VAST-AI/AniGen-Sample-Dataset

---
license: mit
pretty_name: AniGen Sample Data
size_categories:
- n<1K
tags:
- 3d
- image
configs:
- config_name: default
default: true
data_files:
- split: train
path: samples.csv
---

# AniGen Sample Data

This directory is a compact example subset of the AniGen training dataset.

## What Is Included

- 10 examples
- 10 unique raw assets
- Full cross-modal files for each example
- A subset `metadata.csv` with 10 rows

The retained directory layout follows the core structure of the reference test set:

```text
raw/
renders/
renders_cond/
skeleton/
voxels/
features/
metadata.csv
statistics.txt

latents/ (encoded by the trained slat auto-encoder)
ss_latents/ (encoded by the trained ss auto-encoder)
```

## How To Read One Example

Each row in `metadata.csv` corresponds to one example identifier in the `sha256` column. In practice this value is the sample key used across modalities.

For a row with sample key `<file_identifier>`:

- raw asset: `local_path` field, for example `raw/<raw_file>`
- rendered views: `renders/<file_identifier>/`
- conditional rendered views: `renders_cond/<file_identifier>/`
- skeleton files: `skeleton/<file_identifier>/`
- voxel files: `voxels/<file_identifier>.ply` and `voxels/<file_identifier>_skeleton.ply`
- image feature: `features/dinov2_vitl14_reg/<file_identifier>.npz`
- mesh latents: files under `latents/*/<file_identifier>.npz`
- structure latents: files under `ss_latents/*/<file_identifier>.npz`