Haoyu123123/loom-of-time-models

Loom of Time — Model Weights

Model checkpoints for Loom of Time (岁月织机), an interactive art installation presented at ACM Multimedia 2026 (Interactive Art Track).

GitHub repository: HAoyu-baller/Loom-of-Time-Music-Viz

---

Contents

models/
  separator/
    best.pt       # BSRoformer 5-stem separator (best val loss)
    last.pt       # BSRoformer 5-stem separator (final epoch)
  classifier/
    loom_best.pth # ResNet18 regional classifier (5 Chinese folk regions)

---

Models

BSRoformer Separator (models/separator/best.pt)

Separates a Chinese folk music recording into 5 stems: vocal · erhu · plucked strings · wind · percussion

Config	Value
Architecture	Band-Split Rotary Transformer (BSRoformer)
dim	256
depth	6
heads	8
Parameters	~15–20 M
Sample rate	22,050 Hz
Chunk size	132,300 samples (6 s) with 22,050-sample overlap

ResNet18 Classifier (models/classifier/loom_best.pth)

Classifies the regional origin of a Chinese folk song into one of 5 categories: Jiangnan · Shaanxi · Yungui · Huanan · Dongbei

Config	Value
Backbone	ResNet18 (ImageNet pretrained)
Head	Dropout(0.4) → Linear(512,256) → ReLU → Dropout(0.3) → Linear(256,5)
Input	Mel-spectrogram image (128 mel bins, magma colormap, 224×224 px)
Inference	Chunk-based hard voting across 6-second clips

---

Usage

See the GitHub repository for full installation instructions and inference code.

from huggingface_hub import hf_hub_download

hf_hub_download('Haoyu123123/loom-of-time-models', 'models/separator/best.pt',
                local_dir='backend/training/checkpoints')
hf_hub_download('Haoyu123123/loom-of-time-models', 'models/classifier/loom_best.pth',
                local_dir='backend/classifier/models')

---

Citation

@inproceedings{ruan2026loom,
  title     = {Loom of Time: Real-Time Chinese Folk Music Decomposition
               and Generative Weaving Visualization},
  author    = {Ruan, Haoyu},
  booktitle = {Proceedings of the 34th ACM International Conference on
               Multimedia (Interactive Art Track)},
  year      = {2026},
  publisher = {ACM}
}