---
tags:
  - gaussian-splatting
  - 3d-reconstruction
---

# LonGaussian

Pretrained weights for **LonGaussian: Tracking Foundation Priors for
Long-Context Feedforward Gaussian Reconstruction**.

[Paper](https://arxiv.org/abs/2610.23250) ·
[Code](https://github.com/hustvl/LonGaussian) ·
[Project page](https://abearkeeper.github.io/LonGaussian.github.io/)

## Checkpoint

`stage2b.pt` provides the pretrained Gaussian modules for LonGaussian.
Use it with the standard `merge_mode=neural` reconstruction pipeline.

| File | Size | SHA256 |
| --- | --- | --- |
| `stage2b.pt` | 272,167,298 bytes (272.2 MB / 259.6 MiB) | `9f49fab3604ac174e323c8d74bdc44966e257f3b6be32ef35c338f6fb0bb1feb` |

The file contains the Gaussian decoder, Gaussian attribute head and learned
track merger. Frozen LoGeR, AllTracker and sky-segmentation weights are obtained
separately from their original hosts. This is a PyTorch checkpoint for the
LonGaussian codebase, not a self-contained Transformers model.

## Download and run

Download [stage2b.pt](https://huggingface.co/abkp/LonGaussian/resolve/main/stage2b.pt?download=true)
or use Python:

```python
from huggingface_hub import hf_hub_download

checkpoint = hf_hub_download(
    repo_id="abkp/LonGaussian",
    filename="stage2b.pt",
    local_dir="/path/to/weights",
)
```

Install the code repository's dependencies, then download the frozen models:

```bash
python scripts/download_assets.py weights --output /path/to/weights
```

Run the full reconstruction pipeline using the downloaded checkpoint:

```bash
INPUT=/path/to/calibrated_scene NUM_FRAMES=24 OUTPUT_DIR=/path/to/new_inference \
MODEL_NAME=/path/to/weights/LoGeR_star/latest.pt \
GAUSSIAN_HEAD=/path/to/weights/stage2b.pt \
ALLTRACKER_CHECKPOINT=/path/to/weights/alltracker.pth \
SKY_MASK_ONNX=/path/to/weights/skyseg.onnx \
bash scripts/infer_longaussian.sh
```

Training datasets and Stage-I/IIa checkpoints are not required for inference.
Inputs must follow the repository's calibrated image-sequence format. Camera
poses are predicted by the model. See the repository for data preparation and
the complete evaluation protocol.

## Hardware and scope

CUDA is required by the full reconstruction/rendering pipeline. The validated
reference hardware is an RTX 3090 with 24 GB VRAM. Memory usage depends on
input resolution, sequence length, scene content and runtime settings.

LonGaussian targets static-scene reconstruction. Dynamic objects, inaccurate
camera poses and distant loop revisits may cause artifacts. Frozen third-party
weights are downloaded separately; see the original sources and the code
repository's [third-party notices](https://github.com/hustvl/LonGaussian/blob/main/THIRD_PARTY_LICENSES.md).