# Aligned scene latents at 32³

Publication is complete.
These caches cover 99,642 training views across three datasets.
Each reference voxel retains one token and per-object membership labels.
The `latent_valid` flag marks rows with a native encoder output.
Rows without native output retain their coordinates and zero latent values.

| Dataset | Training views | NPZ data, GB |
| --- | ---: | ---: |
| 3dfront | 39,930 | 4.496 |
| sage | 17,456 | 1.136 |
| foundationpose | 42,256 | 4.948 |

The archives preserve the original files and codec provenance.
Each component becomes available after its `index.json` records `complete=true`.
[Publication status](publication-progress.json) records completion for the full collection.

## Restore

Restore the matching observations before reading scene targets.
The `slat-seg` task downloads those observations without other latent targets.

```bash
python -m preprocess.huggingface download --repo cubec/PartProbe3D-preprocessed --task slat-seg
```

Restore the selected scene caches:

```bash
python -m preprocess.huggingface download --repo cubec/PartProbe3D-preprocessed --dataset 3dfront --bundle-root optional/scene/32 --cleanup
python -m preprocess.huggingface download --repo cubec/PartProbe3D-preprocessed --dataset sage --bundle-root optional/scene/32 --cleanup
python -m preprocess.huggingface download --repo cubec/PartProbe3D-preprocessed --dataset foundationpose --bundle-root optional/scene/32 --cleanup
```

The restored root is `datasets/<dataset>/slat/32/scene_targets/`.
`dataloader.slat_scene.SceneTargetDataset` reads the caches with their matching saved observations.
The reader does not require source meshes.