---
tags:
  - gaussian-splatting
  - novel-view-synthesis
  - dynamic-scene-reconstruction
  - pytorch
---

# Mango-GS Weights

Pretrained inference weights for
[Mango-GS](https://github.com/htx0601/Mango-GS). See the
[paper](https://arxiv.org/abs/2603.11543) for the method and evaluation
protocol.

## Download

Download the repository directly into the public code's `weights/` folder:

```bash
hf download htx0601/Mango-GS --local-dir weights
```

Each scene directory contains `cfg_args`, `point_cloud.ply`, and `deform.pth`.
The release scripts load these files directly without a checkpoint argument.

## Inference Parameters

| Dataset | Scenes | T | K | Nodes | Resolution |
|---|---|---:|---:|---:|---:|
| N3V | coffee_martini, flame_salmon_1 | 4 | 3 | 2048 | 2 |
| N3V | cook_spinach, cut_roasted_beef, flame_steak, sear_steak | 4 | 5 | 4096 | 2 |
| HyperNeRF | broom2, vrig-3dprinter, vrig-chicken | 6 | 3 | 2048 | 2 |
| HyperNeRF | vrig-peel-banana | 8 | 3 | 4096 | 2 |

The same machine-readable parameters are provided in `manifest.json` and in
each scene's `cfg_args`.

## Citation

```bibtex
@inproceedings{huang2026mangogs,
  title     = {Mango-GS: Enhancing Spatio-Temporal Consistency in Dynamic Scenes Reconstruction using Multi-Frame Node-Guided 4D Gaussian Splatting},
  author    = {Huang, Tingxuan and Zhu, Haowei and Yong, Jun-hai and Pan, Hao and Wang, Bin},
  booktitle = {International Conference on Learning Representations},
  year      = {2026}
}
```