๐Ÿ”ฅ MLTNT โ€” Militant Roots Reggae LoRAs for YuE2

Six artist-style LoRAs by becausereasons that push YuE2-3B into modern militant roots reggae โ€” dark raspy patois vocals, steppers grooves, deep sub bass, bubbling Hammond, nyabinghi drums, horn stabs, dub sirens and spring reverb.

Each LoRA patches both halves of YuE2: the autoregressive planner that writes the score and the vocal lines, and the flow-matching decoder that makes the sound. Trigger word: mltnt.

New: Soundclash and Chanter, trained with a different trainer. The voice is far more expressive, and they need one habit: planner strength 0.5 for fast or hard styles, and a 360 s cap. Picking one of them sets both for you.

LoRA variant
0 2
1 3
0.1 1.5
0.1 1.5

Tips from the model card

  • Rapid-fire verses come from the lyric, not the prompt. Write verse lines at 15โ€“17 words (one- and two-syllable words) and keep the chorus at 7โ€“8; asking for "double-time" in the style prompt does not speed the vocal up. MLTNT Frontline is the variant trained for this.
  • Planner strength above 1.0 collapses the vocal โ€” it is capped here.
  • Soundclash and Chanter behave differently. At planner strength 1.0 on a fast style they over-commit: everything gets faster and more frantic, and occasionally the score gets stuck in one section (minutes of groove, no singing). Planner 0.5 keeps the voice and restores the structure. On these two files the phrase "rapid-fire double-time deejay flow" does speed the whole song up, so keep it for one section unless you want the entire track double time. The same prompt at 1.0 and 0.5 gives two different songs: try both. The one drop rhythm is not learned by any of the six files.
  • Production words ("boom-bap", "deejay toasting", "steppers") move the tempo grid and key far more than the numeric BPM in the same sentence.
  • A seed can end a song early. That is the seed, not the cap โ€” change it and re-run.
The author's own recipes

LoRA weights ยฉ becausereasons, CC BY-NC 4.0 (non-commercial; attribute "MLTNT LoRAs by becausereasons"). Base model m-a-p/YuE2-3B, same licence. The LoRAs were trained with ComfyUI-FS_Audio_Suite against Comfy's repack (Soundclash and Chanter with AI Toolkit); this Space merges them into the native yue2_infer pipeline and reproduces the FS_Audio graph's sampling recipe, including CFG on the flow-matching decoder.