MoRE-GS

Motion-Regime Experts for Future Gaussian Splatting

Shima Salehi · Joshua Peeples
Texas A&M University · WACV 2027

Project page · Paper · Supplementary

MoRE-GS is a causal, render-aware framework for forecasting a withheld future suffix of a dynamic 3D Gaussian scene. It freezes the Gaussian representation reconstructed from the observed temporal prefix and routes among four complementary control-keypoint motion experts: persistence, a coefficient-space Transformer, the GaussianPrediction graph convolutional network, and a rigidity-aware structured Transformer.

The routing table varies over soft motion groups and forecast-horizon bins. It is calibrated through the frozen renderer on a chronological tail contained entirely inside the observed prefix. Future images are opened only for final evaluation after the experts and router have been frozen and cold-reloaded.

Paper results

The selected 16-group model reports the following scene-unweighted means. The cross-method rows are system-level comparisons using the protocols described in the paper; matched causal controls are reported separately in the ablation.

Dataset	Protocol	PSNR ↑	SSIM ↑	LPIPS ↓
D-NeRF (8 scenes)	chronological 80/20	27.47	0.948	0.045
HyperNeRF (4 scenes)	chronological 90/10	19.22	0.584	0.349

On D-NeRF, MoRE-GS improves mean PSNR by 0.82 dB over the local ODE-GS reproduction. On HyperNeRF, it improves mean PSNR by 0.39 dB and reduces LPIPS by 0.050, while ODE-GS retains higher mean SSIM. Full per-scene values and comparison qualifications are in docs/RESULTS.md.

Repository layout

MoRE-GS/
├── configs/                 selected D-NeRF and HyperNeRF settings
├── docs/                    protocol, data layout, results, provenance
├── moregs/                  experts, grouping, routing, and evaluation
├── patches/                 pinned GaussianPrediction compatibility patch
├── scripts/                 setup, preparation, run, and validation tools
├── CITATION.cff
├── environment.yml
└── pyproject.toml

Datasets, canonical reconstructions, checkpoints, rendered images, caches, and machine-specific experiment directories are intentionally excluded.

Selected configuration

• eight hard groups internal to the structured expert;
• sixteen soft groups used by the group–horizon router;
• four forecast-horizon bins;
• four experts, including persistence;
• a 256-parameter routing table (16 × 4 × 4).

Changing the soft-router group count does not retrain or resize the expert bank. The hard and soft groupings serve different roles and need not have the same cardinality.

Installation

Create the pinned environment, install the GaussianPrediction backbone at the recorded revision, and verify CUDA dependencies:

conda env create -f environment.yml
conda activate moregs
bash scripts/setup_gaussianprediction.sh
python scripts/check_environment.py

The environment targets Python 3.7, PyTorch 1.13.1, and CUDA 11.7. The setup script prints the upstream CUDA-extension installation step because those extensions must match the active CUDA toolchain.

Data

Download D-NeRF or HyperNeRF from their official sources and keep the datasets outside this repository. See docs/DATA_LAYOUT.md.

HyperNeRF scenes must first be converted into an official-training-ID-only chronological view. The preparation script excludes the released validation camera IDs and records an auditable split manifest:

python scripts/prepare_hypernerf.py \
  --source /path/to/original/cut-lemon \
  --output /path/to/prepared/cut-lemon \
  --scene cut-lemon \
  --time-split 0.9

Run one scene

D-NeRF:

PYTHON_BIN=/path/to/python \
bash scripts/run_dnerf_scene.sh \
  /path/to/dnerf/lego \
  output/lego \
  lego \
  17

HyperNeRF:

PYTHON_BIN=/path/to/python \
bash scripts/run_hypernerf_scene.sh \
  /path/to/prepared/cut-lemon \
  output/cut-lemon \
  cut-lemon \
  42

Each runner reconstructs the observed prefix, trains full-prefix and fit-only experts, calibrates the observed-only router, refits the selected 16-group router, cold-reloads the final checkpoint, and writes a JSON summary in the requested output directory.

For exact reproduction from an archived canonical checkpoint:

MOREGS_CANONICAL_CHECKPOINT=/path/to/chkpnt60000.pth \
PYTHON_BIN=/path/to/python \
bash scripts/run_dnerf_scene.sh /path/to/dnerf/lego output/lego lego 17

A newly trained canonical reconstruction is a fresh replication: different keypoint identities and bindings can change the forecasting problem even when the reconstruction quality is similar. Every run therefore records checkpoint, dataset, source, and runtime hashes in moregs_provenance.json.

Release validation

Before publishing or archiving the source release, run:

python scripts/validate_release.py

The audit rejects datasets, checkpoints, large generated artifacts, malformed Python sources, and changes to the selected group and split contracts.

Citation

@inproceedings{salehi2027moregs,
  title     = {{MoRE-GS}: Motion-Regime Experts for Future Gaussian Splatting},
  author    = {Salehi, Shima and Peeples, Joshua},
  booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision},
  year      = {2027}
}

License and acknowledgments

MoRE-GS builds on GaussianPrediction at the revision recorded in patches/GAUSSIANPREDICTION_COMMIT. Upstream repositories and their submodules retain their own licenses. This source release uses the included research-only Gaussian-Splatting license; review all upstream terms before redistribution or commercial use.