molbal/Minimax-Music3-GGUF

MiniMax Music 3 - GGUF Quantizations

This repository contains GGUF quantizations for MiniMax Music 3 (Hugging Face model), a music generation model capable of synthesizing complete, structurally coherent songs up to 5 minutes long at 32 kHz stereo.

For full architectural details, model specifications, and benchmarks, please refer to the official MiniMax-Music3 GitHub repository.

---

Example Workflow

An example ComfyUI workflow is included in the examples/ directory:

• examples/minimax-music-3.json

---

Quantized Models

Diffusion Transformer (Flow Matching 2.4B DiT)

Place these files in ComfyUI/models/diffusion_models/ or ComfyUI/models/unet/:

File	Quantization	Size	Description
minimax_music3_dit_BF16.gguf	BF16	~4.64 GB	Unquantized reference precision
minimax_music3_dit_Q8_0.gguf	Q8_0	~2.52 GB	Standard 8-bit quantization
minimax_music3_dit_Q8_CR.gguf	Q8_CR	~2.38 GB	8-bit channel-residual quantization
minimax_music3_dit_Q4_0.gguf	Q4_0	~1.39 GB	4-bit quantization for low VRAM

Pruned Text & Language Model Encoder

Place these files in ComfyUI/models/text_encoders/ or ComfyUI/models/clip/:

File	Quantization	Size	Description
minimax_music3_text_encoder_pruned_Q8_0.gguf	Q8_0	~8.93 GB	High precision 8-bit text/audio encoder
minimax_music3_text_encoder_pruned_Q8_CR.gguf	Q8_CR	~8.49 GB	8-bit channel-residual text/audio encoder
minimax_music3_text_encoder_pruned_Q4_0.gguf	Q4_0	~5.38 GB	4-bit text/audio encoder for lower memory

---

ComfyUI Node Requirements

To load and generate music with these GGUF weights in ComfyUI:

1. ComfyUI-GGUF:
   • Repository: molbal/ComfyUI-GGUF
   • Registry: comfyui-gguf-reboot
2. Load the workflow provided at examples/minimax-music-3.json.

---

Prompting Guide

MiniMax Music 3 accepts two primary conditioning inputs: Lyrics and a Music Description.

1. Lyrics & Section Tags

Structure your lyrics using standard uppercase section tags placed on their own lines:

[Intro]
(Soft acoustic guitar melody)

[Verse 1]
Morning light filtering through the pine
Every quiet street is yours and mine

[Pre-Chorus]
The horizon starts to glow

[Chorus]
Softly the world begins to breathe
Waking up under golden leaves

[Bridge]
Time moves slow in the morning haze

[Chorus]
Softly the world begins to breathe
Waking up under golden leaves

[Outro]
Fading into the dawn

Supported section tags include: [Intro], [Verse], [Pre-Chorus], [Chorus], [Post-Chorus], [Bridge], [Instrumental], [Solo], and [Outro].

2. Structured Caption (Music Description)

For precise control over the arrangement and performance, structure your music description across three areas:

• Global Metadata: Genre, subgenre, tempo (BPM), key, scale, emotional progression, listening scenario, and production style.
• Vocal Details: Vocal gender, timbre, performance delivery, harmony, backing vocals, and vocal effects (e.g. reverb, delay).
• Arrangement: Primary and secondary instruments, section-by-section instrument progression, groove/rhythm, bassline, percussion style, and spatial atmosphere.

Example Structured Caption:

Genre: Indie Folk Pop. BPM: 98. Key: G major. 
Mood: Warm, intimate, and uplifting with a gradual dynamic build.
Vocals: Close, breathy female lead vocals with light double-tracking and airy reverb; two-part harmonies enter during the chorus.
Arrangement: Fingerpicked acoustic guitar and soft felt piano throughout. Subdued kick and upright bass join in the second verse. Full folk percussion, tambourine, and subtle ambient strings enter during the chorus.

---

Limitations

• Generation is non-streaming.
• The tokenized text prompt is limited to 5,000 tokens.
• Audio generation is capped at 9,000 acoustic frames (~5 minutes).
• Section tags and descriptions provide generative conditioning; tempo, key, and lyrical timing may vary slightly.

---

Original Model & License

• Original weights & code: MiniMax-AI/MiniMax-Music3
• License: CreativeML Open RAIL-M