🌌 QTSTRM — Quiet Storm R&B LoRAs for YuE2

Five artist-style LoRAs by becausereasons that push YuE2-3B into 1980s–1990s quiet storm: slow, late-night R&B and sophisti-soul with a sweet, sultry female lead — deep pulsing bass, electric piano, synth pads, slow drum machines, and on the sophisti-pop side a tenor saxophone, fretless bass and congas. English lyrics, female lead.

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: qtstrm.

Start with clip 1.0 / model 1.0 and score mode full. The five examples are the author's own published recipes — same LoRA, prompt, lyric and length cap as each render on the model card.

LoRA variant
0 2
1 3
0.1 1.5
0.1 1.5

Tips from the model card

  • File first. v2 (Velvet, Midnight) gives the cleanest, most consistent results; Afterhours has the strongest voice character; Candlelight the most variety.
  • Seeds decide the form as much as the lyric. If a render comes out garbled or ghostly it usually wrote an over-long score — change the seed. Good takes here have ~95–125 score lines.
  • Very short lyrics (under ~150 words) make the planner pad the song and can come out garbled; ~180–270 words is the sweet spot.
  • Afterhours can grit on long sustained high notes — lower decoder strength to ~0.8 or use a v2 file.
  • Steer the voice with the caption. The training captions used a fixed vocabulary: register (low/deep contralto, warm mezzo, light soprano), texture (smoky, breathy, husky, silky), character (sweet, sultry, soulful, intimate), delivery (cool restrained, minimal vibrato, whispery, melismatic runs, strong full-voiced, airy stacked harmonies). The smoky contralto has the most training data and is the most reliable.
  • English, female lead only, and every training song is a love song — very dark or aggressive lyrics are outside what it learned. Belted, raspy rock-soul power vocals do not work.
The author's own recipes (from the model card's Listen section)

LoRA weights © becausereasons, CC BY-NC 4.0 (non-commercial; attribute "QTSTRM LoRAs by becausereasons"). Base model m-a-p/YuE2-3B and decoder m-a-p/YuE2-Vae, same licence. This Space runs the official yue2_infer pipeline and merges the Comfy-layout LoRAs into the native checkpoint; ComfyUI solves the decoder ODE with dpm_2/sgm_uniform while this Space uses the reference 32-step midpoint solver, so renders are not sample-identical to the card's demos.