SF-Flow: 3D Sound Field Estimation

• SF-Flow is a conditional flow-matching framework for 3D acoustic transfer function magnitude estimation, reconstructing dense sound fields from sparse measurements.
• It integrates a 3D U-Net with a Transformer-based permutation-invariant encoder to accurately fuse geometric and spectral data.
• Experimental results indicate SF-Flow achieves lower log-spectral distortion and faster training times at low frequencies compared to baseline methods.

Searching arXiv for “SF-Flow” and closely related usages to ground the article in current literature. SF-Flow is a conditional flow-matching framework for reconstructing dense three-dimensional acoustic fields from sparse microphone observations. In the formulation introduced in “SF-Flow: Sound field magnitude estimation via flow matching guided by sparse measurements” (Erdem et al., 11 May 2026), the target is the magnitude of the Acoustic Transfer Function (ATF), represented as a 3D volumetric field conditioned on an arbitrary, unordered set of sparse measurements. The method casts reconstruction as guided generation rather than direct interpolation or deterministic regression: a 3D U-Net learns a flow-matching vector field, while a permutation-invariant set encoder injects sparse observation evidence. The reported system reconstructs ATF magnitude accurately up to $1$ kHz, trains substantially faster than the autoencoder baseline, and improves significantly with dataset size (Erdem et al., 11 May 2026).

1. Terminology and scope

The designation “SF-Flow” is exact, in the strict sense, for the spatial-audio method on 3D ATF magnitude estimation (Erdem et al., 11 May 2026). A recurring source of confusion is that closely related abbreviations recur across unrelated literatures. In generative modeling, “SFM” may denote Switched Flow Matching (Zhu et al., 2024), Shallow Flow Matching (Yang et al., 18 May 2025), or Smooth Flow Matching (Tan et al., 19 Aug 2025); in stochastic dynamics, “[SF]$^2$M” denotes simulation-free score and flow matching for Schrödinger bridges (Tong et al., 2023); and in datacenter networking, “SFC” denotes Source Flow Control rather than a flow-matching model (Le et al., 2023). SF-Flow, as a named method, refers to the sound-field reconstruction framework of (Erdem et al., 11 May 2026).

Term	Meaning in the cited literature	arXiv id
SF-Flow	Sound field magnitude estimation via flow matching guided by sparse measurements	(Erdem et al., 11 May 2026)
SFM	Switched Flow Matching	(Zhu et al., 2024)
SFM	Shallow Flow Matching	(Yang et al., 18 May 2025)
SFM	Smooth Flow Matching	(Tan et al., 19 Aug 2025)
[SF]$^2$M	Simulation-free score and flow matching	(Tong et al., 2023)
SFC	Source Flow Control	(Le et al., 2023)

This terminological ambiguity matters because SF-Flow shares the broad flow-matching lineage with several of these systems, but its task, data geometry, and conditioning structure are specific to spatial acoustics rather than image, speech, or transport modeling.

2. Problem formulation in spatial acoustics

SF-Flow addresses 3D sound-field reconstruction through ATF magnitude estimation (Erdem et al., 11 May 2026). The ATF is written as

$$H(\mathbf{p}_{\text{src}}, \mathbf{p}_{\text{mic}}, f),$$

where $\mathbf{p}_{\text{src}}$ is the source position, $\mathbf{p}_{\text{mic}}$ is the microphone position, and $f$ is frequency. The method focuses on $|H(\cdot)|$, expressed in logarithmic dB scale, because ATF magnitude already captures modal resonances, spectral coloration, frequency-dependent spatial energy distribution, and energy decay patterns across a room (Erdem et al., 11 May 2026).

For a fixed source position, the target field is a dense cube

$$\mathbf{H} \in \mathbb{R}^{F \times D \times H \times W},$$

with $F$ frequency bins over a 3D spatial grid (Erdem et al., 11 May 2026). Conditioning information is provided as a sparse set

$^2$0

where $^2$1 is a geometry descriptor and $^2$2 is the measured ATF magnitude spectrum at the $^2$3-th microphone (Erdem et al., 11 May 2026). The geometry vector is specified as

$^2$4

with $^2$5 the source-to-microphone vector and $^2$6 the source-to-wall distances in a cuboid room (Erdem et al., 11 May 2026).

The problem is ill-posed for several reasons explicitly emphasized in the paper: severe undersampling, infinitely many completions consistent with the same sparse samples, fine-grained interference structure, stronger oscillation at higher frequencies, and variable-size unordered conditioning sets (Erdem et al., 11 May 2026). This motivates a generative prior over plausible fields rather than a purely local interpolant.

