artwork - each downloaded on demand and checked before
  a run.

Nothing leaves your computer unless you deliberately choose a cloud model for the
text.

One default changed

The standard cover lyrics mode is now instrumental. It needs no speech engine
and no transcription, so a fresh install produces a cover straight away. If you
have a saved workflow from an earlier version, set the mode explicitly when you
want new or original words.

Nothing else was taken away

Existing saved workflows keep loading. The nodes they use keep their inputs, in the
same order, with the same defaults; the new options are appended and switched off
unless you turn them on.

Assets

• Music_Production_Toolkit_v3.1.0.json
• Music_Production_AudioEnhance_v3.1.0.json
• ComfyUI-MiniMax-Music-Production-Toolkit-v3.1.0.zip
• SHA256SUMS.txt

The previous release archives stay where they are.

Upgrading

Replace the toolkit folder and restart ComfyUI. Load
Music_Production_Toolkit.json for new work; it replaces the earlier YuE2 and
MiniMax example workflows. If you have saved copies of those, they still load.

MiniMax Music Production Toolkit 3.0.1 for ComfyUI

MiniMax Music Production Toolkit 3.0.1 for ComfyUI

New in 3.0: Turn an audio file into a YuE2 cover version. Upload your track,
describe a new musical direction and take it through generation, mastering,
artwork and export in one connected workflow.

The prompt follows the music. SheetSage2 extracts a musical score as readable
ABC notation. Your LLM uses that score to adapt the cover's Style and Lyrics
plan, while YuE2 receives the original transcription for generation. This is
the first cover implementation in the toolkit—and a first step toward doing
more with an open, readable musical description.
Try YuE2 Cover · What's new in 3.0.1

3.0.1 adds more control over the sound: 10 Auto-EQ presets, 24 manual EQ
presets (including YuE2 Smooth highs), and switchable experimental artifact
reduction. Auto-EQ starts at Warm - gentle; artifact reduction starts
on / Balanced. Reference matching without a reference now warns and skips
correction instead of stopping production.

From a song idea to music, mastering and cover artwork — in one connected workflow.

Describe the music you want to make. Choose a genre, a mood, a voice or a lyrical
theme, and let the toolkit turn your idea into a production brief for YuE2 or MiniMax
Music 3. Generate the song, refine its sound, shape the final master and save
your audio, artwork and production record together.

Already have a song? Choose YuE2 Cover to create a new arrangement, or open
the Audio Enhancement Lab to enhance and master the existing recording.

Created by Johannes Plenio.
Listen to the demo gallery.

What's new in 3.0

YuE2 cover versions are here. Choose YuE2 Cover in CHOOSE and select an
audio file in SOURCE AUDIO. SheetSage2 extracts its score, the LLM prepares a
source-aware arrangement, and YuE2 generates the cover through the existing
production chain. New songs with YuE2 and MiniMax Music 3 remain available.

• A prompt informed by the source. The ABC transcription goes into the LLM
  brief, so the proposed style, instrumentation and development can follow the
  source's musical phrases. The original ABC passes unchanged to YuE2.
• Two ways to reinterpret a track. full conditions on melody and harmony;
  melody gives a new accompaniment more freedom. One source setting controls
  both transcription and generation.
• Ready for the complete workflow. SheetSage2 BF16 check/autoload is enabled
  for the selected cover mode. YuE2 defaults stay at 40 steps and a 360-second
  ceiling, with Refinement off and Mastering/artwork on. The source filename
  plus -cover becomes the title across tags, exports and reports.
• More developed YuE2 arrangements. All twelve system prompts now request
  a detailed chronological Style plan with corresponding Lyrics sections,
  including instrumental tracks. Motifs, transitions and instrumental roles
  develop over the track. The previous prompts are archived for comparison.
• Working audio selection and preview. The release includes the frontend
  fix for the red/UNKNOWN Cover source node in the first cover build.
• Song length reaches the generator. YuE2 prompts now plan timed sections
  with matching Lyrics and a natural ending near the requested Length. This is
  an approximate target: phrases and decay can finish beyond it, up to the
  separately configured maximum. The production record compares target and
  actual duration. The model can still finish early.

SheetSage2 extracts musical notation, not the original lyric words. Supply
the lyrics you want, request new lyrics, or select instrumental. This first
version is a starting point for experimentation; more ABC-based ideas are in preparation.

