FurE: Efficient 3D Animal Fur Reconstruction via Root-Conditioned Latents

Executive Summary: The reconstruction of photorealistic, physically plausible, and artistically editable 3D animal fur from multi-view images represents one of the most demanding inverse rendering tasks in computer vision. Unlike human hair modeling, which leverages decades of structured synthetic libraries and constrained cranial topographies, animal fur is characterized by pervasive whole-body distribution, micro-scale strand details, intricate self-occlusion, extreme morphological diversity, and a persistent absence of dedicated 3D training datasets. In a groundbreaking preprint, researchers Srinjay Sarkar, Prakhar Kaushik, Soumava Paul, and Alan Yuille introduce FurE. This architecture realizes an efficient, strand-based animal fur reconstruction pipeline capable of recovering an editable, per-strand groom. By parameterizing strand geometry through a root-conditioned latent field decoded via a PCA-based decoder learned from human hair, and stripping body occlusion through a surface-constrained Gaussian Frosting representation, FurE delivers an exceptional 10x training speedup over dense per-strand optimization baselines while fully preserving strand fidelity across synthetic and real-world captures.

1. The Geometry of Animal Fur: Structural Complexities and the Dataset Deficit

In modern visual computing, neural radiance fields, and photogrammetric pipelines, strand-level fidelity provides dynamic realism that standard polygonal meshes with normal maps or volumetric textures cannot emulate. However, translating human hair reconstruction techniques to wild and domestic mammals presents persistent domain hurdles:

Whole-Body Surface Coverage: Human hair is concentrated predominantly across the scalp, allowing algorithms to rely on standard cranial templates, fixed scalp origins, and predictable grooming patterns. In contrast, animal fur envelops almost the entire epidermal boundary of the subject. A mammalian coat consists of fundamentally heterogeneous strand populations: coarse guard hairs interlock with dense, wavy, or curled undercoats across variable muscular contours. High morphological variability both across different animal species (inter-species) and between subjects of the same breed (intra-species) renders hand-crafted geometric priors largely ineffective.

Micro-Scale Density, Self-Occlusion, and Visual Obfuscation: The sheer numerical density of fur strands creates extreme multi-view visual ambiguity. Dense layers of hair continuously occlude underlying layers, causing conventional multi-view stereo (MVS) algorithms and volumetric neural representations to blur individual fibers into indistinct, cloudy hulls. Extracting discrete, high-frequency, per-strand trajectories from standard RGB camera rays under these severe occlusion conditions is an intrinsically ill-posed inverse problem.

The Complete Absence of Animal-Fur Datasets: High-quality human hair digitizing workflows depend heavily on comprehensive ground-truth databases, including CT scans, synthetic strand simulations, and curated artist assets. For animal fur, no equivalent high-resolution 3D dataset exists. Capturing living animals at micro-strand resolution using structured-light rigs or calibrated multi-camera domes is virtually impossible due to involuntary breathing, muscle twitches, and the dense optical scattering characteristic of mammalian coats.

2. The FurE Framework: Algorithmic Foundations

To overcome the acute deficit of animal-specific 3D training data without sacrificing per-strand control, Sarkar and colleagues engineered FurE around two key innovations: root-conditioned latent parameterization and linear subspace geometric decoding.

Root-Conditioned Latent Neural Fields

Standard per-strand optimization frameworks typically instantiate individual 3D curve parameters or optimize independent coordinate representations across millions of surface points. This unconstrained, dense parameterization leads to computational bottlenecks, massive memory footprints, and optimization times spanning multiple days. FurE eliminates discrete coordinate memorization by introducing a continuous, root-conditioned latent field.

By defining the latent representation as a function of the root follicle coordinates anchored to the underlying body mesh, the network learns an organized spatial manifold of grooming properties. Essential geometric behaviors—including strand length, trajectory, orientation, and curliness—vary smoothly across adjacent surface regions. This formulation enforces spatial consistency: neighboring follicles naturally inherit correlated grooming vectors, preventing unphysical crossing, high-frequency visual jitter, and noise while maintaining precise local editability.

PCA-Based Geometric Decoding via Human-Hair Transfer

Instead of passing latent embeddings through an over-parameterized multi-layer perceptron (MLP) at every evaluation point along a strand, FurE reconstructs complete 3D strand trajectories via a Principal Component Analysis (PCA) decoder. By projecting complex spatial curves into a compact linear subspace, the geometric degrees of freedom are drastically compressed without losing structural expressiveness.

A central finding of the research is that a PCA-based decoder trained strictly on human-hair strand data generalizes directly to animal coats. Although human hair and animal fur differ fundamentally in overall volume and spatial arrangement, the underlying geometric primitives that govern physical strand curvature, bending, and spatial progression share common eigenspaces. By transferring the human-trained PCA decoder to animal subjects, FurE completely bypasses the need for animal-specific 3D training data while replacing expensive non-linear neural evaluations with highly optimized matrix-vector multiplications.

