Music Production Toolkit
A locally deployable, open-source AI music production toolbox built on ComfyUI. It integrates multiple AI music models including YuE2 and MiniMax Music, and supports functions such as song cover generation, audio restoration, and mastering. Released under the MIT license.
Timeline (last 90 days)
Music Production Toolkit released version 3.1, introducing an "Interpretation Freedom" slider ranging from 0 to 100 for the YuE2 cover song feature.
All computations in Music Production Toolkit are performed locally within ComfyUI, requiring no cloud service relay or account, and it is open-sourced under the MIT license.
The built-in SheetSage2 first extracts actual musical score information from the audio, including melody, chords, phrases, and segment structure.
At the 0 end of the Interpretation Freedom slider, if the model attempts to rewrite a note, chord, measure, or tempo, the response is directly rejected, and a complete cover at any freedom level will retain its harmonic structure.
Style templates do not weaken as freedom increases, and lyrics never change in response to the Interpretation Freedom slider.
Pure instrumental is the default mode, where vocal notes become rests and the melodic line is transferred to an instrumental part played by the user-selected lead instrument.
Because YuE2 sometimes introduces vocal-like content into scores without vocal notes, the toolkit offers optional "Whisper verification," which transcribes the rendered result and re-renders within a limited number of attempts, retaining the least vocal-sounding version when a clean output cannot be produced.
The toolkit has been officially renamed Music Production Toolkit and is no longer called "MiniMax …", because MiniMax is just one of the many models it drives.
The cover song path is the most demanding part of the entire toolkit, with lengthy and tightly structured prompts, and local small models with only a few billion parameters or heavily quantized ones may produce truncated responses, fabricated notation, or ignored constraints.
The toolkit adopts a 'degrade rather than crash' mechanism, where bad responses are rejected and fall back to deterministic backup scores, and unverified content does not enter the rendering engine.
1 more timeline events
All Facts (11)
The cover song path is the most demanding part of the entire toolkit, with lengthy and tightly structured prompts, and local small models with only a few billion parameters or heavily quantized ones may produce truncated responses, fabricated notation, or ignored constraints.
50%UnverifiedThe toolkit adopts a 'degrade rather than crash' mechanism, where bad responses are rejected and fall back to deterministic backup scores, and unverified content does not enter the rendering engine.
50%UnverifiedThe built-in SheetSage2 first extracts actual musical score information from the audio, including melody, chords, phrases, and segment structure.
50%UnverifiedAt the 0 end of the Interpretation Freedom slider, if the model attempts to rewrite a note, chord, measure, or tempo, the response is directly rejected, and a complete cover at any freedom level will retain its harmonic structure.
50%UnverifiedStyle templates do not weaken as freedom increases, and lyrics never change in response to the Interpretation Freedom slider.
50%UnverifiedPure instrumental is the default mode, where vocal notes become rests and the melodic line is transferred to an instrumental part played by the user-selected lead instrument.
50%UnverifiedBecause YuE2 sometimes introduces vocal-like content into scores without vocal notes, the toolkit offers optional "Whisper verification," which transcribes the rendered result and re-renders within a limited number of attempts, retaining the least vocal-sounding version when a clean output cannot be produced.
50%UnverifiedAll computations in Music Production Toolkit are performed locally within ComfyUI, requiring no cloud service relay or account, and it is open-sourced under the MIT license.
50%UnverifiedVersion 3.1 includes two workflows, Music_Production_Toolkit.json and Music_Production_AudioEnhance.json, and fixes a bug where original-lyrics covers were interrupted before the first note.
50%UnverifiedThe toolkit has been officially renamed Music Production Toolkit and is no longer called "MiniMax …", because MiniMax is just one of the many models it drives.
50%UnverifiedMusic Production Toolkit released version 3.1, introducing an "Interpretation Freedom" slider ranging from 0 to 100 for the YuE2 cover song feature.
50%