File size: 2,666 Bytes
93826ab | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 | ---
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.
|