SF-Flow — pretrained checkpoints
Pretrained models for SF-Flow: Sound field magnitude estimation via flow matching guided by sparse measurements (IWAENC 2026).
SF-Flow reconstructs a dense 3D acoustic transfer function (ATF) magnitude field on an 11×11×11 grid from a sparse, variable-size set of microphone observations, using conditional Flow Matching with a permutation-invariant set encoder and a 3D U-Net.
- Project page: https://egerdem.github.io/sf-flow/
- Code: https://github.com/egerdem/SF-Flow
Checkpoints
Four models trained on dataset R1, one per frequency range. Optimizer state has been stripped, so each file is ~460 MB and is intended for evaluation or fine-tuning from the weights rather than exact resumption of the original training run.
| file | freq bins (max freq) | val LSD (dB) | paper test LSD (dB, M=5) |
|---|---|---|---|
sfflow_r1_freq20.pt |
0–20 (312 Hz) | 1.7555 | 1.75 ± 0.58 |
sfflow_r1_freq30.pt |
0–30 (468 Hz) | 3.1703 | 3.17 ± 0.67 |
sfflow_r1_freq40.pt |
0–40 (625 Hz) | 4.1857 | 4.16 ± 0.63 |
sfflow_r1_freq64.pt |
0–64 (1000 Hz) | 5.5800 | 5.56 ± 0.52 |
Test LSD is averaged over the 102 test sources with M=5 observations, using the fixed microphone-selection protocol described in the code repository.
Usage
hf download egeerdem/sf-flow sfflow_r1_freq30.pt --local-dir .
python evaluate.py --model_path sfflow_r1_freq30.pt \
--data_dir data/ir_fs2000_s1024_m1331_room4.0x6.0x3.0_rt200/
evaluate.py, and the script that regenerates the dataset the checkpoints expect, are in
the code repository.