ImpactMat: Continuous Material Estimation for Inverse Impact Sound Rendering

Four bunnies of different materials (Ceramic, Plastic, Steel, Glass) dropped in a physics sim — the impact sounds are procedurally rendered from each material's estimated parameters at the simulated contacts (FEM modal synthesis, stereo-panned to match the screen). 🔊 Sound is on — if your browser blocks autoplay audio, click anywhere to unmute.

Abstract

Impact sound rendering synthesizes the sound produced when a 3D object is struck, but practical renderers often rely on fixed material presets such as wood, plastic, or steel. These presets limit the range of impact sounds a renderer can express, while manually adjusting the underlying material parameters remains difficult without expertise in material acoustics. We therefore study inverse impact sound rendering: predicting material parameters from a reference impact sound so that a simulator can recreate a similar material response. To support this task, we introduce ImpactMat, a dataset and benchmark of single and blended material impact sounds paired with ground-truth material parameters. We further propose a feed-forward model that predicts these parameters from one or more recordings, using blended materials to learn smooth transitions between material types. Experiments show that our method outperforms competitive baselines and enables re-rendering from real recordings without manual parameter tuning.

Method

A set of K impact recordings from the same object is encoded, pooled as an unordered set, and mapped to continuous, renderer-compatible material parameters. Blend supervision regularizes both the prediction and the latent representation.

K impact soundssame object

→

Audio EncoderPANN CNN14

→

Set Poolingmean & std over hits

→

Regression Head+ aux. material class

→

Parametersθ = [logE, logρ, ν, logα, logβ]

E Young's modulus

ρ density

ν Poisson's ratio

α,β Rayleigh damping

Training combines a target-weighted regression loss, a soft-label material classification head, a blend regression term, and a latent blend consistency loss that aligns a blend's audio latent with the interpolation of its two base-material latents — teaching a continuous acoustic-to-physical mapping instead of a lookup over discrete classes.

Physical modal analysis

The parameters are renderer-compatible: every object is voxelized and solved with FEM modal analysis, and the estimated parameters drive the same modal synthesizer at re-render time — so the whole loop, from geometry to sound, stays physically grounded. Below: a voxelized mesh and several of its FEM vibration modes (displacement magnitude).

Where the parameters enter the sound

Each of the five targets is a coefficient of a classical modal sound synthesis equation, so predicting θ = [E, ρ, ν, α, β] is the same as predicting how the object rings. Each parameter is color-coded wherever it enters the equations.

Modal analysis (FEM)O’Brien et al. 2002 · Ren et al. 2013

K(\htmlClass{pE}{E},\htmlClass{pN}{\nu})\,\boldsymbol{\phi}_i=\omega_i^{2}\,M(\htmlClass{pR}{\rho})\,\boldsymbol{\phi}_i\;\;\Rightarrow\;\;\omega_i=\sqrt{\htmlClass{pE}{E}/\htmlClass{pR}{\rho}}\;\tilde{\omega}_i(\htmlClass{pN}{\nu})

Pitch. A stiffer object (higher E) rings higher and a heavier one (higher ρ) rings lower: every partial scales with \sqrt{\htmlClass{pE}{E}/\htmlClass{pR}{\rho}}. ν only nudges the ratios between partials.

Rayleigh dampingRayleigh 1877 · Ren et al. 2013

C=\htmlClass{pA}{\alpha}M+\htmlClass{pB}{\beta}K\;\;\Rightarrow\;\;d_i=\tfrac{1}{2}\big(\htmlClass{pA}{\alpha}+\htmlClass{pB}{\beta}\,\omega_i^{2}\big)

Decay. α adds the same damping to every mode (it dominates the low ones); β grows with \omega_i^{2}, so it kills the high modes first. Each partial fades as e^{-d_i t}.

Continuous material blends

Because the targets are continuous, a material can be interpolated. Drag the slider to morph one object from Plastic to Ceramic: the impact sound and the underlying physical parameters move smoothly together (ρ, E, α, β by geometric mean, ν arithmetic) — the continuity our latent blend-consistency objective is built for.

Plastic Ceramic

—

Same object & strike throughout — only the material changes. Each step is a full FEM modal re-solve at the interpolated parameters.

