ntc-ai/minimax-music3-concept-sliders

MiniMax Music 3 concept sliders

Unidirectional LoRA sliders for MiniMax Music 3. One scalar adds one musical property while the prompt stays fixed: 0 is off, +1 is one trained unit of the concept. Opposite poles are separate LoRAs (Loud and Quiet are two files, not ± on one file).

Current studio catalog is uni-lyric, language-model only. Each shipped file is weights/<name>-lm-uni-lyric/<name>-lm-uni-lyric_last.safetensors (unit-normalized, drop in at strength 1), except the four v24 lyric-hold refreshes (gender-lm-v24-lyrichold, distortion-lm-v24-uni, breath-lm-v24-uni, joy-lm-v24-uni). Space stays the transformer exception (space-tf-v6); dust stays dust-lm-v1-faithful-kl-pole03.

Earlier bipolar *-lm-v9 weights remain under weights/ for comparison.

Listen (uni-lyric)

Same prompt, same lyrics, same seed. Only the LoRA scale changes. Player clips below are uni-lyric at +1 and +2. Full ladders (0 / +1 / +2 plus caption REFs) live under samples/uni-lyric/.

Energy — Off → Loud

+1 loud

+2 loud

Quiet — Off → Quiet

+1 quiet

+2 quiet

Gender — Off → Female

+1 female

+2 female

Male — Off → Male

+1 male

+2 male

Distortion — Off → Distorted

+1 distorted

+2 distorted

Clean — Off → Clean

+1 clean

+2 clean

Joy — Off → Joy

+1 joy

+2 joy

Somber — Off → Somber

+1 somber

+2 somber

Training code: ntc-ai/sliders-conceptmod (conceptmod/textsliders/train_lm_slider_music3.py).

Objective (uni-lyric)

Plus-only LoRA on Qwen3Attention. The generation prompt is always the neutral caption. At LoRA scale s = +1 the prompt-last hidden h(s) is pulled toward the frozen encode of the + caption h+; at s = 0 it must reproduce the neutral encode h0. A lyric-token hold keeps the yaml lyrics span pinned to encode(neu):

$${\mathcal{L}}_{\mathrm{p}\mathrm{o}\mathrm{l}\mathrm{e}}=\mathrm{\parallel }h(+1)-h_{+}{\mathrm{\parallel }}^2+\mathrm{\parallel }h(0)-h_0{\mathrm{\parallel }}^2\backslash mathcal\{L\}\_\{\backslash mathrm\{pole\}\}\; =\; \backslash |h(+1)\; -\; h\_+\backslash |^2\; +\; \backslash |h(0)\; -\; h\_0\backslash |^2$$

$${\mathcal{L}}_{\mathrm{l}\mathrm{y}\mathrm{r}\mathrm{i}\mathrm{c}}=\mathrm{\parallel }{h}_{\mathcal{L}}(+1)-h_{0,\mathcal{L}}{\mathrm{\parallel }}^2\backslash mathcal\{L\}\_\{\backslash mathrm\{lyric\}\}\; =\; \backslash |h\_\{\backslash mathcal\{L\}\}(+1)\; -\; h\_\{0,\backslash mathcal\{L\}\}\backslash |^2$$

$$\mathcal{L}={\mathcal{L}}_{\mathrm{p}\mathrm{o}\mathrm{l}\mathrm{e}}+{\mathcal{L}}_{\mathrm{l}\mathrm{y}\mathrm{r}\mathrm{i}\mathrm{c}}+{\lambda }_{\mathrm{e}\mathrm{n}\mathrm{d}}{\mathcal{L}}_{\mathrm{e}\mathrm{n}\mathrm{d}}\backslash mathcal\{L\}\; =\; \backslash mathcal\{L\}\_\{\backslash mathrm\{pole\}\}\; +\; \backslash mathcal\{L\}\_\{\backslash mathrm\{lyric\}\}\; +\; \backslash lambda\_\{\backslash mathrm\{end\}\}\backslash ,\backslash mathcal\{L\}\_\{\backslash mathrm\{end\}\}$$

Teacher is raw h+ at +1 (last token), h0 at scale 0, and encode(neu) on the lyrics span only (subscript L). Vocal Details / Global Metadata / Arrangement are not held. No minus pole, no leftover-gate, no pair-odd. Infer with the neutral caption + LoRA — not the + caption.