Open Yue2_MM3_Production_Toolkit.json
and follow the cover guide.

What's new in 2.6

Choose YuE2 or MiniMax Music 3 in one workflow. The brand-new YuE2 model brings
excellent musical quality to the toolkit: its developers report competitive
results with leading song generators in their September 2026 evaluation.
Official YuE2 results.
The default YuE2 production path is faster to finish because it skips the entire
Restore/FlashSR section and proceeds straight to mastering. Actual generator
speed depends on hardware, duration and settings; no MiniMax/YuE2 timing benchmark
is claimed for this toolkit.

• Everything starts in CHOOSE. YuE2 is the default. Cover and Mastering are
  on; Refinement follows the model (MiniMax on, YuE2 off), with manual overrides.
• Prompts suited to each model. Twelve YuE2 system prompts generate Style and
  Lyrics with native ABC planning. Instrumental requests explicitly exclude sung
  and spoken text; their Lyrics section contains only instrumental structure tags.
  At most, explicitly requested background humming is described in Style.
• Ready-to-use YuE2 settings. 40 steps, a separate 360-second ceiling and
  yue2_3b_bf16.safetensors, with missing-model download enabled by default.
• More readable reports and mastering choices. Lyrics preserve line breaks in
  Markdown and Windows files. Twelve compressor presets plus Custom include
  Balanced - gentle glue, which retains the previous settings.

Open Yue2_MM3_Production_Toolkit.json
and read the YuE2 guide or 2.6.0 release notes.
The classic MiniMax and Audio Enhancement Lab workflows remain available.

What's new in 2.5.2

A bug-fix release: no new controls, no changed settings.

• Safer audio export. Both audio savers check the whole batch before peak
  handling and encoding, so invalid samples can no longer end in a blank
  SoundFile assertion, and the encoder messages name the format and library
  involved. Both decoder branches check their input and their output, retry the
  decoding once with smaller tiles instead of writing broken audio, and never
  replace bad samples with silence. The music sampler no longer repeats an
  identical non-finite run; it stops immediately and names the remedy.
• Preview image and song. The generated cover and the finished track stay
  visible right in the workflow: PreviewImage shows the artwork as soon as it
  is saved, and PreviewAudio plays the exported song.

What's new in 2.5.1

• Use any language model. The LLM node runs the bundled GGUF inside ComfyUI,
  a model served by a local app (LM Studio, Ollama, llama.cpp, Unsloth Studio,
  vLLM) or a cloud service (OpenAI, Claude, Gemini, DeepSeek, Qwen, MiniMax,
  OpenRouter, Groq) — with model discovery, a connection setup dialog and
  session-only API keys that are never written into the workflow.
• Switch the cover off. A dedicated control in the cover artwork area
  turns FLUX.2 artwork off completely: no rendering, no cover file, no FLUX
  download. Your audio export continues as usual.
• Smaller fixes. The cover switch now sits where the documentation says it
  does, and the release checks no longer depend on timing.

What's new in 2.5

This release brings the improvements together into a clearer, more capable music
production toolkit:

• Workflows that are easier to follow. Numbered stages, labelled controls and
  dedicated areas for song creation, audio processing, mastering, artwork and export.
• A complete mastering section. Auto-EQ, an editable manual 8-band
  parametric EQ, and a stereo-linked compressor with LUFS targeting and a
  true-peak limiter.
• Independent control. Use automatic tonal shaping, make your own EQ
  adjustments, combine both, or switch either off. Disable compression while
  keeping loudness and peak control active.
• More efficient processing. Improvements to model/resource handling,
  block-based audio processing and bounded working buffers help avoid unnecessary
  work and memory use. The new mastering tools run on the CPU and need no extra VRAM.
• More reliable everyday operation. Improvements across model downloads,
  prompt handling, file output, metadata and regression testing.

The redesigned workflows now use the familiar filenames. There is no separate
“optimized” version to choose between.

Choose your workflow

Workflow	Start with	What it produces
YuE2 / MiniMax Production + YuE2 Cover	A song idea, prompt template or source audio	New songs or cover versions, optional artwork/refinement/mastering, original and release audio, tags and production record
Classic MiniMax Production	A song idea or prompt template	MiniMax music, original and mastered audio, cover artwork, tags and a production record
Audio Enhancement Lab	An existing audio file	Enhanced and mastered FLAC with your tags

Open the JSON in ComfyUI or drag it onto the canvas. The notes inside each
workflow explain where to start and which controls matter.