3. Flow-matching formulation

SF-Flow formulates reconstruction as conditional generation using Flow Matching (FM) and, specifically, Conditional Flow Matching (CFM) (Erdem et al., 11 May 2026). The latent trajectory $^2$7 evolves according to

$^2$8

with $^2$9 taken as a standard Gaussian prior and $^2$0 approaching a valid ATF magnitude cube (Erdem et al., 11 May 2026).

The paper uses the CFM objective

$^2$1

with conditional target vector field

$^2$2

An Optimal Transport path is chosen with

$^2$3

so that

$^2$4

and the target field simplifies to

$^2$5

The training loss therefore reduces to

$^2$6

The paper interprets this as learning a straight-line transport field from Gaussian noise to a valid ATF cube, rather than a denoising chain of the diffusion type (Erdem et al., 11 May 2026).

In this respect SF-Flow is structurally aligned with the broader FM literature, where a vector field generates a probability path, but its novelty lies in adapting that construction to 3D spatial acoustics with sparse-set conditioning (Zhu et al., 2024, Erdem et al., 11 May 2026).

4. Conditioning architecture and reconstruction pipeline

A central architectural requirement is support for an arbitrary number of sparse microphone observations. SF-Flow therefore uses a Transformer-based permutation-invariant set encoder coupled to a 3D U-Net (Erdem et al., 11 May 2026).

Each observation $^2$7 is projected by two-layer MLPs into a shared embedding of dimension $^2$8; coordinate and magnitude embeddings are summed to form tokens; and a 3-layer Transformer encoder with 8 attention heads refines the set representation (Erdem et al., 11 May 2026). Padding with a learned null token and masking handles variable $^2$9 (Erdem et al., 11 May 2026). The encoder emits two conditioning outputs: a token sequence

$$H(\mathbf{p}_{\text{src}}, \mathbf{p}_{\text{mic}}, f),$$0

used for cross-attention in the generator, and a global pooled context

$$H(\mathbf{p}_{\text{src}}, \mathbf{p}_{\text{mic}}, f),$$1

obtained by masked mean pooling (Erdem et al., 11 May 2026).

The generator is a 3D U-Net operating on

$$H(\mathbf{p}_{\text{src}}, \mathbf{p}_{\text{mic}}, f),$$2

and predicts a vector field of the same shape (Erdem et al., 11 May 2026). The U-Net receives a sinusoidal time embedding, the global context $$H(\mathbf{p}_{\text{src}}, \mathbf{p}_{\text{mic}}, f),$$3, and the token sequence $$H(\mathbf{p}_{\text{src}}, \mathbf{p}_{\text{mic}}, f),$$4 (Erdem et al., 11 May 2026). Its specified structure includes two residual downsampling stages with channels $$H(\mathbf{p}_{\text{src}}, \mathbf{p}_{\text{mic}}, f),$$5 and $$H(\mathbf{p}_{\text{src}}, \mathbf{p}_{\text{mic}}, f),$$6, a bottleneck at $$H(\mathbf{p}_{\text{src}}, \mathbf{p}_{\text{mic}}, f),$$7 channels, mirrored decoder stages with transposed-convolution upsampling, skip connections, a final $$H(\mathbf{p}_{\text{src}}, \mathbf{p}_{\text{mic}}, f),$$8 convolution to $$H(\mathbf{p}_{\text{src}}, \mathbf{p}_{\text{mic}}, f),$$9 output channels, and reflective padding to $\mathbf{p}_{\text{src}}$0 followed by cropping back (Erdem et al., 11 May 2026). Residual blocks inject time embedding and global context as biases after group norm, while cross-attention allows spatial features to attend directly to observation tokens (Erdem et al., 11 May 2026). A per-frequency FiLM modulation provides frequency-specific conditioning (Erdem et al., 11 May 2026).

Training samples a ground-truth cube $\mathbf{p}_{\text{src}}$1, a time $\mathbf{p}_{\text{src}}$2, and noise $\mathbf{p}_{\text{src}}$3, forms

$\mathbf{p}_{\text{src}}$4

samples an observation count

$\mathbf{p}_{\text{src}}$5

and randomly selects $\mathbf{p}_{\text{src}}$6 microphone positions and spectra to define $\mathbf{p}_{\text{src}}$7 (Erdem et al., 11 May 2026). Inference uses 10 Euler steps (Erdem et al., 11 May 2026).

5. Experimental regime and reported performance

The reported experiments are conducted in simulation using pyroomacoustics, with room size $\mathbf{p}_{\text{src}}$8 m, reverberation time $\mathbf{p}_{\text{src}}$9 s, random source positions, an $\mathbf{p}_{\text{mic}}$0 target grid, sampling rate 2000 Hz, room impulse response length 128 samples, and 64 Fourier bins up to 1000 Hz (Erdem et al., 11 May 2026). The main dataset, R1, contains 1024 source positions split into 820 train, 102 validation, and 102 test examples; larger variants R2, R3, and R3 Long scale to 4096 and 8192 sources (Erdem et al., 11 May 2026).

