FurE uses human-hair priors to cut per-strand fur training from 10.5 hours to 52 minutes and works on a real bison sequence

FurE moves expensive strand learning into a low-dimensional latent space: at each sampled root it predicts compact shape coefficients, which a PCA decoder pretrained on human-hair data maps to local strand geometry, scaled to target length and rendered with strand-aligned cylindrical Gaussians. Prior SOTA such as NeuralFur performs dense per-strand optimization in full 3D strand space; FurE instead optimizes root-conditioned latent codes while keeping an explicit, editable strand representation. The paper reports training on five synthetic fur styles from the Artemis dataset, with strand training at 52 minutes and preprocessing at 1 hour versus NeuralFur's 10.5 hours and 10 hours; on a synthetic tiger asset with artist-generated ground-truth strands, FurE's precision/recall/F-score exceed GaussianHairCut and NeuralFur at the 2/20, 3/30 and 4/40 thresholds, for example F-score 50.86 versus 46.84 and 37.93 at 4/40.

The paper frames defurring, estimating the skin beneath the fur, as a local shell-estimation problem: Gaussian Frosting shell widths serve as local fur-thickness cues, calibrated with part-level priors, and smooth inward vertex displacements yield the strand-root mesh without SMAL fitting. NeuralFur's defurring depends on template fitting and part-level semantic estimates; FurE instead uses local Frosting cues plus part-aware priors and direct part segmentation via ALIGN-Parts, removing the SMAL-fitting dependency. The paper states that shell width is not a direct measurement of skin depth or strand length, hence calibration with part-level references and greater reliance on part estimates in weakly supported regions; ablations show that removing the defurred mesh gives visually similar results but inconsistent direction and curvature metrics, while fixed length clearly fails on the body and belly where strands are naturally longer.

The paper shows human-hair priors can transfer to animal fur, sidestepping the lack of animal-fur training data while keeping reconstruction instance-specific: each target animal is still optimized and fine-tuned on its own multi-view images. Human hair and animal fur differ in length, density and growth direction, and animal-fur datasets were previously unavailable; FurE initializes its decoder from PERM's human-hair PCA basis and treats that prior as a transferable compact representation. The paper ablates fixed length, shape prior only (no defurred mesh) and the full method, and reports unsupervised geometric metrics on four Artemis synthetic scenes covering length consistency, curvature and orientation, plus PSNR/LPIPS/SSIM rendering metrics, with FurE slightly ahead of NeuralFur and GaussianHairCut on most metrics.

The paper reports what it describes as the first instance-specific, strand-based animal fur reconstruction from noisy real-world multi-view images, and releases the associated multi-view data. Earlier work was validated mainly on synthetic data; the paper notes NeuralFur fails to capture the underlying geometric surface on this real bison sequence and therefore cannot reconstruct coherent fur. The evidence is qualitative results and comparisons on a real bison sequence, with the paper explaining that the underlying-geometry failure blocks fur reconstruction; the authors also note that capturing dense static multi-view images of living animals is itself difficult.

The result targets static animal instances with calibrated multi-view images and an available metric scale reference: in that setting FurE completes strand training in under an hour and outputs explicit, editable strands that import directly into engines such as Blender and Unreal Engine and support physics simulation such as wind. For researchers and content creators who want to avoid animal-fur datasets while still producing instance-specific fur assets, it offers a path that substitutes human-hair priors for animal data and low-dimensional latent codes for dense per-strand optimization; the paper also releases the real bison multi-view data so real-sample validation can continue.

The boundaries the paper itself notes are worth watching: evaluation lacks ground-truth cases with strong local variation within body parts, such as shaved patches, injuries or irregular grooming, which are exactly where local Frosting cues and part-level priors would be tested against each other; FurE represents fur as a single strand layer on one root surface estimated before strand optimization, and multilayer coats plus joint optimization of root placement and strand geometry are listed as next steps. Defurring also relies on initial Frosting cues and dense camera viewpoints, so severe self-occlusion can degrade shell thickness and fur-length estimates; centimeter-valued length initialization requires a metric reference, and the paper states there is no validated automatic fallback. Real bison results are mainly qualitative, and reproducibility depends on the released multi-view data. This summary is based on the paper's full text and abstract and does not verify every figure or table value in the appendix.