ntc-ai/minimax-music3-particle-sliders

MiniMax Music 3 particle sliders

Try the MiniMax Music 3 Space · YuE2 slider weights · YuE2 Space

NEW — September 16, 2026

16 newly retrained voice and genre LoRAs, with new matching samples and ComfyUI exports. The downloads below are the selected checkpoints from the new four-prompt training runs.

New voice and genre controls · Selected checkpoints · Native and ComfyUI downloads

Steer the lead toward a female vocal, a piano-led arrangement toward heavy riffs, or a small band toward a club pulse. These LoRA adapters change the language model that plans music in MiniMax Music 3.

September 16, 2026: all sixteen retrains are complete. This release uses the checkpoints selected by the final per-slider audit: steps 1,000, 2,000, 3,000 or 3,400 depending on the control. All voice and genre downloads and On samples below refer to those selected checkpoints.

Choose a control · ComfyUI · Native loading · Selection criteria · Training · How it works + math

Press play

Off uses the base model. On adds the named adapter at strength +1. The prompt, lyrics and seed match within each pair; composition and phrasing can still change. Recordings retain their original levels and lengths.

Female

Same prompt · same lyrics · seed 1709 · selected step 1000

OFF · 0	ON · +1

Off WAV · On WAV · Target-caption reference

Metal

Same prompt · same lyrics · seed 1709 · selected step 2000

OFF · 0	ON · +1

Off WAV · On WAV · Target-caption reference

House

Same prompt · same lyrics · seed 1709 · selected step 1000

OFF · 0	ON · +1

Off WAV · On WAV · Target-caption reference

Disco Funk

Same prompt · same lyrics · seed 1709 · selected step 2000

OFF · 0	ON · +1

Off WAV · On WAV · Target-caption reference

Pop Punk

Same prompt · same lyrics · seed 1709 · selected step 3400

OFF · 0	ON · +1

Off WAV · On WAV · Target-caption reference

Lo-fi

Same prompt · same lyrics · seed 1709 · selected step 3000

OFF · 0	ON · +1

Off WAV · On WAV · Target-caption reference

Hear all 64 Off/On comparisons. Every selected checkpoint has two arrangements × two seeds. Each featured example uses the first fixed arrangement and seed; examples were not chosen by their individual scores. Target-caption references are also included.

Choose a control

All sixteen downloads in this table are NEW as of September 16, 2026. Each file adds one target sound. All are language-model attention adapters with rank 8 and alpha 8. Female and Male are separate controls; negative strength is not a trained opposite.

New control	Target sound	Selected step	Native	ComfyUI	New samples
Female	Encourages one adult female lead vocal	1,000	Weights · Settings	Weights	4 pairs
Lo-fi	Soft swung drums, mellow keys and gentle tape warmth	3,000	Weights · Settings	Weights	4 pairs
Male	Encourages one adult male lead vocal	3,400	Weights · Settings	Weights	4 pairs
Pop	Clear hooks, crisp drums and a polished chorus	2,000	Weights · Settings	Weights	4 pairs
Hip-Hop	Rapped verses, deep sub bass and nimble hats	3,000	Weights · Settings	Weights	4 pairs
R&B	Warm keys, deep pocket and fluid vocal phrasing	3,400	Weights · Settings	Weights	4 pairs
Indie Rock	Chiming guitars, moving bass and a human drum kit	1,000	Weights · Settings	Weights	4 pairs
Pop Punk	Palm-muted power chords and driving chorus drums	3,400	Weights · Settings	Weights	4 pairs
Metal	Heavy guitar riffs, tight kicks and big melodic choruses	2,000	Weights · Settings	Weights	4 pairs
Country	Acoustic strum, twangy fills and an easy backbeat	1,000	Weights · Settings	Weights	4 pairs
Acoustic Folk	Fingerpicked strings and a warm small-room performance	2,000	Weights · Settings	Weights	4 pairs
House	Steady club kick, offbeat hats and a rolling bass line	1,000	Weights · Settings	Weights	4 pairs
Disco Funk	Elastic bass, clipped guitar and bright dance-floor strings	2,000	Weights · Settings	Weights	4 pairs
K-pop	Sharp synth hooks, tight edits and a big chorus lift	2,000	Weights · Settings	Weights	4 pairs
Reggaeton	Dembow drums, rounded sub bass and clipped melodic hooks	1,000	Weights · Settings	Weights	4 pairs
Afrobeats	Interlocking percussion, melodic bass and buoyant guitar	1,000	Weights · Settings	Weights	4 pairs