Create a new song

1. Choose YuE2 or MiniMax Music 3 in CHOOSE. Set your output folder, artist and album.
2. Choose a prompt template or describe your own idea. Adjust genre, tempo,
   language, voice and length as needed.
3. Check the model settings for your computer, th...

ComfyUI Music Production Toolkit 3.0.0

MiniMax Music Production Toolkit 3.0 for ComfyUI

New in 3.0: Turn an audio file into a YuE2 cover version. Upload your track,
describe a new musical direction and take it through generation, mastering,
artwork and export in one connected workflow.

The prompt follows the music. SheetSage2 extracts a musical score as readable
ABC notation. Your LLM uses that score to adapt the cover's Style and Lyrics
plan, while YuE2 receives the original transcription for generation. This is
the first cover implementation in the toolkit—and a first step toward doing
more with an open, readable musical description.
Try YuE2 Cover · What's new in 3.0.0

From a song idea to music, mastering and cover artwork — in one connected workflow.

Describe the music you want to make. Choose a genre, a mood, a voice or a lyrical
theme, and let the toolkit turn your idea into a production brief for YuE2 or MiniMax
Music 3. Generate the song, refine its sound, shape the final master and save
your audio, artwork and production record together.

Already have a song? Choose YuE2 Cover to create a new arrangement, or open
the Audio Enhancement Lab to enhance and master the existing recording.

Created by Johannes Plenio.
Listen to the demo gallery.

What's new in 3.0

YuE2 cover versions are here. Choose YuE2 Cover in CHOOSE and select an
audio file in SOURCE AUDIO. SheetSage2 extracts its score, the LLM prepares a
source-aware arrangement, and YuE2 generates the cover through the existing
production chain. New songs with YuE2 and MiniMax Music 3 remain available.

• A prompt informed by the source. The ABC transcription goes into the LLM
  brief, so the proposed style, instrumentation and development can follow the
  source's musical phrases. The original ABC passes unchanged to YuE2.
• Two ways to reinterpret a track. full conditions on melody and harmony;
  melody gives a new accompaniment more freedom. One source setting controls
  both transcription and generation.
• Ready for the complete workflow. SheetSage2 BF16 check/autoload is enabled
  for the selected cover mode. YuE2 defaults stay at 40 steps and a 360-second
  ceiling, with Refinement off and Mastering/artwork on. The source filename
  plus -cover becomes the title across tags, exports and reports.
• More developed YuE2 arrangements. All twelve system prompts now request
  a detailed chronological Style plan with corresponding Lyrics sections,
  including instrumental tracks. Motifs, transitions and instrumental roles
  develop over the track. The previous prompts are archived for comparison.
• Working audio selection and preview. The release includes the frontend
  fix for the red/UNKNOWN Cover source node in the first cover build.
• Song length reaches the generator. YuE2 prompts now plan timed sections
  with matching Lyrics and a natural ending near the requested Length. This is
  an approximate target: phrases and decay can finish beyond it, up to the
  separately configured maximum. The production record compares target and
  actual duration. The model can still finish early.

SheetSage2 extracts musical notation, not the original lyric words. Supply
the lyrics you want, request new lyrics, or select instrumental. This first
version is a starting point for experimentation; more ABC-based ideas are in preparation.

Open Yue2_MM3_Production_Toolkit.json
and follow the cover guide.

What's new in 2.6

Choose YuE2 or MiniMax Music 3 in one workflow. The brand-new YuE2 model brings
excellent musical quality to the toolkit: its developers report competitive
results with leading song generators in their September 2026 evaluation.
Official YuE2 results.
The default YuE2 production path is faster to finish because it skips the entire
Restore/FlashSR section and proceeds straight to mastering. Actual generator
speed depends on hardware, duration and settings; no MiniMax/YuE2 timing benchmark
is claimed for this toolkit.

