yue2-steps-from-hell

YuE2 Two Steps From Hell LoRA

A set of LoRA adapters for m-a-p/YuE2-3B, fine-tuned to push the base model toward epic cinematic orchestral trailer music in the spirit of Two Steps from Hell: massive symphonic orchestra with soaring heroic strings, thunderous taiko drums and pounding orchestral percussion, dramatic brass fanfares, powerful angelic and dark choir, and uplifting heroic cinematic climaxes.

Description

YuE2 generates music in two stages, and there is one adapter per stage:

• AR adapter — shapes the plan: structure, melody, and genre direction (the autoregressive planning stage).
• NAR adapter — shapes the timbre: the actual instrumental/audio texture (the non-autoregressive decoding stage).

Adapters

Both adapters sit on top of m-a-p/YuE2-3B. The training pool was a full-track TSFH-style orchestral trailer library (flat 25 fps, full-length tracks, no slicing).

Naming: the numbers in the adapter names are the number of tracks in the training pool. adapter-nar-194 excludes the single longest track (~7.5 min, 11358 frames at 25 fps): the NAR trainer consumes the full track in a single forward pass and that one track does not fit on the 8 GB GPU even when trained on its own (a single isolated step hits CUDA OOM).

The AR adapter was first trained on 179 tracks (1200 steps) and then finetuned for 400 steps exclusively on the 16 longest tracks so that every track in the library contributed.

Samples

All samples: seed 777.

<audio src="https://huggingface.co/monsterovich/yue2-steps-from-hell/resolve/main/tsfh_ar05_nar0375_seed777.flac" controls></audio> tsfh_ar05_nar0375_seed777.flac — new original track in the TSFH style (no reference melody), AR≈0.5 / NAR≈0.375 (alpha/r × scale), seed 777.

Recommended pair

from yue2_inference import YuE2Model  # YuE2 inference code

model = YuE2Model.from_pretrained("m-a-p/YuE2-3B")
model.load_lora("adapter-ar-195/lora.safetensors")     # AR stage (plan)
model.load_lora("adapter-nar-194/lora.safetensors")    # NAR stage (timbre)

audio = model.generate(
    prompt="Epic cinematic orchestral trailer music, massive symphonic orchestra "
           "with soaring heroic strings, thunderous taiko drums and pounding "
           "orchestral percussion, dramatic brass fanfares, powerful angelic and "
           "dark choir, uplifting heroic cinematic soundscapes, massive epic "
           "climaxes, professional studio recording",
    ar_scale=0.5,   # AR LoRA strength
    nar_scale=0.375,  # NAR LoRA strength
    seed=777,
)

Usage

The adapters sit on top of the m-a-p/YuE2-3B base model following the standard YuE2 inference pipeline (autoregressive planning → non-autoregressive decoding). Load the base model, apply the AR adapter to the AR stage and the NAR adapter to the NAR stage, then run generation with lyrics geared toward epic orchestral trailer music. Each adapter's strength can be tuned independently via its LoRA scale at generation time; the recommended start is AR≈0.5, NAR≈0.375.

Tuning: effective strength = alpha / r × scale. Both adapters ship with alpha / r = 1.0. Keep NAR at ≈0.3–0.7 for clean tracks; push higher only if you want a stronger source timbre, but stop before it turns harsh/hollow. If artifacts persist, halve alpha (e.g. 32 → 16) or lower nar_scale — both reduce effective strength identically.

Limitations

• Trained on a single epic-orchestral library; output quality varies outside that slot, and the exact tracks are not recreated (the model composes new originals).
• Requires the m-a-p/YuE2-3B base model — the adapters do not work on their own.
• The single longest training track (~7.5 min) is not covered by the NAR adapter: even a single isolated training step on it exceeds 8 GB of GPU memory.