README.md

---
tags:
- 3d
- fur
- hair
- strands
- synthetic
pretty_name: PatchLocks data
---

# PatchLocks — data

Data for **PatchLocks**, which reconstructs animal fur as individual strands from multi-view
images with one learned patch prior. Code, documentation and usage guide ([docs/USAGE.md](https://github.com/CCdydy/PatchLocks/blob/main/docs/USAGE.md)):
**[github.com/CCdydy/PatchLocks](https://github.com/CCdydy/PatchLocks)**.
Weights: [zengzzy/PatchLocks](https://huggingface.co/zengzzy/PatchLocks).

The repository layout is a **data root** for the code: download it and link it in.

```bash
# from the code repository: everything, or subsets (training | testsets | animals | figures)
bash scripts/download_data.sh /path/to/DATA_ROOT
bash scripts/download_data.sh /path/to/DATA_ROOT testsets
# or directly with the Hugging Face CLI (pip install huggingface_hub), then link:
hf download zengzzy/PatchLocks-data --repo-type dataset --local-dir /path/to/DATA_ROOT
bash pipeline/00_link_data.sh /path/to/DATA_ROOT
```

| Path | Size | Contents | Used for |
|---|---|---|---|
| `patchlocks/datasets/v11_longfur/renders_mv/` | 13 GB | 1,600 synthetic fur patches (nine coat archetypes), each with 3D strands (`patch.npz`), 4 calibrated 256 px renders, codec latents, field targets and pose features | training; its 80-patch validation subset is Table 1 |
| `patchlocks/experiments/rebuttal_2027/testB_*` | 2.3 GB | frozen in-distribution test set, n=160 | Table 2 (ID), supplementary ceiling table |
| `patchlocks/experiments/rebuttal_2027/ood_*` | 1.1 GB | frozen extrapolation test set, n=80, four beyond-range coats | Table 2 (EXT) |
| `patchlocks/experiments/rebuttal_2027/grid5_B/` | 91 MB | per-predictor patch renders | `fig_baselines_qual` |
| `patchlocks/datasets/v16/v16_visual_gate_agx/` | 67 MB | patch-level input / ground truth / prediction renders | supplementary visual-gate figures |
| `v2_results/refine_v16_fluffy/` | 5 GB | the four reconstructed animals of the paper: per-step outputs (diffusion predictions, character, swatch, fitted parameters) and the final grooms (`<animal>/grow/surfpack_v2fur.npz`, 1.1–1.2M strands × 32 points), plus the renders used by the figures | paper figures, `tab:animals` |
| `v2_results/refine_v16_4animals/` | 5 GB | an earlier deployment of the same four animals | supplementary seam table |
| `v2_results/refine_v16_paper_smooth/global4_champagne/` | 3 MB | renders of the smooth-loft panda variant | supplementary turntables |
| `v2_results/paper_cover_patch/beagle_p855/` | 12 MB | one predicted patch and its renders | supplementary patch-result figure |
| `patchlocks/deploy/nf_official_render/render_*_ours*/` | 30 MB | our side of the NeuralFur comparison renders | Fig. 3, supplementary comparison |

The synthetic patches use seeds 210000+ (training), 610000+ (in-distribution test) and 810000+
(extrapolation); the test sets were generated once and frozen (`*_FROZEN_manifest.txt` in the
code repository). Generation and evaluation protocols: `docs/DATA.md` and `docs/EVALUATION.md`
in the code repository.

**Not included.** The real-animal inputs (multi-view captures, masks, orientation maps, furless
meshes and the surface patch packs preprocessed from them) come from
[NeuralFur](https://github.com/Vanessik/NeuralFur) (CC BY-NC-SA 4.0) and
[ARTEMIS](https://github.com/HaiminLuo/Artemis); obtain them from those projects.
Files that embed their imagery or reconstructions were removed from this release:
photo-overlay diagnostics of the orientation fit, photo-projected albedo, capture photos,
renders of NeuralFur's grooms, and the NeuS / GenZoo baseline renders. Figures that need those
inputs (Fig. 3 and the baseline grid) can only be rebuilt with the third-party data.
The reconstructed grooms and renders of the four real animals (`v2_results/`, `nf_official_render/`) were fitted to those captures and are shared under
[CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/) with attribution to NeuralFur and ARTEMIS. The license of the synthetic data will be stated with the code release.