• Everything starts in CHOOSE. YuE2 is the default. Cover and Mastering are
  on; Refinement follows the model (MiniMax on, YuE2 off), with manual overrides.
• Prompts suited to each model. Twelve YuE2 system prompts generate Style and
  Lyrics with native ABC planning. Instrumental requests explicitly exclude sung
  and spoken text; their Lyrics section contains only instrumental structure tags.
  At most, explicitly requested background humming is described in Style.
• Ready-to-use YuE2 settings. 40 steps, a separate 360-second ceiling and
  yue2_3b_bf16.safetensors, with missing-model download enabled by default.
• More readable reports and mastering choices. Lyrics preserve line breaks in
  Markdown and Windows files. Twelve compressor presets plus Custom include
  Balanced - gentle glue, which retains the previous settings.

Open Yue2_MM3_Production_Toolkit.json
and read the YuE2 guide or release notes.
The classic MiniMax and Audio Enhancement Lab workflows remain available.

What's new in 2.5.2

A bug-fix release: no new controls, no changed settings.

• Safer audio export. Both audio savers check the whole batch before peak
  handling and encoding, so invalid samples can no longer end in a blank
  SoundFile assertion, and the encoder messages name the format and library
  involved. Both decoder branches check their input and their output, retry the
  decoding once with smaller tiles instead of writing broken audio, and never
  replace bad samples with silence. The music sampler no longer repeats an
  identical non-finite run; it stops immediately and names the remedy.
• Preview image and song. The generated cover and the finished track stay
  visible right in the workflow: PreviewImage shows the artwork as soon as it
  is saved, and PreviewAudio plays the exported song.

What's new in 2.5.1

• Use any language model. The LLM node runs the bundled GGUF inside ComfyUI,
  a model served by a local app (LM Studio, Ollama, llama.cpp, Unsloth Studio,
  vLLM) or a cloud service (OpenAI, Claude, Gemini, DeepSeek, Qwen, MiniMax,
  OpenRouter, Groq) — with model discovery, a connection setup dialog and
  session-only API keys that are never written into the workflow.
• Switch the cover off. A dedicated control in the cover artwork area
  turns FLUX.2 artwork off completely: no rendering, no cover file, no FLUX
  download. Your audio export continues as usual.
• Smaller fixes. The cover switch now sits where the documentation says it
  does, and the release checks no longer depend on timing.

What's new in 2.5

This release brings the improvements together into a clearer, more capable music
production toolkit:

• Workflows that are easier to follow. Numbered stages, labelled controls and
  dedicated areas for song creation, audio processing, mastering, artwork and export.
• A complete mastering section. Auto-EQ, an editable manual 8-band
  parametric EQ, and a stereo-linked compressor with LUFS targeting and a
  true-peak limiter.
• Independent control. Use automatic tonal shaping, make your own EQ
  adjustments, combine both, or switch either off. Disable compression while
  keeping loudness and peak control active.
• More efficient processing. Improvements to model/resource handling,
  block-based audio processing and bounded working buffers help avoid unnecessary
  work and memory use. The new mastering tools run on the CPU and need no extra VRAM.
• More reliable everyday operation. Improvements across model downloads,
  prompt handling, file output, metadata and regression testing.

The redesigned workflows now use the familiar filenames. There is no separate
“optimized” version to choose between.

Choose your workflow

Workflow	Start with	What it produces
YuE2 / MiniMax Production + YuE2 Cover	A song idea, prompt template or source audio	New songs or cover versions, optional artwork/refinement/mastering, original and release audio, tags and production record
Classic MiniMax Production	A song idea or prompt template	MiniMax music, original and mastered audio, cover artwork, tags and a production record
Audio Enhancement Lab	An existing audio file	Enhanced and mastered FLAC with your tags

Open the JSON in ComfyUI or drag it onto the canvas. The notes inside each
workflow explain where to start and which controls matter.

Create a new song

1. Choose YuE2 or MiniMax Music 3 in CHOOSE. Set your output folder, artist and album.
2. Choose a prompt template or describe your own idea. Adjust genre, tempo,
   language, voice and length as needed.
3. Check the model settings for your computer, then queue the workflow.
4. Listen to the result and adjust the restoration or mastering to taste.

Your chosen LLM prepares the caption, lyrics, title and cover idea: use a GGUF
inside ComfyUI, a model in another local app, or a cloud provider.
The selected YuE2 or MiniMax Music 3 model generates the music. Audio restoration and mastering prepare the
release ...

MiniMax Music Production Toolkit 2.5.2

Release Notes – v2.5.2

Release date: 2026-09-13

Summary

2.5.2 is a bug-fix release. It hardens the audio export path and the music
sampler against invalid numerical output, and it makes the documented
runtime-safety mode work as described. Nothing else changes: node interfaces,
workflow layout, stored settings and the LLM defaults stay as they are.

