---
language: en
tags:
- pytorch
- autonomous-driving
- occupancy-prediction
- lidar
---

# LO-World

Pretrained checkpoints for joint future 3D occupancy and LiDAR prediction in a shared BEV latent space, with a GenieDrive transformer for six-camera video generation.

## Checkpoints

| File | Model |
| --- | --- |
| `bev_world_6f.pth` | 6-frame nuScenes World Model |
| `bev_world_12f_autoregressive.pth` | 12-frame autoregressive World Model |
| `bev_world_12f_layout_autoregressive.pth` | 12-frame autoregressive World Model with BEVLayout conditioning |
| `bev_world_6f_stc.pth` | 6-frame World Model fine-tuned with STCOcc inputs |
| `bev_world_6f_waymo.pth` | 6-frame Waymo World Model |
| `occ_lidar_vae.pth` | Joint occupancy-LiDAR tokenizer |
| [`geniedrive_448x800.safetensors`](geniedrive_448x800.safetensors) | GenieDrive transformer for 448 × 800 six-camera video generation; requires the Wan base model (see [Video generation](#video-generation)) |

Run the following from the LO-World code repository with `huggingface_hub` installed in your Conda environment:

```bash
hf download JimmyLiaoo/LO-World \
  bev_world_6f.pth bev_world_12f_autoregressive.pth \
  bev_world_12f_layout_autoregressive.pth bev_world_6f_stc.pth \
  bev_world_6f_waymo.pth occ_lidar_vae.pth \
  geniedrive_448x800.safetensors SHA256SUMS --local-dir ckpts
(cd ckpts && sha256sum -c SHA256SUMS)
```

Weights are saved to `ckpts/`. Use the matching configuration and data preparation instructions in the code repository. World Model checkpoints include tokenizer parameters.

## Video generation

The released GenieDrive transformer generates six-camera videos at 448 × 800. Install the `loworld-genie` environment and prepare scene RGB/semantic conditions using the LO-World code repository.

Download the transformer and its required Wan base model from the repository root:

```bash
hf download JimmyLiaoo/LO-World geniedrive_448x800.safetensors --local-dir ckpts
hf download alibaba-pai/Wan2.1-Fun-V1.1-1.3B-Control \
  --local-dir ckpts/geniedrive/Wan2.1-Fun-V1.1-1.3B-Control
```

The transformer is saved to `ckpts/geniedrive_448x800.safetensors`, matching the default inference path. Generate videos for a prepared scene:

```bash
conda run --no-capture-output -n loworld-genie python \
  tools/genie/inference/generate_scene.py \
  --scene-dir work_dirs/visualization/scene_0001
```

Output: six camera MP4s and a surround-view GIF under `work_dirs/visualization/scene_0001/video/`.

## Data downloads

Download nuScenes train/validation annotation infos to `data/nus_infos/`:

```bash
hf download JimmyLiaoo/LO-World --include "nus_infos/*.pkl" --local-dir data
```

Download and extract 2 Hz BEVLayout arrays to `data/bev_layout/` for Layout-conditioned models. Extraction requires `tar` with Zstandard support:

```bash
hf download JimmyLiaoo/LO-World \
  data/bev_layout/2hz/train.tar.zst data/bev_layout/2hz/val.tar.zst --local-dir .
tar --zstd -xf data/bev_layout/2hz/train.tar.zst -C data/bev_layout
tar --zstd -xf data/bev_layout/2hz/val.tar.zst -C data/bev_layout
```

Optional 12 Hz condition packages provide one 81-frame clip per scene for 700 training and 150 validation scenes, sampled at uniform physical times:

```bash
hf download JimmyLiaoo/LO-World --include "data/bev_layout/12hz_*_81f/**" --local-dir .
```

Outputs: `data/bev_layout/12hz_train700_81f/` and `data/bev_layout/12hz_val150_81f/`. See the [training](data/bev_layout/12hz_train700_81f/README.md) and [validation](data/bev_layout/12hz_val150_81f/README.md) package cards for usage.

Original nuScenes sensor data and Occ3D labels are required separately. STCOcc models also require [STCOcc predictions](https://drive.google.com/file/d/1dXB9mtROLWChycBZlhYIf_JBLshXogBs/view?usp=drive_link) encoded with the released tokenizer.

BEVLayout data is derived from nuScenes and is for non-commercial research under the [nuScenes terms](https://www.nuscenes.org/terms-of-use). See the [data card](data/bev_layout/README.md) for details.