m-a-p

/

YuE2-3B

Like 1.23k

Follow

Multimodal Art Projection 1.62k

Hello m-a-p, and thank you again for YuE2. Our studio is built around it.

Ruach Studio 2.0.0 is out, its first stable release. It is a free, open studio that runs YuE2 on your own hardware through yue2.cpp: Linux with an NVIDIA card, Windows through WSL2, Intel Arc as a beta. The model, the score, the adapters and every take stay on your machine. Six rooms as one:

• the Creator writes and edits songs as YuE2 plans them: Full mode with the score, Only the opening, Extend, re-singing a line, a section or a window, chords and a chord bed, and a piano roll for YuE2's ABC scores (beta);
• the Writer, a lyrics room with a language model of your choice;
• the Refiner, a canvas of nodes from eight stems to mastering, with a DAW (beta);
• the Artist, for covers;
• the Librarian, with a playlist;
• the Trainer, LoRA training for YuE2 on one card or several.

The guide is in twelve languages.

YuE2-RS-FW is our first fine-tune of YuE2-3B, the First Wave, published only as merged weights. Each merge comes as GGUF (BF16, Q8_0, Q6_K) and as BF16 safetensors in your own layout, under CC BY-NC-SA 4.0 with credit to YuE2-3B.

• Data: 32,633 tracks, about 1,800 hours. MTG-Jamendo and FMA above an Audiobox Aesthetics bar (PQ ≥ 7.0), two community sets of songs with lyrics, and a guard of sung and spoken material. 31,801 tracks have their own ABC score (SheetSage2), so Full mode is trained too.
• Training: LoRA on both halves (music half rank 160, sound half rank 320), four epochs, EMA, three RTX 3090s.
• Merges: seven checkpoint and strength pairs. Ruach Studio takes 2.4 × 0.75 as its default for now; the blind listening that picks the best one is under way.

Findings that may interest you:

• A later save weighs more. Measured as the norm of what a save adds, the change grows about 3× on the music half and 2.5× on the sound half from epoch 0.7 to 4.0. At one strength, a late save pushes much harder, so we merge later saves lighter.
• The music half keeps its scores. 320 plans in Full mode (20 cards × saves 0.7–4.0 × strengths 0.6–0.75) passed the score guard, none broken; a second seed at 0.75 gave 80 of 80 with a sung vocal line.
• A grain on high female voices. From about 1.5 epochs on, a fine grain (a «sand») comes back on high female voices, as we hear it in the stock weights; we are testing lighter strengths and a lighter sound half alone.
• A ring against two closed services, the same 20 cards (style and lyrics) sent to each, raw takes, one yardstick. The table is on the model card.
  • On lyrics (Whisper's word error), the 0.7 save is at 18 % against 17 % and 23 % for the two closed services; the stock weights are at 21 %.
  • The 0.7 save measures the least sand of all five columns, and half the stray tones of the stock weights.
  • The decoder's 25 Hz frame buzz in the highs measures about 3 dB above the closed services in every YuE2 column, stock and fine-tuned alike. We thought you might want that number.

The merges and their measurements are on the model card. Questions and corrections are welcome.

Viktor Zhuromskyy and Ruach Studio