Machine-readable catalog · Selection scores and checkpoint comparisons · Release file hashes

ComfyUI

1. Download the ComfyUI file for your chosen control and place it in ComfyUI/models/loras/.
2. Use Load LoRA with your Music 3 text encoder connected to CLIP.
3. Set strength_model = 0 and strength_clip = 1 for the published On treatment. Compare with CLIP strength 0; start with one adapter.

Use a Music 3 text encoder with separate q_proj, k_proj, v_proj and o_proj layers. Text encoders with merged qkv_proj layers cannot apply the separate query/key/value factors in these files. See the ComfyUI loading guide.

All sixteen ComfyUI files were produced by the upstream conversion script from the selected native exports. Factors are BF16 and alpha scalars are FP32. The published recordings use the native adapters; ComfyUI precision and generation settings can change the resulting audio.

Native loading

Download the native weight and its matching JSON settings file. Use the included loading example with your existing Music 3 pipeline.

from huggingface_hub import snapshot_download

folder = snapshot_download(
    "ntc-ai/minimax-music3-particle-sliders",
    allow_patterns=[
        "catalog.json", "usage.md", "source/lora.py",
        "weights/uni16-fresh-selected-v2/female/*",
        "prompts/uni16-fresh-selected-v2/*",
    ],
)

Replace female with a control ID, or use weights/uni16-fresh-selected-v2/*/* for all native files. For ComfyUI only, download comfyui/uni16-fresh-selected-v2/*.

Start at direct multiplier 1 and compare with 0 using the same caption, lyrics and seed. The studio normalizes slider mixtures by host energy, so its fader position may differ from the direct multiplier used here.

How these checkpoints were chosen

The final audit scored 80 checkpoint candidates, with four matched clips per candidate. It excluded checkpoints with technical flags, then evaluated Content Enjoyment (CE) and Production Quality (PQ) separately using Audiobox Aesthetics.

For each control, a checkpoint qualifies when its mean CE is within 0.2 points of the best clean CE and its mean PQ is within 0.2 points of the best clean PQ. Among qualifying checkpoints, the latest checkpoint from the new training run is preferred. The two best values can come from different checkpoints. This is the quality-later-v2 policy.

Description similarity has no influence on selection. CLAP measurements are kept as a separate diagnostic. Cached lyric checks are informational only. There is no combined style/quality score in this selection rule.

The 0.2-point tolerance is a preference rule, not a calibrated audible threshold. The selection report shows how picks change at 0.1, 0.2 and 0.3 points. Four short clips support a shortlist; they do not establish universal quality, broad musical diversity, full-song reliability or behavior when stacking adapters.

All sixteen selected native exports contain finite tensors and match their saved full training states. Original-waveform checks cover silence, clipping, shortening, high-frequency excess, stereo issues, interior gaps and repeated or near-identical audio. These checks do not rule out subtle musical defects.

Training

Each slider uses all four existing prompt pairs. A fresh rank-8 adapter first trains for 600 updates with all four rows per batch. Further updates use one row at a time in balanced shuffled passes, with a fresh frozen-base continuation on every update. Continuations supply ending supervision; the prompt-state style targets remain fixed per row.

Eight runs reached 3,400 updates, and the other eight stopped at the approved 2,000-update budget. The published checkpoints were selected afterward. Generator and critic learning rates stay at 0.0005 and 0.00075, with adversarial, feature-matching and ending terms each weighted 1. Milestone samples retain their first draws without seed retries.

