---
tags:
- 3d
- fur
- hair
- diffusion
- strands
library_name: pytorch
---

# PatchLocks — model weights

Weights 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)**.
Data: [zengzzy/PatchLocks-data](https://huggingface.co/datasets/zengzzy/PatchLocks-data).

Files keep the paths the code expects. From a clone of the code repository:

```bash
bash scripts/download_weights.sh            # curl + sha256 check, no extra dependencies
# or with the Hugging Face CLI (pip install huggingface_hub):
hf download zengzzy/PatchLocks --local-dir . --exclude "README.md" ".gitattributes"
```

| File | Size | Used for |
|---|---|---|
| `patchlocks/datasets/v16/e2a_cw15/v14_best.pt` | 148 MB | paper diffusion model: multi-view EDM over 64-channel strand latents with clump-channel weighting (all whole-animal results, Table 1 "ours") |
| `patchlocks/datasets/v16/e2a_cw15/b3_fuser.pt` | 8.6 MB | multi-view fuser paired with the paper model |
| `patchlocks/models/fur_strand_codec_v6upright.pt` | 6.8 MB | strand codec (z64 ↔ strand), fine-tuned from DiffLocks |
| `patchlocks/v2_locked/V10_PATCH_STABLE/stage3_model.pt`, `b3_fuser.pt` | 72 MB | single-view baseline (Table 1 "1-view") |
| `patchlocks/datasets/v14/e1_full_sm80/v14_best.pt`, `b3_fuser.pt` | 157 MB | recipe ablation without clump-channel weighting (supplementary recipe table) |
| `patchlocks/experiments/rebuttal_2027/det_z64/det_z64_best.pt` | 152 MB | deterministic-latent baseline (Table 2) |
| `patchlocks/experiments/rebuttal_2027/direct7/direct7_best.pt` | 6.5 MB | scalar-regression baseline (Table 2) |

Training: 1,600 synthetic fur patches × 4 views (nine coat archetypes), seed 0, ~30 min on one
NVIDIA RTX PRO 6000 (`pipeline/02_train_diffusion.sh`). Re-running the released checkpoint
reproduces the paper's diffusion predictions bit for bit and every paper table; see
`docs/REPRODUCIBILITY.md` in the code repository.

The strand codec architecture and its initial weights come from
[DiffLocks](https://difflocks.is.tue.mpg.de/). The conditioning backbone is DINOv2, loaded
through `torch.hub`.