Ruach Studio · LoRAs for YuE2: instruments and voices

Website · GitHub · The guide · ☕ Buy me a Coffee Machine

Adapters that teach YuE2 an instrument it plays weakly or not at all, made in the LoRA Trainer of Ruach Studio (its own trainer, bf16 throughout) and chosen by ear, epoch by epoch, by the studio's author. The repo has four rooms: instruments/ (below), voices/ (six voices by their kind, below), and styles/ and vocals/ (singing voices) for what comes next; each adapter's folder is as the studio keeps it in its loras/.

Where to start. The shofar: e110, the author's choice by ear (a long Yemenite horn, a whole meditation); e099 and e121 are its neighbours, e167 the run's last epoch (thinner, weaker). The duduk: any of the five; the samples below put each beside the base model on the same seeds.

Voices

Six speaking voices sung, each trained on one Russian audiobook reader and named by the kind of voice it carries, never by the person; all about them, how they were made and their samples: voices/README.md.

In Ruach Studio

1. Put an adapter's folder from instruments/ (or one file of it) into the studio's loras/: loras/shofar-yt-r16/shofar-yt-r16-e110.safetensors.
2. In the Creator, the LoRA picker shows it. The samples here are Direct (no score), both halves at 1.0. With a planned score (Full plan, Melody only) keep the music half low (0.5 and less): stacked music strengths past the studio's measured ceiling break the score.
3. The style: begin with the instrument's name, as its captions did (shofar, …, duduk, …). The samples' styles:
   • shofar: shofar, Instrumental, solo shofar, a ram's horn blown in long sustained blasts and broken calls, tekiah, shevarim, teruah, raw piercing horn tone with breath, huge reverberant space, ritual and solemn, nothing else playing
   • duduk: Instrumental, solo duduk. A slow, meditative piece played by a single duduk, close-miked and in front, nothing else playing. Natural room, no vocals.

Both halves are in one file, with AI-Toolkit's key names (text_encoders.* the music half, diffusion_model.* the sound half), so other YuE2 front ends that read AI-Toolkit adapters load them as they are.

How they were made

Ruach Studio's LoRA Trainer, its own engine: the bf16 model, nothing quantized; rank 16, learning rate 1e-4; music half (AR) loss weight 0.5 with a KL anchor of 0.2 and half the learning rate; 60-second windows; Direct (no score).

The recordings are public videos of the instruments (YouTube -0ajtTFYm2k for the duduk; wXrt6S9aGUI, PtiEU7ZEKgs, TXQV2bMnGlk, gjlh864Mhbg, o8-AFAcNl0E, qckg_d1mYO8 for the shofar). Only the stretches that do not repeat were used: some of them loop the same minutes for hours.

Licence

The adapters: CC BY-NC 4.0, as the YuE2 weights they adapt. The samples (audio and pictures): CC BY-NC-ND 4.0, as everything published with Ruach Studio; the studio's code is AGPL-3.0-or-later.

Samples

Each epoch on two seeds, A and B: most on 5101/6101 (A) and 5102/6102 (B), so they compare; a sample made again with new seeds says its own. The length is the take's own.