Preview of image and song. The generated cover and the finished song stay
visible right in the workflow: PreviewImage shows the artwork as soon as the
cover saver has written it, and PreviewAudio plays the exported track. Both are
part of the production example and need no setup.

Fixed

• Audio export never writes invalid samples. Both audio savers validate the
  entire batch before peak handling and encoding. NaN/Infinity audio used to end
  in a blank SoundFile assertion while writing FLAC; the remaining encoder
  assertions now carry format, sample-rate and library diagnostics, and the
  incomplete staging file is removed.
• Both decoder branches check their input and their output. The production
  subgraph uses MiniMaxSafeAudioDecode for the normal and the tiled decoder.
  It validates sampler latents, decoded audio and the normalization step,
  retries decoding once with conservative tiles when only the decoding failed
  numerically, and never replaces bad samples with silence.
• The music sampler no longer repeats an identical non-finite run.
  KSamplerWithConfig failed twice on the same seed in the reporting run: the
  noise follows the seed and the conditioning and weights are unchanged, so the
  second pass reproduced the same latents and only doubled the longest stage of
  the prompt. Non-finite latents now stop immediately with the concrete remedy
  (--bf16-unet, --fp32-unet, or a new seed). The one retry for
  capture/CUDA-graph backend errors — where clearing the captured state can
  help — is unchanged.
• MINIMAX_MUSIC3_RUNTIME_SAFETY=auto works as documented. The policy
  parser folded auto into off, so the risky-backend branch (Blackwell-class
  CUDA or ROCm) could never run and the automatic mode silently did nothing.
  off is still the default and stays inert.
• TROUBLESHOOTING.md documents the remaining non-finite DiT class with the
  dtype decision measured on the reporting machine (float16 without flags;
  --bf16-unet changes only the DiT and is VRAM-neutral; --fp32-unet also
  switches the FLUX.2 cover to fp32) and the experiment order, so the upstream
  issue is not re-diagnosed from scratch.
• Release-gate hygiene. The FFmpeg pipe timeout test no longer races the
  process start (it could fail a release build at random), and the regression
  coverage now includes decoder recovery, input ownership, real FLAC/WAV
  roundtrips, batch validation, failed-export cleanup and the sampler guard.

Breaking changes

None. Node identifiers, inputs, outputs, return orders, widget names and stored
values are unchanged; the example workflows keep their layout and settings.

Upgrade notes

• Restart ComfyUI and hard-refresh the browser after updating.
• Reopen the bundled example workflows; saved personal workflows are not
  rewritten automatically.
• The non-finite latents of the fp16 diffusion model are a ComfyUI/MiniMax issue
  that this release detects early and explains. TROUBLESHOOTING.md lists the
  measured --bf16-unet / --fp32-unet experiments.

Validation

• Full test suite: 866 tests, OK (1 skipped: mutagen is not installed in
  the development environment).
• scripts/validate_release.py: Release validation OK, privacy scan clean.
• scripts/dump_node_contracts.py --check: contract snapshot up to date.
• python -m compileall -q .: clean.
• scripts/package_release.py --dry-run: 34 registered nodes, 850 files,
  privacy scan CLEAN.

Assets

• ComfyUI-MiniMax-Music-Production-Toolkit-v2.5.2.zip
• MiniMax_Music3_Production_Toolkit_v2.5.2.json
• MiniMax_Music3_Production_Toolkit_AudioEnhance_v2.5.2.json
• SHA256SUMS.txt

MiniMax Music Production Toolkit 2.5.1

Release Notes – v2.5.1

Release date: 2026-09-13

Summary

2.5.1 makes the language-model stage a first-class choice. The LLM node now runs
either the bundled GGUF inside ComfyUI, a model served by a local app, or a cloud
service — selected in the node, with the same outputs as before. A new switch
turns FLUX.2 cover generation off completely (rendering, saving and preflight
downloads), and the example workflow's layout, node ownership and release gates
were brought back in line with the documentation.

Everything in 2.5.0 still applies: the redesigned workflows, the mastering
section with auto-EQ, parametric EQ, compressor and true-peak limiter, and the
Audio Enhancement Lab.

