README.md

metadata

title: MLTNT Militant Roots Reggae (YuE2)
emoji: 🔥
colorFrom: blue
colorTo: green
sdk: gradio
sdk_version: 6.17.3
app_file: app.py
python_version: '3.12'
short_description: Militant roots reggae LoRAs for YuE2-3B
startup_duration_timeout: 1h
license: cc-by-nc-4.0
models:
  - becausereasons/yue2-mltnt-militant-reggae
  - m-a-p/YuE2-3B
  - m-a-p/YuE2-Vae
tags:
  - music-generation
  - text-to-music
  - reggae
  - lora
  - yue2

🔥 MLTNT — Militant Roots Reggae LoRAs for YuE2

Text + lyrics → a full militant roots reggae song, using the six MLTNT LoRAs by becausereasons on top of YuE2-3B.

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

	LoRA	Character	Recipe
💥	Soundclash	new-trainer generation; most expressive voice: hip-hop, dancehall, lovers rock	clip 0.5 / model 1.0 for fast styles, clip 1.0 for slow songs, 360 s cap
🦁	Chanter	new-trainer generation; settled signature voice	clip 1.0 or 0.5 / model 1.0, 360 s cap
🥁	Steppers	flagship; tight hard-hitting steppers, big anthemic chorus	clip 1.0 / model 1.0, cfg 1.0
🔥	Frontline	most distinctive voice of the earlier four, trained for rapid-fire lyrics	clip 1.0 / model 1.0, cfg 1.0
🎤	Fusion	reggae hip-hop, boom-bap over one-drop, deejay toasting	clip 1.0 / model 1.5, cfg 1.4
🌿	Roots	the purist; straightest roots timbre	clip 1.0 / model 1.0, cfg 1.0

How this Space runs the LoRAs

The MLTNT files were trained with the ComfyUI FS_Audio Suite against Comfy-Org's repack of YuE2, so they ship in Comfy's fused-key layout (text_encoders.* for the planner, diffusion_model.* for the decoder).

This Space runs the official yue2_infer pipeline from m-a-p instead of ComfyUI, and merges the LoRA into the native checkpoint:

• Comfy's fused self_attn.qkv_proj [4096, 2048] is split back into native q_proj (0:2048), k_proj (2048:3072) and v_proj (3072:4096).
• Comfy's fused mlp.gate_up_proj [12288, 2048] is split into gate_proj (0:6144) and up_proj (6144:12288).
• diffusion_model.* maps onto the native nar_self_attn / nar_mlp branch, text_encoders.* onto the AR branch. The full-rank llm2vae / vae2llm projection diffs are applied too.
• Merge math follows ComfyUI (these files carry no alpha key): W += strength · (up @ down).

Both key layouts were compared tensor-by-tensor before the mapping was written, and the corresponding slices are bit-identical.

The FS_Audio sampler's knobs are reproduced against the native pipeline:

FS_Audio widget	here
score_temperature	ABC planner sampling temperature (top-p 0.9, top-k 30, rep 1.005)
music_temperature	semantic codec sampling temperature (top-p 0.95, top-k 100)
song_length_cap	semantic max_tokens = cap × 25
Weirdness (cfg)	classifier-free guidance on the flow-matching decoder against a zeroed AR context — which is exactly what the Comfy graph does (it passes zeros_like(cond) as the negative)
steps	flow-matching ODE steps

Known deviation: ComfyUI solves the decoder ODE with dpm_2 / sgm_uniform; this Space uses the reference m-a-p 32-step midpoint solver that ships with yue2_infer. Both are second-order solvers over the same velocity field, but the renders will not be sample-identical to the demos on the model card.

Prompting

• Style prompt — one descriptive sentence, trigger first, in this order: language → genre → vocal → instruments → mood → BPM → production. Tag lists and [Intro]…[Outro] scaffolding in the style field produce "odd, not reggae" output.
• Lyrics — tagged blocks ([Verse], [Pre-Chorus], [Chorus], [Bridge], [Outro]), ~4 lines a verse at ~8 words. No empty [Intro] — the planner writes its own.
• Rapid-fire delivery comes from the lyric, not the prompt: write verse lines at 15–17 words with simple one- and two-syllable words. Frontline is the variant trained for it.
• Planner strength above 1.0 collapses the vocal — it is capped at 1.0 here on purpose.
• A seed can end a song early; that is the seed, not the length cap.

The five style prompts in the examples are the author's own, taken verbatim from prompts/. The lyric sheets are original, written to the shapes documented on the model card (a standard 8-word-line sheet and a dense 15–17-word-line sheet for the rapid-fire comparison).

Licence and credits

LoRA weights © becausereasons, released under CC BY-NC 4.0 — non-commercial use only; attribute "MLTNT LoRAs by becausereasons". The base model m-a-p/YuE2-3B carries the same licence.

• YuE2 — Multimodal Art Projection (m-a-p).
• ComfyUI-FS_Audio_Suite — KytraScript / The Fixed Seed Company: the inference nodes and artist trainer the LoRAs were trained with.
• LoRAs trained and documented by becausereasons (Ko-fi).