Two baselines are reported: a conditioned autoencoder trained with Log-Spectral Distortion (LSD) loss, and a Gaussian kernel ridge regression (KRR) baseline (Erdem et al., 11 May 2026). The principal metric is

$\mathbf{p}_{\text{mic}}$1

with $\mathbf{p}_{\text{mic}}$2 spatial locations (Erdem et al., 11 May 2026).

For $\mathbf{p}_{\text{mic}}$3 observations on R1, the paper reports the following LSD values (Erdem et al., 11 May 2026):

Method	0–20 bins (312 Hz)	0–30 (468 Hz)	0–40 (625 Hz)	0–64 (1000 Hz)
KRR	6.59 ± 1.48	8.11 ± 1.35	9.05 ± 1.14	10.67 ± 1.11
AE	2.69 ± 1.22	3.71 ± 0.58	4.06 ± 0.43	4.55 ± 0.41
SF-Flow	1.76 ± 0.71	3.17 ± 0.67	4.16 ± 0.63	5.56 ± 0.52

These results support a nuanced interpretation. SF-Flow outperforms the autoencoder at low and mid frequencies up to 468 Hz, while the autoencoder becomes slightly better in LSD at higher frequencies because it tends to oversmooth toward the spatial mean, thereby lowering average error but losing detail (Erdem et al., 11 May 2026). The visual comparisons reported in the paper indicate that SF-Flow preserves more fine spatial structure, whereas the autoencoder becomes near-uniform at higher frequencies (Erdem et al., 11 May 2026).

The method is also reported to work with very few observations. On the 0–20 bin setting, SF-Flow yields LSD values of 1.99, 1.76, 1.73, 1.71, and 1.71 for $\mathbf{p}_{\text{mic}}$4, respectively, while the autoencoder changes only slightly from 2.71 to 2.66 over the same range (Erdem et al., 11 May 2026). The paper interprets this as evidence that SF-Flow makes more effective use of additional sparse evidence.

Training-speed comparisons are similarly explicit. On R1, the autoencoder requires about 87–108 seconds per epoch, whereas SF-Flow requires about 20 seconds per epoch; best-checkpoint wall-clock time is about 19–24 hours for the autoencoder and about 2.4–5.8 hours for SF-Flow (Erdem et al., 11 May 2026). Increasing dataset size improves performance substantially: for the 0–64 bin setting, SF-Flow decreases from 5.56 on R1 to 4.44 on R2, 4.08 on R3, and 3.67 on R3 Long (Erdem et al., 11 May 2026).

6. Relation to adjacent flow-matching research and present limitations

SF-Flow belongs to a wider class of methods that use flow matching to learn transport fields, but its role within that family is highly domain-specific. Switched Flow Matching replaces a single ODE with multiple conditional ODEs selected by a switching signal to address singularities induced by heterogeneous source and target distributions (Zhu et al., 2024). Shallow Flow Matching for text-to-speech starts the ODE from an intermediate coarse representation rather than pure noise and integrates only the later part of the path (Yang et al., 18 May 2025). Smooth Flow Matching for functional data learns a scalar flow $\mathbf{p}_{\text{mic}}$5 over irregularly observed curves and guarantees smooth generated trajectories (Tan et al., 19 Aug 2025). Simulation-free score and flow matching, or [SF]$\mathbf{p}_{\text{mic}}$6M, unifies score matching and flow matching in a Schrödinger-bridge framework without simulating learned SDEs during training (Tong et al., 2023). SF-Flow shares the FM paradigm with these methods but does not reuse their problem settings, priors, or conditioning mechanisms.

Several misconceptions can therefore be excluded. SF-Flow is not a generic synonym for SFM, not a networking control mechanism, and not a shallow-start or switched-ODE variant by name (Le et al., 2023, Zhu et al., 2024, Yang et al., 18 May 2025). It is a conditional generative model for 3D ATF magnitude fields (Erdem et al., 11 May 2026).

The current formulation also has clear limitations stated in the paper. Experiments are confined to simulated single-room data; only magnitude is modeled, with phase left for future work; evaluation is limited to LSD; generalization to multiple rooms and real recordings remains untested; and the reported setup is frequency-limited to 1 kHz (Erdem et al., 11 May 2026). A plausible implication is that the method’s present empirical claims should be read as establishing a strong proof of concept for sparse-measurement-guided sound-field generation, rather than a complete solution to broadband real-room reconstruction.