YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
Polymorphic
Compute-efficient video-to-3D research by Modular Computing.
This repository contains the Polymorphic video-to-3D experiments and releases. The current public experiment combines AnySplat for scene/camera initialization with Fourier Splatting for per-scene optimization and novel-view synthesis.
Important: This is an experimental reconstruction pipeline, not a feed-forward model that instantly converts any video into a finished asset. Each new scene requires optimization. The released representation is a renderable splat scene; mesh/GLB conversion is not included as a validated workflow.
Featured experiment: Fourier Splatting
Open the full experiment model card and run instructions →
The experiment directory contains:
- Runnable preparation and evaluation scripts
- Apartment and kitchen scene checkpoints
- PLY exports at Fourier frequency levels K=1–6
- Per-scene metrics, experiment report, and third-party license notices
Held-out novel-view results
Metrics are from two indoor scenes, using the same VGG LPIPS implementation for both methods. Higher PSNR/SSIM is better; lower LPIPS is better.
| Scene | Method | PSNR ↑ | SSIM ↑ | VGG LPIPS ↓ | Primitives |
|---|---|---|---|---|---|
| Apartment | AnySplat baseline | 19.1235 | 0.6612 | 0.3154 | 1,858,896 |
| Apartment | Fourier Splatting K=6 | 20.3469 | 0.7041 | 0.3173 | 99,483 |
| Kitchen | AnySplat baseline | 21.6155 | 0.6660 | 0.2678 | 1,983,229 |
| Kitchen | Fourier Splatting K=6 | 25.3078 | 0.8212 | 0.2586 | 120,000 |
Across these two scenes, Fourier Splatting used approximately 16.5–18.7× fewer primitives than the AnySplat baselines. Apartment LPIPS was slightly worse, so the results do not indicate uniform improvement. This is a small exploratory evaluation, not a generalization or state-of-the-art claim.
Quick links
- Fourier Splatting model card and instructions
- Experiment report
- Metrics JSON
- Generic video-frame preparation script
- Apartment PLY, K=6
- Kitchen PLY, K=6
Reproducing the experiment
Start with the experiment README, which documents dependencies, frame extraction, scene preparation, optimization, rendering, results, and known limitations. The research run was conducted on an NVIDIA H100 80 GB with PyTorch/CUDA 12.9 at 448×448 resolution. The complete workflow has not yet been validated from a clean environment, and dependency compatibility may require adjustments.
Licensing
This is a research release that incorporates upstream AnySplat and Fourier Splatting code. Their respective license terms apply; see the license notices in the experiment directory. In particular, review Fourier Splatting's non-commercial research/evaluation terms before any commercial use.