Find the root surface.
NeuS2 reconstructs the visible coat. Local Frosting shell widths and verified body parts guide a bounded, smooth estimate of the hidden surface where strands attach.
Johns Hopkins University Computer Vision Research
Efficient instance-specific 3D fur reconstruction
without animal-fur datasets.
Reconstructed geometry. Individual
strands.
Six animals. Over 1.26 million rendered strands.Explicit · Editable · Animatable
01 The idea
Not just the look of fur.
The geometry behind it.
Fur is more than a surface. It is hundreds of thousands of fine strands, each with its own shape, direction, and attachment to the body. FurE recovers this explicit structure from calibrated multiview images.
We combine local Gaussian Frosting cues with a compact PCA-based strand decoder initialized from human-hair data. The result is an instance-specific, editable groom, without requiring a separate animal-fur training dataset.
02 The collection
A closer look, strand by strand.
Six animals. Distinct coats.
One explicit representation.
01 / 06
Full-density render · original teaser palette
Full density: 15–44 MB per animal. Desktop recommended.
Artemis
Cat
Fine strands, a full coat, and a distinct silhouette. Inspect the explicit geometry beneath the studio rendering.
310,000source strands
Studio images retain all source strands, with the approved illustrative teaser colors. Explore 3D starts with a lightweight sample; full density loads every original teaser strand, without interpolation. The body stays simplified. Browser lines do not reproduce studio hair shading.
03 How it works
From photographs
to fur you can work with.
Learn less. Reconstruct more.
A root-conditioned latent field predicts compact shape coefficients. A PCA decoder initialized from human hair turns them into explicit strands, scaled and oriented at their roots.
Bring every view together.
Strand-aligned Gaussians make the groom differentiably renderable. Optimizing against the input views refines the instance-specific strand representation.
A local cue, not a direct fur-length measurement
Frosting shell width indicates relative volumetric support around the surface. We calibrate it with part-level priors rather than treating it as measured skin depth or strand arc length. Root displacement and strand-length initialization are separate estimates.
04 Efficiency
Compact codes.
Shorter training.
Optimize a small representation of strand shape instead of every 3D point independently.
NeuralFur10.5 hours
FurE52 minutes
Strand-training times reported in the current manuscript.
36
views · NVIDIA RTX A5000 · excludes preprocessing.
05 Beyond the reconstruction
A little wind.
A world of detail.
Explicit strands become a groom
you can animate and art-direct.
265,000 strands. All in motion.
Root-pinned strands respond to changing wind intensity and direction. This is a downstream wind simulation of the reconstructed asset, not motion recovered from the input photographs.
Download 4K film ↓
06 Beyond synthetic scenes
Into the
real world.
Real captures bring clutter, uneven views, and complex coats. The bison sequence explores FurE beyond synthetic animals, connecting multiview observations to an explicit, editable asset.
07 Reference
Build on FurE.
Publication details are coming soon.
This BibTeX entry is a
placeholder.
BibTeX Placeholder
@misc{sarkar2026fureefficientinstancespecific3d,
title={FurE: Efficient Instance-Specific 3D Fur Reconstruction without Animal-Fur Datasets},
author={Srinjay Sarkar and Prakhar Kaushik and Soumava Paul and Alan Yuille},
year={2026},
eprint={2609.35770},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2609.35770},
}