minimax-music3-training-tournament-round1

MiniMax Music 3 training tournament round 1

A central comparison of MiniMax Music 3 language-model LoRA tournament round 1 runs. The previous campaigns are preserved below, followed by the Caption v2 training-window and singer-tag comparisons. Tournament and prompt sections are collapsed so individual audio comparisons can be opened as needed.

The card stores 2,144 FLAC renders directly in this repository. Each comparison holds its lyrics, caption conditions, LoRA-strength sweep, and random stream constant within that tournament.

Tournament overview

Machine-readable comparison manifest

<details> <summary><strong>Previous tournament: original objective comparison</strong> (256 renders)</summary>

The original 1,000-step Vanilla, XM, NextLat, and NextLat + XM comparison.

All renders use a 40-second maximum duration, 30 flow steps, flow CFG 1.7, and seed 20260822. MiniMax Music 3 may end a song before the duration cap.

Training legs

The source model repos were consolidated into this repository; audio renders are stored below each comparison.

• Vanilla: soad-mm3-vanilla-20260822
• Explorative Modeling (XM): soad-mm3-xm-20260822
• NextLat: soad-mm3-nextlat-20260822
• NextLat + XM: soad-mm3-nextlat-xm-20260822

No trigger word / off-genre descriptors

Caption: Alternative pop-jazz, 102 BPM, brushed drums, upright bass, clean electric piano, airy male vocals, smooth verse-chorus songwriting, restrained dynamics, polished intimate mix.

<details> <summary><strong>LoRA strength 0.25</strong></summary> </details> <details> <summary><strong>LoRA strength 0.50</strong></summary> </details> <details> <summary><strong>LoRA strength 0.75</strong></summary> </details> <details> <summary><strong>LoRA strength 1.00</strong></summary> </details>

Trigger word only

Caption: system of a down, soad style music

<details> <summary><strong>LoRA strength 0.25</strong></summary> </details> <details> <summary><strong>LoRA strength 0.50</strong></summary> </details> <details> <summary><strong>LoRA strength 0.75</strong></summary> </details> <details> <summary><strong>LoRA strength 1.00</strong></summary> </details>

Matching descriptors, no trigger word

Caption: Lilting hypnotic arpeggios, a gentle melodic mid-tempo groove with layered harmonies, sudden explosive distorted riffs, unstable meter changes, theatrical male vocals, tightly syncopated bass and drums, abrupt shifts between fragile verses and frantic heavy choruses.

<details> <summary><strong>LoRA strength 0.25</strong></summary> </details> <details> <summary><strong>LoRA strength 0.50</strong></summary> </details> <details> <summary><strong>LoRA strength 0.75</strong></summary> </details> <details> <summary><strong>LoRA strength 1.00</strong></summary> </details>

Trigger word + matching descriptors

Caption: system of a down, soad style music, lilting hypnotic arpeggios, a gentle melodic mid-tempo groove with layered harmonies, sudden explosive distorted riffs, unstable meter changes, theatrical male vocals, tightly syncopated bass and drums, abrupt shifts between fragile verses and frantic heavy choruses.

<details> <summary><strong>LoRA strength 0.25</strong></summary> </details> <details> <summary><strong>LoRA strength 0.50</strong></summary> </details> <details> <summary><strong>LoRA strength 0.75</strong></summary> </details> <details> <summary><strong>LoRA strength 1.00</strong></summary> </details> </details> <details> <summary><strong>Full song without regularisation</strong> (576 renders)</summary>

AdamW 8e-5, full-song training windows, no caption dropout, and no regularisation set.

All renders use a 40-second maximum duration, 30 flow steps, flow CFG 1.7, and seed 20260822. MiniMax Music 3 may end a song before the duration cap.

Training legs

• Vanilla: soad-mm3-vanilla-20260823-10k-adamw8e-5-fullsong-nocapdrop
• Explorative Modeling (XM): soad-mm3-xm-20260823-10k-adamw8e-5-fullsong-nocapdrop
• NextLat: soad-mm3-nextlat-20260823-10k-adamw8e-5-fullsong-nocapdrop
• NextLat + XM: soad-mm3-nextlat-xm-20260823-10k-adamw8e-5-fullsong-nocapdrop

Run note: These four runs resumed their learned LoRA weights from checkpoint 500 with fresh optimizer, sampler, and RNG state after the original local trainer state was lost.

No trigger word / off-genre descriptors

Caption: Alternative pop-jazz, 102 BPM, brushed drums, upright bass, clean electric piano, airy male vocals, smooth verse-chorus songwriting, restrained dynamics, polished intimate mix.

<details> <summary><strong>LoRA strength 0.25</strong></summary> </details> <details> <summary><strong>LoRA strength 0.50</strong></summary> </details> <details> <summary><strong>LoRA strength 0.75</strong></summary> </details> <details> <summary><strong>LoRA strength 1.00</strong></summary>

The model card is truncated. Read the rest on Hugging Face.