---
license: cc-by-nc-sa-4.0
pretty_name: Sensor2Sensor precomputed caches (nuScenes 850 scenes)
tags:
  - lidar
  - nuscenes
  - diffusion
  - dinov3
  - sensor-synthesis
extra_gated_prompt: >-
  These are features/latents DERIVED from the nuScenes dataset
  (CC BY-NC-SA 4.0, non-commercial). By requesting access you agree to the
  nuScenes Terms of Use (https://www.nuscenes.org/terms-of-use) and to
  non-commercial use only. No raw nuScenes images or LiDAR points are included,
  and none are recoverable from these lossy encoder outputs.
extra_gated_fields:
  I agree to the nuScenes non-commercial Terms of Use: checkbox
  I will use these caches for non-commercial research only: checkbox
---

# Sensor2Sensor — precomputed training caches (nuScenes, 850 scenes)

Precomputed **frozen-encoder outputs** for training the DINOv3-conditioned LiDAR
diffusion model in **[github.com/skr3178/sensor2sensor](https://github.com/skr3178/sensor2sensor)**
**without downloading the 57 GB raw nuScenes dataset**.

Model weights: **[huggingface.co/sangramrout/sensor2sensor](https://huggingface.co/sangramrout/sensor2sensor)**.

## Contents

Two **matched** caches — 34,149 CAM_FRONT + LIDAR_TOP keyframes (all 850 nuScenes
v1.0-trainval scenes), joined by the `sample_token` filename. Each is a tarball of
per-token `.npz` files.

### `cached_latents_v5_850scenes.tar` (3.3 GB)
| key | shape / dtype | role |
|---|---|---|
| `mu` | `[8,8,256]` f32 | **diffusion target** — v5 LiDAR-VAE latent |
| `raymap` | `[6,32,56]` f32 | conditioning — ray directions |
| `image_latent` | `[4,32,56]` f32 | legacy SD-1.5 image cond (unused by the DINOv3 model) |
| `sample_token` | scalar str | nuScenes sample token |

### `cached_dinov3_v5_850scenes.tar` (7.7 GB)
| key | shape / dtype | role |
|---|---|---|
| `feat` | `[384,14,24]` f16 | **image conditioning** — DINOv3 features (upsampled → 32×56 at train time) |
| `sample_token` | scalar str | nuScenes sample token |

**You need both**: `feat` + `raymap` (conditioning) → predict `mu` (target). The DINOv3
cache has no target; the latents cache has no DINOv3 conditioning.

## Provenance (frozen encoders)

- **Image latent** (`image_latent`): SD-1.5 VAE
- **LiDAR latent** (`mu`): v5 LiDAR VAE @ step 3933 (`lidar_vae_best.pt`, on the model repo)
- **DINOv3** (`feat`): `vit_small_patch16_dinov3.lvd1689m`, input 224×384

Built with `s2s_min/train/cache_latents.py` and `s2s_min/train/cache_dinov3.py` in the
GitHub repo. Subset: `np.random.default_rng(0).choice(850, 850, replace=False)`.

## Usage

```bash
hf download sangramrout/sensor2sensor-cache --repo-type dataset --local-dir ./cache
tar xf ./cache/cached_latents_v5_850scenes.tar -C s2s_min/out/
tar xf ./cache/cached_dinov3_v5_850scenes.tar -C s2s_min/out/
# then train per the GitHub README
```

## License

Derived from **nuScenes** (CC BY-NC-SA 4.0), **non-commercial**, subject to the
[nuScenes Terms of Use](https://www.nuscenes.org/terms-of-use). No raw sensor data is
included or recoverable from these encodings.