⚡ MiniMax Music 3 Turbo FP8 & 8-Step Distillation (ComfyUI Optimized)
🎉 Production Release (Master 8-Step Distillation):
- ✅ 8-Step Turbo LoRA (
minimax_music3_turbo_lora_8step.safetensors): Fully distilled over 5,000 real optimization steps through all 36 DiT layers using true Teacher-Student Euler trajectory matching. All 144 adapter matrices (94.3M active parameters) are trained and calibrated for fast, high-fidelity 8-step sampling at Strength 1.00.- ✅ Pure FP8 Models (
minimax_music3_text_encoder_fp8_e4m3fn.safetensors&minimax_music3_dit_fp8_e4m3fn.safetensors): Pure unrotated FP8 models delivering the primary ~3.8x generation speedup (cutting full song generation from ~15 minutes down to 233 seconds).
🎧 Audio Showcase Comparisons (A/B Test)
Listen to full audio samples generated with the exact same seed:
| Version | Audio Player / File | Description |
|---|---|---|
| 🚀 With Turbo LoRA (8 Steps, Strength 1.00) | 🔊 With_LoRA_8Steps.mp3 | Crisp vocals, full dynamic range, generated in only 8 steps. |
| Standard Baseline (Without LoRA) | 🔊 Without_LoRA.mp3 | Baseline unaccelerated generation. |
🚀 Benchmark Highlights (RTX 4090)
| Configuration | Full Song Generation Time | Total Speedup | VRAM Footprint |
|---|---|---|---|
| Baseline (INT8 non-optimized) | ~895s (~15 minutes) | 1.0x (Base) | ~16 GB |
Text Encoder & DiT in FP8 (e4m3fn) |
~315s (~5 minutes 15s) | ⚡ 2.8x Faster | ~11 GB |
| 🔥 FULL STACK: FP8 + Master Turbo LoRA (8 steps) + SageAttention | ⚡ 233.76s (~3 min 53s) | 🚀 ~3.8x Faster (-11 min!) | ~11.5 GB |
🎛️ Recommended Workflow Configuration in ComfyUI
1. Optimal Node Stack & Settings
LoraLoaderModelOnly:lora_name:minimax_music3_turbo_lora_8step.safetensorsstrength_model:1.00(Verified ideal for 8-step sampling)
Patch Sage Attention KJ:sage_attention:autoallow_compile:true
KSampler:steps:8sampler_name:eulerscheduler:simplecfg:1.5 - 1.7
MiniMax Music3 Text Encode:clip:minimax_music3_text_encoder_fp8_e4m3fn.safetensorscfg_scale:1.5
📦 Verified Production Weights
| File Name | Size | Format | ComfyUI Target Folder | Status |
|---|---|---|---|---|
minimax_music3_text_encoder_fp8_e4m3fn.safetensors |
8.48 GB | Pure FP8 e4m3fn (from clean BF16 source) |
ComfyUI/models/text_encoders/ |
✅ Verified & Production Ready |
minimax_music3_dit_fp8_e4m3fn.safetensors |
2.46 GB | Pure FP8 e4m3fn DiT Flow-Matching |
ComfyUI/models/diffusion_models/ |
✅ Verified & Production Ready |
minimax_music3_turbo_lora_8step.safetensors |
180 MB | Master LoRA (5,000 Real Distillation Steps, Rank 64) | ComfyUI/models/loras/ |
✅ Master Distillation Complete |
🤝 Open Source & Code Repository
Conversion scripts and the complete teacher-student distillation training engine are available on GitHub:
👉 GitHub: Guillaume-127/Minimax-music-3-Turbo-8-steps
Created by Guillaume-127. Released for the open-source AI music community.