Added

• LLM backend selection. The MiniMaxLLMChat node offers three modes:
  In ComfyUI (GGUF) (unchanged default), Local app / server, and
  Cloud service. The node shows only the controls the selected mode needs.
• Local app / server presets: LM Studio (127.0.0.1:1234), Ollama
  (11434), llama.cpp (8080), Unsloth Studio (8888), vLLM (8000), and any
  other OpenAI-compatible server.
• Cloud service presets: OpenAI (Responses API), Claude (Anthropic), Gemini
  (Google), DeepSeek, Qwen (Alibaba Cloud), MiniMax, OpenRouter and Groq, plus
  any other OpenAI-compatible cloud endpoint. Each provider has its documented
  credential environment variable (OPENAI_API_KEY, ANTHROPIC_API_KEY,
  GEMINI_API_KEY, DEEPSEEK_API_KEY, DASHSCOPE_API_KEY, MINIMAX_API_KEY,
  OPENROUTER_API_KEY, GROQ_API_KEY).
• Model discovery and a setup dialog. Available models can be listed from the
  selected endpoint, or entered by hand. The connection setup (backend, provider,
  address, model, key) is reachable from the node.
• Session-only API keys. A key entered in the browser is held in memory for
  the session and is never written into the workflow or to disk.
• FLUX.2 cover switch (MiniMaxCoverControl). One boolean, enabled by
  default, wired to both SaveImageSmartPrefix.enabled and
  MiniMaxModelAutodownload.flux2_models. Switched off, no cover is rendered or
  saved, the FLUX preflight download is skipped, and the audio export continues
  with an empty artwork path. Previously written JPGs are not deleted.
• Documentation and tests. New LLM_PROVIDERS.md setup guide (local apps,
  cloud, keys, output length, optional cover, troubleshooting), updated node
  pages for the three LLM nodes and the cover switch, plus provider, transport,
  UI and cover regression tests including two frontend suites.

Changed

• LLM execution uses ComfyUI's cache invalidation. An enabled run generates
  fresh text instead of replaying the previous completion.
• The session helper is gone. MiniMaxLLMSessionId was removed from the
  example workflow and the LLM node no longer takes a session input; stored
  workflows with those wires migrate on load without disturbing other links.
• The production workflow got a visual pass. The LLM node is titled
  LLM · In ComfyUI / Local app / Cloud, its panel and the prompt-report panel
  are taller so the new controls fit without scrolling, and node positions and
  group sizes were tidied. The saved view opens on the mastering chain and the
  artwork lane, where the cover switch and the master controls live.
• The workflow's generic LLM preflight is off. The integrated LLM downloads
  its selected GGUF on demand; external providers manage their own models.
• The production workflow grew from 48 to 50 nodes (cover switch, preview nodes).

Fixed

• The cover-switch documentation matches the shipped layout. The switch
  lives in 05 · ILLUSTRATE / Cover artwork with the cover nodes it controls —
  its cover_enabled output feeds the image saver and the FLUX.2 preflight
  download next to it. The node page, the README, WORKFLOW_OPTIMIZED.md and
  LLM_PROVIDERS.md now all describe that placement, and the layout test pins
  it together with both connections.
• PreviewImage and PreviewAudio are registered as ComfyUI-core node types
  in the node-ownership test, so the preview nodes in the example workflow are
  covered instead of reported as unknown owners.
• The FFmpeg pipe timeout test no longer races the process start. It used
  0.2 s of audio with a 1 ms deadline, so FFmpeg occasionally finished in time
  and the expected timeout never occurred — a flaky test that could fail a
  release build at random. The workload is now long enough that the deadline is
  always reached while FFmpeg is still working.

Breaking changes

None. Node identifiers, input and output names, return orders and the GGUF
defaults are unchanged. Existing workflows load as before; the removed session
wires migrate automatically.

Upgrade notes

• Open the updated bundled workflow to get the cover switch and the new LLM
  menus. An update never inserts nodes into personal workflows — search for
  FLUX.2 Cover to add the switch to your own, and connect cover_enabled to
  SaveImageSmartPrefix.enabled and MiniMaxModelAutodownload.flux2_models.
• For cloud or local-app use, set the provider's environment variable (or type
  the key into the node for the current session). Unattended and API runs should
  use the environment variable.
• Keep the cover saver active even when the switch is off; some ComfyUI versions
  validate missing model dropdowns before execution. See LLM_PROVIDERS.md.