Method and equations · Training record · Checkpoint integrity · Conversion record

How the sliders learn

The new September 16 voice and genre adapters use LoRA to change the language model that plans the music.

One strength control

For an adapted weight matrix, the slider applies:

$$W(s)=W_0+s\frac{\alpha }{r}BA\mathrm{.}W(s)\; =\; W\_0\; +\; s\backslash ,\backslash frac\{\backslash alpha\}\{r\}BA.$$

Here $W_{0}W\_0$ is the frozen base weight, $AA$ and $BB$ are the learned low-rank factors, and $ss$ is the strength. The new exports use $r=8r=8$ and $\alpha =8\backslash alpha=8$, so $s=0s=0$ switches the adapter off and $s=1s=1$ applies its exported change. Intermediate strengths scale the weight change smoothly; the resulting music need not change linearly. Negative strength is not a trained opposite sound.

Voice and genre: learn the change between descriptions

Each of the sixteen new controls learns from four caption pairs. One caption describes a starting arrangement and the other adds the requested voice or genre. Lyrics, tempo and section tags stay fixed within each pair.

1. Build a target change. The frozen language model encodes the starting and target captions. We compare corresponding lyric-token states and the audio-start state. Let $H_{0}H\_0$ be the starting-caption sequence, $H_{+}H\_+$ the target-caption sequence, and $H_{\theta }H\_\backslash theta$ the adapted model's sequence on the starting caption:

$$x_{+}=\frac{H_{+}-H_0}{\sigma },x_{\theta }=\frac{H_{\theta }-H_0}{\sigma }\mathrm{.}x\_+\; =\; \backslash frac\{H\_+\; -\; H\_0\}\{\backslash sigma\},\; \backslash qquad\; x\_\backslash theta\; =\; \backslash frac\{H\_\backslash theta\; -\; H\_0\}\{\backslash sigma\}.$$

The fixed scale $\sigma \backslash sigma$ is calibrated from the teacher changes using root mean square (RMS). It puts the target change $x_{+}x\_+$ and the adapter's change $x_{\theta }x\_\backslash theta$ in the same units. Padding is excluded.

2. Train a critic to recognize that change. A small two-layer transformer $DD$ learns to score target changes above adapted changes. It combines the mean of valid token features with the final audio-start feature. Its relativistic pairing objective is:

$${\mathcal{L}}_D=\mathbb{E}\text{\hspace{0.17em}⁣}[\mathrm{softplus}(D(x_{\theta })-D(x_{+})\left)\right]+{\mathcal{R}}_{\mathrm{c}\mathrm{a}\mathrm{p}}\mathrm{.}\backslash mathcal\; L\_D\; =\; \backslash mathbb\; E\backslash !\backslash left[\backslash operatorname\{softplus\}\backslash big(D(x\_\backslash theta)-D(x\_+)\backslash big)\backslash right]\; +\; \backslash mathcal\; R\_\{\backslash mathrm\{cap\}\}.$$

Here $\mathbb{E}\backslash mathbb\; E$ denotes a batch average, and $\mathrm{softplus}(u)=\log (1+e^u)\backslash operatorname\{softplus\}(u)=\backslash log(1+e^u)$ penalizes the wrong ordering smoothly. The critic is used during training only.

3. Train the adapter for the target change, its features and stopping behavior. The adapter minimizes three terms, each with coefficient 1:

Term	What it teaches
Adversarial comparison	Make the adapted change resemble the target-caption change.
Feature matching	Match the batch means of the critic's learned features $\varphi \backslash phi$. MSE is mean squared error over feature coordinates.
Ending supervision	Preserve the base model's balance between continuing and ending on a supplied token history.

Warm-up uses four rows per batch. Fresh-continuation training uses one row per update, so feature matching then compares that update's student and teacher features. The adapter has no separate explicit lyric-hold term.

