affectscore-ace-step-r32-20260629-no-lora

AffectScore -- Vanilla ACE-Step baseline (no-lora)

Unmodified ACE-Step v1 3.5B base model weights, saved as a reconstruction-quality reference for the AffectScore ablation study. No LoRA fine-tuning was applied. This is not an adapter -- it is the base model checkpoint used to establish a pre-training performance ceiling for FAD-MERT, KLD, and PCE comparisons.

Included here to make the full ablation suite reproducible without requiring a separate download of the base model. For actual music generation, use ACE-Step directly or the full AffectScore r=32 adapter.

Code and Colab notebooks: github.com/LeeTgk/affectscore (link active after publication)

---

Evaluation

Metrics on 200 held-out clips at 40 diffusion steps, included as the reference row in the AffectScore ablation table.

Audio quality (EVAL-02)

The lower FAD-MERT relative to fine-tuned variants reflects the base model's broader pretraining distribution, not superior output quality for game audio use cases. Fine-tuning specializes the model toward the training corpus at the cost of some generalization -- this is expected and discussed in the paper.

Emotion controllability (EVAL-03)

---

Ablation variants

See affectscore-ace-step-r32-20260629 for the full variant table and evaluation discussion.

---

Citation

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