Validation

• Full test suite: 837 tests, OK (1 skipped: mutagen is not installed in
  the development environment).
• scripts/validate_release.py: Release validation OK, privacy scan clean.
• scripts/dump_node_contracts.py --check: contract snapshot up to date.
• Both example workflows re-validated: graph integrity, group containment, no
  overlapping nodes, canonical node ownership.

Assets

• ComfyUI-MiniMax-Music-Production-Toolkit-v2.5.1.zip
• MiniMax_Music3_Production_Toolkit_v2.5.1.json
• SHA256SUMS.txt

MiniMax Music Production Toolkit 2.5.0

MiniMax Music Production Toolkit 2.5

Release 2.5 brings a clearer workflow, more efficient everyday operation and a
complete mastering section to music generation and existing-audio enhancement.

Highlights

• Redesigned, numbered working areas make both workflows easier to navigate.
• Auto-EQ is on by default, with a gentle Warm tilt preset (35%, maximum 2 dB).
• A separate manual 8-band EQ stays editable with Auto-EQ on or off.
• Stereo-linked compression, LUFS targeting and true-peak limiting finish the
  release signal. Compression can be disabled independently of the limiter.
• Final output defaults to 44.1 kHz; 48 kHz is selectable before mastering.
• Improvements throughout memory/resource handling, audio processing, downloads,
  prompts, metadata and file output support a wider range of computers.

Updating

Update the toolkit and dependencies in the ComfyUI environment, restart ComfyUI
and refresh the browser. Open the bundled production or AudioEnhance JSON again.
The redesigned examples now have the original filenames; the temporary optimized
copies and the older example contents are retired. Keep personal workflows under
separate filenames. Existing saved workflows are not silently replaced.

The production release chain now uses dynamic mastering in place of static gain;
the resulting audio can differ. Auto-EQ can be switched off. Manual EQ starts
neutral. Check the master at matched loudness; target loudness may remain below
the requested value when peak/gain-reduction limits take priority.

Mastering runs on CPU. Generation models still require adequate memory; the
bundled large LLM settings should be reduced for smaller machines. This release
does not promise that every model fits every GPU.

Distribution

The release package contains the toolkit and both canonical workflows. Standalone
versioned workflow JSON files and SHA-256 checksums accompany the ZIP. Model
weights are not included. See INSTALLATION.md for dependencies and model setup.

Prepared locally for tag v2.5.0. Building these assets does not publish a GitHub
release or submit the package to Comfy Registry.

Local validation

The Python suite completed 809 tests with no failures and one skipped test
(optional metadata dependency unavailable in the validation interpreter).
The EQ coefficient parity, prompt UI, workflow migration and headless EQ browser
tests passed. Release structure/version/privacy validation passed as well.
This does not constitute a fresh full-model GPU generation or listening test;
the workflows were previously tested interactively by the maintainer.

MiniMax Music Production Toolkit 2.1.1

Release Notes – v2.1.1

Release date: 2026-09-09

Summary

The branding and README refresh. This patch adds the project icon and banner to the Comfy Registry metadata and the README, and brings the README description up to date with the current prompt counts (239 user templates, 11 system-prompt variants). No runtime behavior changed. The main feature work shipped in v2.1.0 (see RELEASE_NOTES_v2.1.0.md) — the structured system prompt, the reworked MiniMaxOutputPaths layout and the eleven bundled system-prompt variants.

Added

• Branding assets: assets/branding/icon.png (400×400) and assets/branding/banner.png (1680×720), referenced from pyproject.toml as [tool.comfy] Icon / Banner (raw GitHub URLs on main).
• README banner: the banner is now shown at the very top of README.md (centered, full-width).

Changed

• README refreshed: the production system-prompt bullet now mentions the 11 bundled focus variants, and the bundled genre prompt library bullet states the exact 239 templates (was the rounded "230+"). The rest of the description was kept current with the released feature set.

Fixed

• None.

Breaking changes

• None.

Upgrade notes

• None beyond a normal restart + browser hard-refresh (Ctrl+F5). If v2.1.0 was not published separately, this v2.1.1 release is the one to publish and includes all v2.1.0 changes as well.

Assets

• ComfyUI-MiniMax-Music-Production-Toolkit-v2.1.1.zip
• MiniMax_Music3_Production_Toolkit_v2.1.1.json
• SHA256SUMS.txt