⚡ MiniMax Music 3 Turbo FP8 & 8-Step LoRA (ComfyUI Optimized)

⚠️ Notice (v1.0 Community Release - Work in Progress / Experimental Calibration):
This is a v1.0 calibration release created to accelerate local generation on NVIDIA RTX GPUs. Further refinements and community feedback are welcome as we calibrate audio hyperparameters.

🚀 Overview & Benchmark Highlights

This repository contains the ultra-optimized FP8 (torch.float8_e4m3fn) weights and the Turbo 8-Step Step-Distillation LoRA for MiniMax Music 3, tailored specifically for lightning-fast inference in ComfyUI.

Configuration	Full 190s Song Generation Time (RTX 4090)	Speedup	VRAM Footprint
Baseline (INT8 non-optimized)	~895s (~15 minutes)	1.0x	~16 GB
Text Encoder & DiT FP8	~315s (~5 minutes 15s)	2.8x faster	~11 GB
🔥 FULL STACK: FP8 + Turbo LoRA 8-Step + SageAttention	⚡ ~245s (~4 minutes 05s)	🚀 ~3.7x Faster (-11 min!)	~11.5 GB

📦 Downloadable Files in this Repo

File Name	Size	Type	Destination Folder in ComfyUI
`minimax_music3_text_encoder_fp8_e4m3fn.safetensors`	8.48 GB	Pure FP8 Text Encoder (Clean, unrotated BF16 base)	`ComfyUI/models/text_encoders/`
`minimax_music3_dit_fp8_e4m3fn.safetensors`	2.46 GB	Pure FP8 DiT Flow-Matching Model	`ComfyUI/models/diffusion_models/`
`minimax_music3_turbo_lora_8step.safetensors`	180 MB	8-Step Consistency Distillation LoRA (Rank 64 / Alpha 64)	`ComfyUI/models/loras/`

🎛️ Recommended ComfyUI Workflow & Settings

1. Model Connections

[Load Diffusion Model] (minimax_music3_dit_fp8_e4m3fn.safetensors)
       │
       ▼
[Load LoRA] (minimax_music3_turbo_lora_8step.safetensors | strength: 0.85 ⭐ IDEAL)
       │
       ▼
[Patch Sage Attention KJ] (sage_attention: auto)
       │
       ▼
[KSampler]

💡 Pro-Tip on LoRA Strength (0.85 is Ideal)

⭐ LoRA Strength = 0.85 (Recommended Sweet Spot):
Setting the LoRA strength to 0.85 (rather than 1.0) provides the absolute best balance: it accelerates the diffusion sampling to 8-10 steps while letting the base model inject 100% crystal-clear vocal formants, diction, and delicate high-frequency percussion.

2. Optimal Node Parameters

A. `MiniMax Music3 Text Encode`

clip : Connect to Load Text Encoder with minimax_music3_text_encoder_fp8_e4m3fn.safetensors.
cfg_scale : 1.5 (Recommended: keeps 100% voice clarity, instrument separation and lyrical articulation).
max_duration : Adjust to your desired song length (e.g. 120.0 for 2 min, 190.0 for 3 min 10s).

B. `KSampler`

steps : 8 (Ultra-fast) or 10 - 12 (Studio Master Quality)
sampler_name : euler
scheduler : simple
cfg : 1.7
denoise : 1.0

🔬 Technical Details & Methodology

Text Encoder Quantization: Converted from the pristine unrotated minimax_music3_text_encoder_pruned_bf16.safetensors directly to float8_e4m3fn. All 166 sensitive layers (audio decoder heads, RMSNorms, token embeddings) remain in high-precision BF16 to guarantee zero early-stop anomalies and pitch accuracy.
DiT Flow-Matching Quantization: Converted 150 2D matrix weights to FP8 while keeping 224 modulation/norm parameters in FP16.
Turbo LoRA Distillation: Trained using Consistency Flow-Matching & Mean Flow Distillation across 144 attention and MLP projection matrices over 20,000 steps with Cosine Warmup scheduling on an NVIDIA RTX 4090.

🤝 GitHub Repository

Training scripts and conversion tools are open-sourced at:
👉 GitHub: Guillaume-127/Minimax-music-3-Turbo-8-steps

Created by Guillaume-127. Released for the open-source AI music community.

Downloads last month: -; Downloads are not tracked for this model. How to track