SCAV

SCAV Checkpoints

Official pretrained checkpoints for SCAV from Whence the Voice? Self-supervised Dual-source Audio-Visual Localisation via Selective Convergence (ECCV 2026).

Project page · Paper · Code · VGGSound-Duet^Mask

Model description

Localising multiple simultaneous sound sources is challenging because source separation benefits from knowing source locations, while localisation benefits from separated audio. SCAV uses the selective convergence behaviour of contrastive audio-visual models in a progressive two-stage framework:

1. Stage 1 learns dominant-source localisation from mixed audio using contrastive learning.
2. Stage 2 uses the Stage 1 spatial prior to decouple visual and audio features and reveal the subdominant source.

The method is self-supervised and does not require manual localisation annotations during training. The paper reports single-forward-pass localisation of both sources at 43.2 FPS on one NVIDIA A100 GPU.

Files

The checkpoints are provided together because Stage 2 is part of the same progressive SCAV pipeline and depends on heatmaps produced by Stage 1.

Download

The repository is public and does not require an access token.

Hugging Face CLI

pip install -U "huggingface_hub[cli]"
hf download lemonweed6312/SCAV --local-dir ./SCAV-weights

To download one checkpoint only:

hf download lemonweed6312/SCAV checkpoints/best-stage1.pth \
  --local-dir ./SCAV-weights

Python

from huggingface_hub import hf_hub_download

stage1_path = hf_hub_download(
    repo_id="lemonweed6312/SCAV",
    filename="checkpoints/best-stage1.pth",
)

stage2_path = hf_hub_download(
    repo_id="lemonweed6312/SCAV",
    filename="checkpoints/best-stage2.pth",
)

Requirements

Use these checkpoints with the official SCAV repository:

git clone https://github.com/happy-new-bears/SCAV.git
cd SCAV

The main requirements are Python 3.8 or newer, PyTorch 1.12 or newer, NumPy, h5py, OpenCV, scikit-learn, matplotlib, tqdm, and FFmpeg. CUDA is recommended. See the code repository for data preprocessing and the complete runtime setup.

Loading the checkpoints

The official evaluation scripts instantiate the corresponding model, load the checkpoint with torch.load, and accept either a wrapped model_state_dict or a plain state dictionary:

import torch

checkpoint = torch.load("/path/to/best-stage1.pth", map_location="cpu")
state_dict = checkpoint.get("model_state_dict", checkpoint)

# Instantiate FlowAttnHardWayModel exactly as in stage1_vggss/eval.py
# or stage1_vggss/eval_seg.py, then:
model.load_state_dict(state_dict)
model.eval()

Use FlowAttnHardWayModel for best-stage1.pth and ASLNet for best-stage2.pth. Model definitions and constructor settings are maintained in the official code repository.

Citation

If you use these checkpoints, please cite:

@inproceedings{hu2026scav,
  title     = {Whence the Voice? Self-supervised Dual-source Audio-Visual Localisation via Selective Convergence},
  author    = {Hu, Han and Lin, Dongheng and Hou, Yuqi and Li, Haotian and Chang, Hyung Jin and Jiao, Jianbo},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026}
}