4. Preserve the stopping margin on fresh histories. Let $\mathrm{\ell }\backslash ell$ denote token logits and $\mathcal{S}\backslash mathcal\; S$ the semantic audio-token vocabulary. On the same supplied history, compare the audio-end logit with the total semantic-continuation mass:

The first 600 updates use fixed base histories. From update 601 onward, each update samples a fresh frozen-base continuation with the adapter disabled, cycling through all four training rows in balanced shuffled passes. Seeds do not repeat; duplicate continuation tensors are rejected without substituting a new seed. These histories diversify ending supervision. The prompt-state style targets remain fixed per row.

The ending term preserves a stopping decision on the supplied histories. It does not establish that every newly generated full song will end naturally.

5. Keep the critic's gradients bounded. In the calibrated input coordinates, the cap penalizes gradient norms only above 1:

$${\mathcal{R}}_{\mathrm{c}\mathrm{a}\mathrm{p}}=\frac{1}{2}\sum _{z\in \{x_{+},x_{\theta }\}}\mathbb{E}\text{\hspace{0.17em}⁣}[\max \text{\hspace{0.17em}⁣}{(0,\parallel {\mathrm{\nabla }}_{z}D(z){\parallel }_2-1)}^2]\mathrm{.}\backslash mathcal\; R\_\{\backslash mathrm\{cap\}\}\; =\; \backslash frac\{1\}\{2\}\backslash sum\_\{z\backslash in\backslash \{x\_+,x\_\backslash theta\backslash \}\}\; \backslash mathbb\; E\backslash !\backslash left[\; \backslash max\backslash !\backslash left(0,\backslash lVert\backslash nabla\_zD(z)\backslash rVert\_2-1\backslash right)^2\; \backslash right].$$

The cap coefficient and threshold are both 1. Small gradients are allowed, and the initially zero adapter change remains included.

Training setting	New September 16 release
Adapter location	Language-model attention; 36 layers × four projections
Starting point	A fresh zero-output rank-8, alpha-8 adapter and critic per control
Warm-up	600 updates; all four prompt pairs per batch
Further training	One row per update; balanced shuffled passes; one fresh base continuation per update
Learning rates	Adapter 0.0005; critic 0.00075; constant
Continuation step bound	Adapter parameter-update L2 norm at most 2
Training budgets	Eight runs to 3,400; eight runs to 2,000
Published checkpoints	Selected per control at 1,000, 2,000, 3,000 or 3,400
Published samples	64 Off/On pairs: two arrangements × two seeds for each of all sixteen controls

The critic, optimizers, sampler and RNG state carry across milestones. Every selected native export was checked tensor-for-tensor against its saved full state. The full training record gives the recipe and budgets.

Checkpoint selection: quality bounds, then later steps

Selection happens after training. For one slider, let $\mathcal{C}\backslash mathcal\; C$ contain checkpoints that pass the technical screen, with four-clip mean enjoyment $E_{c}E\_c$, mean production quality $P_{c}P\_c$, and step $t_{c}t\_c$. The quality anchors and eligible set are:

Among qualifying checkpoints from the new run, choose:

$$c_{\mathrm{r}\mathrm{e}\mathrm{l}\mathrm{e}\mathrm{a}\mathrm{s}\mathrm{e}}=\underset{c\in {\mathcal{Q}}_{\mathrm{n}\mathrm{e}\mathrm{w}}}{\mathrm{arg\hspace{0.17em}max}}{t}_c\mathrm{.}c\_\{\backslash mathrm\{release\}\}\; =\; \backslash operatorname*\{arg\backslash ,max\}\_\{c\backslash in\backslash mathcal\; Q\_\{\backslash mathrm\{new\}\}\}\; t\_c.$$

If no new-run checkpoint qualifies, the legacy release can be considered under the same quality bounds; an empty eligible set requires listening to resolve the tradeoff. All sixteen current selections are from the new runs. Description similarity and cached lyric checks have zero influence on this selection. The 0.2-point tolerance is a preference rule, not statistical equivalence or a proven audible boundary. Selection policy and sensitivity.

Earlier releases

The update-660 UNI16 release and older uni-lyric, v24 and bipolar weights remain at their existing paths. The root catalog and download table now point to the selected fresh-continuation release. Previous model card.