HiiragiLee/affectscore-ace-step-r32-20260629

AffectScore -- LoRA r=32 (full, selected)

LoRA adapter for ACE-Step v1 3.5B fine-tuned for real-time affect-conditioned music generation in narrative-driven games. This is the primary checkpoint selected by held-out MER accuracy.

Part of the AffectScore system. Code and Colab notebooks: github.com/LeeTgk/affectscore (link active after publication)

---

What this model does

AffectScore conditions music generation on a two-layer signal:

• Layer 1 -- Designer intent (authored per scene): scene_valence, scene_arousal, arc_position on the Russell circumplex. Converted to mood-word text prompts (e.g. "melancholic, tense") via a lookup table.
• Layer 2 -- Player engagement (runtime behavioral): choice_latency_norm, dwell_deviation_norm, interaction_rate_norm. Applied as an inference-time text suffix that modulates intensity and texture around the Layer 1 anchor without overriding it.

LoRA targets cross_attn.q_proj and cross_attn.v_proj only. Self-attention and FFN stay frozen.

---

Training

Parameter	Value
Base model	ACE-Step/ACE-Step-v1-3.5B
LoRA rank	32
LoRA targets	cross_attn.q_proj, cross_attn.v_proj
use_rslora	True
lora_dropout	0.05
Trainable parameters	7.86M (0.237% of total)
Training clips	1,571 (CC0/CC-BY, CLAP-quality filtered)
Epochs	50
Learning rate	1e-4
Batch size	8
Quadrant weighting	max-quadrant-weight 5.0 (compensates Q2 scarcity)
Hardware	NVIDIA A100 (Google Colab)

Training dataset: https://zenodo.org/records/21830658

---

Evaluation

All metrics computed on 200 held-out clips (50 per V-A quadrant) at 40 diffusion steps.

Audio quality (EVAL-02)

FAD-MERT ↓	CLAP-score ↑	KLD ↓	PCE ↑
33.06	−0.45	1.17	3.57

Note: FAD-MERT is higher than the no-lora baseline because LoRA fine-tuning specializes the model distribution toward the training corpus. This is expected and not evidence of degraded quality -- see the paper for full discussion.

Emotion controllability (EVAL-03)

Pearson r between designer-intended and music2emo-predicted V-A on 200 held-out clips. No variant achieves significant positive correlation (critical value |r| ≥ 0.138 at α=0.05); null result attributed to Q3-heavy training distribution (78.1% melancholic clips) and text-space ceiling.

r_valence	RMSE_valence	r_arousal	RMSE_arousal
−0.171	0.520	−0.087	0.292

---

How to use

This adapter is designed to run inside the AffectScore FastAPI server, which handles conditioning, generation, and WAV encoding end-to-end. Direct PEFT loading is possible but the conditioning pipeline is non-trivial -- refer to server/affectscore_server.py in the code repository.

python server/affectscore_server.py \
  --lora HiiragiLee/affectscore-ace-step-r32-20260629

The server exposes a /generate endpoint accepting a 6-float affect embedding (Layer 1 + Layer 2), style prompt, chunk duration, and step count.

---

Ablation variants

Repo	Description
This model	Full two-layer conditioning, r=32 (selected)
affectscore-ace-step-r16-20260629	Lower capacity, r=16
affectscore-ace-step-r64-20260629	r=64, diverged at epoch 40 -- not recommended
affectscore-ace-step-r32-20260629-no-style	No Layer 2 style descriptor (ablation)
affectscore-ace-step-r32-20260629-no-affect	No Layer 1 V-A conditioning (ablation)
affectscore-ace-step-r32-20260629-no-lora	Vanilla ACE-Step base (reconstruction reference)

---

Citation

Citation will be added once the paper is published. If you use this model before then, please link to this repository.