The end term teacher-forces the LoRA'd LM over a frozen base-model composition and penalizes drift of the stop margin:

$$m=\mathrm{l}\mathrm{o}\mathrm{g}\mathrm{i}\mathrm{t}(⟨\mathrm{a}\mathrm{u}\mathrm{d}\mathrm{i}\mathrm{o}\mathrm{\_}\mathrm{e}\mathrm{n}\mathrm{d}⟩)-\mathrm{l}\mathrm{o}\mathrm{g}\mathrm{s}\mathrm{u}\mathrm{m}\mathrm{e}\mathrm{x}\mathrm{p}(\text{semantic band}),{\mathcal{L}}_{\mathrm{e}\mathrm{n}\mathrm{d}}=\mathrm{\parallel }m(+1)-m_{\mathrm{b}\mathrm{a}\mathrm{s}\mathrm{e}}{\mathrm{\parallel }}^{2}m\; =\; \backslash mathrm\{logit\}(\backslash langle\backslash mathrm\{audio\backslash \_end\}\backslash rangle)\; -\; \backslash mathrm\{logsumexp\}(\backslash text\{semantic\; band\}),\; \backslash qquad\; \backslash mathcal\{L\}\_\{\backslash mathrm\{end\}\}\; =\; \backslash |m(+1)\; -\; m\_\{\backslash mathrm\{base\}\}\backslash |^2$$

The slider may move the musical plan; it must not move the stop decision.

Flags:

--lm_target faithful_plus_neu_lyric --pole_mode hidden --endreg_weight 1.0

Opposite concepts ship as separate weights (e.g. energy-lm-uni-lyric and energy-quiet-lm-uni-lyric). There is no trained −1 on a plus pole.

Catalog

id	concept	weights
gender	Female	weights/gender-lm-v24-lyrichold/
male	Male	weights/gender-male-lm-uni-lyric/
energy	Loud	weights/energy-lm-uni-lyric/
quiet	Quiet	weights/energy-quiet-lm-uni-lyric/
distortion	Distorted	weights/distortion-lm-v24-uni/
clean	Clean	weights/distortion-clean-lm-uni-lyric/
tempo	Fast	weights/tempo-lm-uni-lyric/
slow	Slow	weights/tempo-slow-lm-uni-lyric/
live	Live	weights/live-lm-uni-lyric/
studio	Studio	weights/live-studio-lm-uni-lyric/
breath	Breathy	weights/breath-lm-v24-uni/
rapslow	Rap	weights/rapslow-lm-uni-lyric/
triphop	Trip-hop	weights/triphop-lm-uni-lyric/
pop	Pop	weights/triphop-pop-lm-uni-lyric/
rhyme	Rhyme	weights/rhyme-lm-uni-lyric/
prose	Prose	weights/rhyme-prose-lm-uni-lyric/
sexy	Sexy	weights/sexy-lm-uni-lyric/
plain	Plain	weights/sexy-plain-lm-uni-lyric/
tender	Tender	weights/tender-lm-uni-lyric/
fierce	Fierce	weights/tender-fierce-lm-uni-lyric/
grit	Grit	weights/grit-lm-uni-lyric/
smooth	Smooth	weights/grit-smooth-lm-uni-lyric/
joy	Joy	weights/joy-lm-v24-uni/
somber	Somber	weights/joy-somber-lm-uni-lyric/
yearn	Yearning	weights/yearn-lm-uni-lyric/
settled	Settled	weights/yearn-settled-lm-uni-lyric/
hurt	Hurt	weights/hurt-lm-uni-lyric/
numb	Numb	weights/hurt-numb-lm-uni-lyric/
dust	Tape	weights/dust-lm-v1-faithful-kl-pole03/
space	Wet	weights/space-tf-v6/ (transformer)

Play samples/uni-lyric/<name>-lm-uni-lyric/ in filename order (0, +1, +2, then caption REFs).

ComfyUI

Each shipped file has a sibling *_comfyui.safetensors (LM files use text_encoders.model.layers.*.self_attn.*). Drop the Comfy file in ComfyUI/models/loras/ and load it with Load LoRA. Strength is the slider scale (0 off, +1 the baked unit).

ComfyUI support is untested.

Which stage a slider attaches to

MiniMax Music 3 generates in two stages:

1. A Qwen3 language model writes a plan — who is singing, melody, rhythm, arrangement.
2. A flow transformer renders that plan — timbre, loudness, tone, space.

uni-lyric attaches to stage 1. Space is the catalog exception on stage 2 (reverb is a rendering property).