Blend rule

\log p_{\text{blend}}=w\log p_1+(1-w)\log p_2,\qquad p\in\{\htmlClass{pR}{\rho},\htmlClass{pE}{E},\htmlClass{pA}{\alpha},\htmlClass{pB}{\beta}\}

\htmlClass{pN}{\nu}_{\text{blend}}=w\,\htmlClass{pN}{\nu}_1+(1-w)\,\htmlClass{pN}{\nu}_2

Geometric mean for the quantities that span orders of magnitude, arithmetic for ν — exactly what the slider does at every step.

Listen · Synthetic re-rendering

For each object we play the input impact and the sound re-rendered from our predicted parameters (FEM modal synthesis, same object & hit). Close match = accurate parameters.

Ceramic

Object · pure

Input impact

Re-rendered (ours)

InputRe-render

Glass

Object · pure

Input impact

Re-rendered (ours)

InputRe-render

Plastic

Object · pure

Input impact

Re-rendered (ours)

InputRe-render

Log-mel spectrograms (onset-aligned): the re-render reproduces the input's modal structure and decay from the predicted parameters alone.

Listen · Real-world objects

Real recordings and 3D models from ObjectFolder-Real: each panel shows the log-mel of the real recording (left) next to the sound re-rendered from the material parameters our model predicted from that recording (right, playable). Hit point is arbitrary and hit time is estimated (no metadata used).

Ceramic

Vase (green)

Re-rendered (ours)

Real inputRe-render

Glass

Salad Bowl

Re-rendered (ours)

Real inputRe-render

Iron

Kettlebell

Re-rendered (ours)

Real inputRe-render

Results

Our model achieves lower MAE than each baseline across all jointly evaluated parameters, with Avg. NMAE of 0.078 on single-material impacts and 0.116 on blends. It runs in 6.1 ms per query, compared with seconds for the audio–language baselines and 190 ms for optimization-based estimation.

Method	Setting	Test subset	ρ	E	ν	α	β	Avg. NMAE	Infer. time	VRAM
DSP + Ridge Reg.	Conventional	Single	0.190	0.443	0.047	9.93	3.26×10^-7	1.405	16.1 (ms)	CPU
DSP + Ridge Reg.	Conventional	Blend	0.131	0.321	0.032	5.19	1.83×10^-7	0.961	16.1 (ms)	CPU
Qwen2.5-Omni-7B	ICL	Single	0.338	0.614	0.071	13.38	1.61×10^-1	1.270	5.6 (s)	16.1 GB
Qwen2.5-Omni-7B	ICL	Blend	0.257	0.495	0.067	8.23	1.73×10^-1	1.162	5.6 (s)	16.1 GB
Audio Flamingo 3	ICL	Single	14.1	2.14	0.130	21.7	1.8×10^-2	12.42	4.4 (s)	18.3 GB
Audio Flamingo 3	ICL	Blend	13.4	2.33	0.123	11.3	4.4×10^-2	10.92	4.4 (s)	18.3 GB
DiffSound	Optimization	Single	N/A	0.893	0.890	N/A	N/A	–	190 (ms)	2.0 GB
DiffSound	Optimization	Blend	N/A	0.487	0.696	N/A	N/A	–	190 (ms)	2.0 GB
Ours	Feed-forward	Single	0.081	0.140	0.008	0.59	2.39×10^-8	0.078	6.1 (ms)	1.33 GB
Ours	Feed-forward	Blend	0.119	0.206	0.012	1.08	1.32×10^-8	0.116	6.1 (ms)	1.33 GB

Comparison of physical-parameter estimation methods on ImpactMat. ρ and E are reported in the log-space target scale, ν is linear, and α and β are in physical units. Avg. NMAE is computed over all five normalized target dimensions and is omitted for DiffSound, which estimates only E and ν in our evaluation. Evaluated with K = 1; lower is better.

BibTeX

@article{cho2026impactmat,
  title   = {ImpactMat: Continuous Material Estimation for Inverse Impact Sound Rendering},
  author  = {Cho, Hyebin and Kim, Bumsoo and Chung, Joon Son},
  journal = {arXiv preprint arXiv:2610.07061},
  year    = {2026}
}