yue2-industrial-rock-lora

YuE2 Industrial Rock LoRA

A set of LoRA adapters for m-a-p/YuE2-3B, fine-tuned to push the base model toward industrial rock: heavy distorted guitars, mechanical/industrial sound textures, a dense mix, and aggressive rhythms.

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).

Each stage has an adapter trained on a large pool (179 tracks) and a small pool (86 tracks). Use the large-pool ones for the best quality; the small-pool ones are kept for A/B comparisons and as the historical "old sound" reference.

adapter-nar-179-v2 is a retrain of the NAR stage with a training bug fixed — it is the recommended NAR adapter (see Known issue & fix). The earlier adapter-nar-179 is kept as the pre-fix reference.

Adapters

All adapters are on top of m-a-p/YuE2-3B. Train band: Celldweller + Blue Stahli (industrial / electronic rock), roughly 1200 steps each.

Naming: the numbers in the adapter names are the number of tracks / clips in the training pool, not steps. The earlier "1200"-style names (adapter-ar-1200, adapter-nar-1200) are gone — replaced by the unambiguous track-count naming above.

Samples

All samples: seed 777.

<audio src="https://huggingface.co/monsterovich/yue2-industrial-rock-lora/resolve/main/ar1_nar05_rockdrive_v2_seed777.flac" controls></audio> ar1_nar05_rockdrive_v2_seed777.flac — upbeat rock-drive example, NAR LoRA retrained with the gradient fix, AR≈1.0 / NAR≈0.5, seed 777.

<audio src="https://huggingface.co/monsterovich/yue2-industrial-rock-lora/resolve/main/ar05_nar1200_seed777.flac" controls></audio> ar05_nar1200_seed777.flac

<audio src="https://huggingface.co/monsterovich/yue2-industrial-rock-lora/resolve/main/ar05_nar257_seed777.flac" controls></audio> ar05_nar257_seed777.flac

Recommended pair

from yue2_inference import YuE2Model  # YuE2 inference code

model = YuE2Model.from_pretrained("m-a-p/YuE2-3B")
model.load_lora("adapter-ar-179/lora.safetensors")       # AR stage (large pool)
model.load_lora("adapter-nar-179-v2/lora.safetensors")   # NAR stage (large pool, bug-fixed retrain)

audio = model.generate(
    prompt="industrial rock, heavy distorted guitars, mechanical rhythm",
    ar_scale=0.5,    # AR LoRA strength
    nar_scale=1.0,   # 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 industrial rock. Each adapter's strength can be tuned independently via its LoRA scale at generation time; the recommended start is AR≈0.5, NAR≈1.0.

Known issue & fix

The first NAR train (adapter-nar-179) shipped with training-bug-silent-looser-timbre: the flow-matching trainer wrapped every transformer layer in torch.utils.checkpoint and detached the layer input before each call (change_requires_grad). With use_reentrant=False that cut the autograd graph between layers, so the loss gradient reached the LoRA weights only along a truncated path instead of back-propagating through all 28 layers. The adapter learned, but wrong — good loss reduction was impossible (mse stuck around 1.38) and the timbre imprint came out exaggerated and brittle.

The fix: remove the detach between layers so the checkpoint graph stays connected and every layer's LoRA receives gradients. adapter-nar-179-v2 is retrained with that fix, achieving mse≈0.59 on the same data. In practice it sounds cleaner and more stable than the old adapter, at the cost of a slightly softer dataset imprint — push the NAR LoRA scale up if you want more of the source timbre.

Tuning: effective strength = alpha / r × scale. Keep NAR low (≈0.3–0.7) for clean tracks; raise until you like the timbre but stop before it turns harsh/hollow. Pushing NAR well above alpha / r = 1.0 degrades into audible noise/artifacts.

Limitations

• The adapters were trained on a limited set of industrial rock tracks, so output quality can vary and may not always hit the genre precisely.
• Requires the m-a-p/YuE2-3B base model — the adapters do not work on their own.