diffusion_pet: v18 fur strand prior

These are the weights for v18, a geometry-conditioned diffusion prior for animal fur strands. The code, data generator and usage instructions are at https://github.com/CCdydy/diffusion_pet.

What the model does

The body is split into small surface patches. For each patch, v18 samples the strand shapes that grow there, along a given fur-flow direction: how the fur rises, droops and clumps. Each sample is conditioned on:

  • Image features. Multi-view DINO features of the patch, fused by the v17 fuser.
  • Local skin geometry. A per-texel height map, normal map and the world up direction, in the patch frame.

The geometry condition enters the v17 UNet through a zero-initialised input layer.

Files

Path Size Use
v18/fuser.pt 35 MB Multi-view fuser and parameter heads (v18 stage A)
v18/v18_best.pt 80 MB UNet18 (EMA weights)
v17_r4b/fuser.pt 35 MB v17 r4b fuser: the initialisation for training v18
v17_r4b/v17_best.pt 80 MB v17 r4b UNet: the initialisation for training v18
  • Inference needs only v18/: pass --geo_dir v18 --model v18/v18_best.pt to diag/v17/infer_v18_surfpack_proposals.py, or use scripts/apply_v18.sh.
  • Training (scripts/train_v18.sh) starts from v17_r4b/.

Checksums are in SHA256SUMS.

Training data

All training data is synthetic. Bodies are random dogs from the public SMAL 39-dog shape space; no fitted parameters of real animals are used. Synthetic grooms were grown on these bodies and rendered from 58 cameras. The v17 base model was trained on procedurally generated flat fur patches.

Split Bodies Dogs Patches
Training (incl. validation) 0–89 135 8,640
Validation body % 10 == 7 β€” 576
Test 90–99 20 1,280

The released checkpoint is from step 16,000 of stage C.

Results on held-out synthetic test bodies

Values are medians per patch over 1,280 patches.

  • Shape: resampled point error divided by arc length (lower is better).
  • Yaw: azimuth error in degrees.
Method Shape ↓ Yaw ↓
v17 (flat-patch prior) 0.430 54Β°
Procedural growth, same parameters 0.220 16Β°
v18 without geometry 0.181 13Β°
v18 0.163 11Β°

Licence

Not yet decided.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support