README.md

---
license: other
license_name: source-dataset-terms
license_link: https://github.com/NVlabs/ffhq-dataset
pretty_name: Marcus Avatar Data
size_categories:
- 100K<n<1M
tags:
- face
- relighting
- 3dmm
- uv-texture
- webdataset
configs:
- config_name: ffhq1024_rendered
  data_files: FFHQ1024/rendered/*.tar
- config_name: ffhq1024_unwrap_texture
  data_files: FFHQ1024/unwrap_texture/*.tar
- config_name: celebamask_hq_rendered
  data_files: CelebAMask-HQ/rendered/*.tar
- config_name: celebamask_hq_unwrap_texture
  data_files: CelebAMask-HQ/unwrap_texture/*.tar
---

# Marcus Avatar Data

Relit renders and unwrapped UV textures derived from **FFHQ-1024** and
**CelebAMask-HQ**. Everything is stored as uncompressed
[WebDataset](https://github.com/webdataset/webdataset) tar shards; the sample key
is the source dataset's image id (zero-padded for FFHQ, unpadded for CelebA).

## Subsets

| config | samples | shards | ids/shard | size |
|---|---|---|---|---|
| `ffhq1024_rendered` | 69,999 | 140 | 500 | ~755 GB |
| `ffhq1024_unwrap_texture` | 69,999 | 35 | 2000 | ~194 GB |
| `celebamask_hq_rendered` | 29,994 | 60 | 500 | ~330 GB |
| `celebamask_hq_unwrap_texture` | 29,994 | 15 | 2000 | ~83 GB |

Each `<subset>/shard_index.json` lists every shard with its first/last id, plus
any sample whose file count deviates from the expected 11 (rendered) or 9
(unwrap_texture).

## Per-sample contents

`rendered` — 3 HDRI lighting conditions plus one even-lit bake:

| key | description |
|---|---|
| `.render_hdri_{0,1,2}.png` | relit render under HDRI *i* |
| `.baked_hdri_{0,1,2}.png` | baked texture for HDRI *i* |
| `.env_hdri_{0,1,2}.png` | environment map used for HDRI *i* |
| `.baked_even.png` | baked texture under even illumination |
| `.info.json` | render metadata |

`unwrap_texture` — 3DMM fit and UV unwrap:

| key | description |
|---|---|
| `.obj` / `.left_eye.obj` / `.right_eye.obj` | fitted meshes |
| `.uv.png` | unwrapped UV texture |
| `.coeffs.pt` | 3DMM coefficients |
| `.lm68_2d.pt` | 68 2-D landmarks |
| `.trans_params.pt` | crop/align transform |
| `.mask.png`, `.seg_result.png` | face mask and segmentation |

## Loading

```python
from datasets import load_dataset

ds = load_dataset("luh0502/Marcus-avatar-data", "ffhq1024_rendered",
                  split="train", streaming=True)
sample = next(iter(ds))
print(sample["__key__"], sample["render_hdri_0.png"].size)
```

Or with `webdataset` directly:

```python
import webdataset as wds
url = ("pipe:curl -sL -H 'Authorization: Bearer $HF_TOKEN' "
       "https://huggingface.co/datasets/luh0502/Marcus-avatar-data/resolve/main/"
       "FFHQ1024/rendered/rendered-{000000..000139}.tar")
ds = wds.WebDataset(url).decode("pil")
```

## Notes and known gaps

- **FFHQ id `45650` is absent** from both FFHQ subsets — the upstream 3DMM fit
  never produced a texture for it, so nothing was rendered. Hence 69,999 samples
  rather than 70,000.
- **Six CelebAMask-HQ ids are absent** for the same reason: `1011`, `10106`,
  `10108`, `10111`, `10113`, `10215`. Hence 29,994 samples rather than 30,000.
- Renders marked `failed` in the manifests are included as-is and may be
  incomplete — FFHQ `00091`, `21931`, `32511`, `52431` and CelebA `2796`,
  `3511`, `5974` (all but FFHQ `00091` segfaulted). Check the `anomalies` list in
  the relevant `shard_index.json` before using them.
- `<dataset>/rendered/render_results.csv` is the merged per-id status manifest
  (69,999 / 29,994 rows); the raw per-rank logs it was reconciled from are kept
  under `<dataset>/rendered/_logs/`.

## Licensing

This repository contains **derived** data only, not the source images. FFHQ is
distributed under CC BY-NC-SA 4.0 with individual photographs remaining under
their respective Flickr licenses; CelebAMask-HQ is for non-commercial research
use. Both sources' terms carry over — non-commercial research only.