---
license: mit
library_name: pytorch
tags:
  - depth-estimation
  - geometry
  - world-model
  - video
---

# GeoState

Inference checkpoints for **GeoState: Learning View-Decoupled Geometry States for 3D World Modeling**.

[Project page](https://hongchang-liu.github.io/GeoState-project-page/)

GeoState represents a scene using a compact, view-decoupled geometry state. A world model predicts future states from four historical states, and a camera-conditioned geometry decoder produces depth maps from a specified viewpoint.

## Checkpoints

| File | Contents | Approximate size |
| --- | --- | --- |
| `geostate.pt` | Scene Tokenizer and Geometry Decoder; frozen VGGT-Ω weights are excluded | 824 MB |
| `world_model.pt` | A-best residual World Model, using four historical 512 × 512 states | 1.03 GB |
| `camera_predictor.pt` | Camera Predictor, including its motion backbone and pose statistics | 36 MB |


## Download and use

From the GeoState source repository root, download the three files into `checkpoints/`:

```bash
hf download HongchangLiu/GeoState \
  geostate.pt world_model.pt camera_predictor.pt CHECKSUMS.json SHA256SUMS \
  --local-dir checkpoints
```

Verify the downloads:

```bash
cd checkpoints
sha256sum -c SHA256SUMS
```

Use the GeoState source repository's installation instructions, `configs/geostate.yaml`, and the evaluation entry points `scripts/eval_waymo.py`, `scripts/eval_kitti.py`, and `scripts/eval_nuscenes.py`. Each evaluation script accepts:

```text
--scene-checkpoint checkpoints/geostate.pt
--world-checkpoint checkpoints/world_model.pt
--camera-checkpoint checkpoints/camera_predictor.pt
```


## Required upstream dependency

The frozen **VGGT-Ω 1B 512** backbone and DPT head are required separately and are not redistributed in this repository. Obtain the unmodified weights from the [official VGGT-Ω model repository](https://huggingface.co/facebook/VGGT-Omega), and install the [official source](https://github.com/facebookresearch/vggt-omega) at commit:

```text
39a0cb8af88554f15ddcb5354cd52bde588fa014
```

The GeoState configuration expects the weight file at `checkpoints/vggt_omega_1b_512.pt`. Alternatively, set `VGGT_OMEGA_CHECKPOINT_OVERRIDE` to the local weight path and `VGGT_OMEGA_IMPORT_ROOT_OVERRIDE` to the source checkout. The expected upstream weight SHA-256 is:

```text
c02da418b18bb01d0392598d3f6147366bcde1bb70fd08a5e3bf7925b0667934
```

## License

The GeoState artifacts in this repository are released under the MIT License; see `LICENSE`. Separately obtained VGGT-Ω code and weights remain subject to their upstream FAIR Noncommercial Research License and Acceptable Use Policy.