KnockGS

Interaction-Grounded Calibration of Physical Gaussian Representations

Chenchen Ge^*, Hanwen Shen^*, Bowen Jing, Jiyuan Cai, Xiaofeng Wang,
Hongsen Lei, Weitao Zhou, Dandan Zhang, Haibao Yu^†

^*Equal contribution    ^†Corresponding author

arXiv | Paper

KnockGS calibrates the effective elasticity and density scales of a physics-integrated 3D Gaussian asset from its response to a known interaction. The estimated scales are frozen, written back into the same simulator, and evaluated by predicting the response to a different, held-out interaction.

Highlights

• Interaction-grounded calibration. KnockGS turns the response to a known Probe A into physical evidence for estimating elasticity and density scales.
• Simple response-space estimator. A shared deterministic descriptor, candidate-only standardization, hard top-k retrieval, and local ridge regression produce continuous material estimates without MPM gradients.
• Frozen cross-interaction prediction. The estimate is frozen before Probe B and evaluated on interactions that differ in direction, magnitude, or both.
• Object-specific but reusable. The Probe-A response library is built once for a fixed Gaussian asset and simulator contract, then reused across target calibrations.

The candidate-library size is denoted by J and is not fixed by the method. The main paper benchmark uses J = 54, while the supplementary study also evaluates smaller libraries.

Core code (partial release)

The response descriptor and calibration implementation are now available. This small CPU-only release includes five-feature extraction, response retrieval, local/global ridge estimation, a real precomputed Ficus example, and regression tests against archived predictions. It does not include the full simulator, Gaussian assets, or rendering pipeline.

See the code release guide for setup, data conventions, and scope. Quick start from the repository root:

python -m pip install -r requirements.txt
python -m response_calibration --library examples/ficus/library.csv --queries examples/ficus/queries.csv --output predictions.json
python -m unittest discover -s tests -v

Results

Across five held-out Pillow targets, local ridge reduces mean joint scale error to 1.13%, compared with 2.37% for response KNN and 2.45% for global ridge under the same Probe-A evidence. The frozen estimate also yields the lowest trajectory error under the held-out direction- and magnitude-shifted probes.

The rendered comparison follows the same ordering. On the illustrated unseen Probe-B sequence, KnockGS reaches 41.2 dB PSNR and 0.998 SSIM, with the smallest final-frame absolute error.

The same object-specific calibration procedure is evaluated on three geometrically distinct Gaussian assets: Pillow, Ficus, and Vasedeck.

Citation

If you find KnockGS useful, please cite:

@article{ge2026knockgs,
  title         = {KnockGS: Interaction-Grounded Calibration of Physical Gaussian Representations},
  author        = {Ge, Chenchen and Shen, Hanwen and Jing, Bowen and Cai, Jiyuan and Wang, Xiaofeng and Lei, Hongsen and Zhou, Weitao and Zhang, Dandan and Yu, Haibao},
  journal       = {arXiv preprint arXiv:2608.27365},
  year          = {2026},
  eprint        = {2608.27365},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2608.27365}
}

Acknowledgments

KnockGS is built on PhysGaussian, 3D Gaussian Splatting, and Warp-MPM. We thank the authors for releasing their work.