3. Defurred Body Estimation: Surface-Constrained Gaussian Frosting

A critical prerequisite for realistic strand-based reconstruction is accurately determining the underlying epidermal boundary. If the naked skin surface is reconstructed inaccurately, hair roots will either hover unanchored in free space or sink into the interior volume of the animal, invalidating physical grooming simulations.

FurE addresses this challenge through an integrated two-step body recovery process:

• Surface-Constrained Gaussian Frosting: The pipeline adopts a surface-constrained Gaussian Frosting representation that models the fuzzy, volumetric outer boundary of the coat while maintaining rigorous geometric proximity to the underlying surface.
• Extraction of Local Fur-Thickness Cues: By evaluating the spatial differential between the outer volumetric boundary and inner density boundaries, the framework algorithmically extracts local fur-thickness cues across distinct body segments, accurately distinguishing thin facial fur from dense dorsal ruffs.
• Part-Based Priors and Defurring: By combining these local thickness measurements with robust part-based structural priors, FurE computationally "defurrs" the input subject, synthesizing a clean, bare skin surface upon which hair roots are consistently positioned.

4. Empirical Performance, Speedup, and Qualitative Validation

The operational effectiveness of the FurE architecture was validated across diverse synthetic models and unconstrained real-world multi-view video captures:

The 10x Optimization Speedup

State-of-the-art dense per-strand optimization techniques require massive computational budgets, often requiring extensive optimization times for a single capture. FurE achieves a verified 10x training speedup over prior dense per-strand baselines. This major efficiency gain is made possible by the PCA-based linear decoder, which reduces gradient backpropagation overhead during the iterative optimization loop.

Strand Fidelity and Post-Reconstruction Editability

Importantly, the tenfold acceleration in training does not compromise reconstruction fidelity. The recovered fur coats do not deteriorate into blurred volumes or coarse polygonal meshes; instead, they maintain discrete, distinct 3D strand curves. Because the geometry remains fully decoupled into root coordinates and latent vectors, artists and simulation engines can perform direct downstream grooming operations—such as grooming, trimming, combing, and physics-based dynamic simulations.

Robust Generalization Across Real-World Sequences

FurE successfully resolves the domain gap between controlled synthetic environments and in-the-wild video captures. The framework demonstrated reliable hair-strand recovery across various animal morphologies without requiring manual tuning or fine-tuning on animal-specific data, confirming that human-hair geometric priors act as a universal basis for fiber geometry.

5. Strategic Impact on Computer Vision and Digital Production

FurE represents a significant paradigm shift in inverse rendering and visual computing. By demonstrating that cross-domain representations (human hair) can resolve severe domain-specific data scarcity (animal fur), Sarkar and co-authors establish an efficient pathway for digitizing complex fibrous subjects.

For visual effects (VFX) facilities, game development studios, biological preservation initiatives, and spatial computing pipelines, FurE eliminates the labor-intensive requirement of grooming digital animal doubles by hand. High-fidelity, editable, and simulation-ready 3D grooms can now be extracted directly from multi-view imagery in a fraction of previous compute times.

6. Primary Research Record and Formal Citations

Primary Research Abstract:

Realistic and editable animal fur reconstruction from multi-view images is challenging due to fine-scale detail, self-occlusion and obfuscation, and, unlike human hair, the lack of animal-fur datasets. Fur usually covers most of an animal's body, with large inter-species and intra-species variability. We present FurE, an efficient strand-based animal fur reconstruction method that recovers a per-strand, editable groom by optimizing a root-conditioned latent field, decoded into strand geometry via a PCA-based decoder. We reconstruct a defurred animal body using local fur-thickness cues from a surface-constrained Gaussian Frosting representation together with part-based priors. We further show that a PCA-based decoder learned from human-hair strand data can alleviate animal-data scarcity while enabling substantially faster optimization. FurE achieves a 10x speedup in strand training over current SOTA dense per-strand optimization while retaining strand fidelity and generalizing across synthetic and real-world sequences, with quantitative and qualitative validation despite the reduction in training time.

• Contributing Researchers: Srinjay Sarkar, Prakhar Kaushik, Soumava Paul, Alan Yuille
• Submission Date: September 28, 2026
• Permanent Archive Record: arXiv:2609.35770v1 [cs.CV]
• Full Preprint Document: Download Official PDF Document

Standard Reference (APA Format):

Srinjay Sarkar, Prakhar Kaushik, Soumava Paul, and Alan Yuille (2026). FurE: Efficient Instance-Specific 3D Fur Reconstruction without Animal-Fur Datasets. arXiv:2609.35770v1. https://arxiv.org/abs/2609.35770v1

Academic Field: Computer Vision | Archive ID: arXiv:2609.35770v1

BibTeX Entry:

@article{arxiv_2609.35770v1,
  author    = {Srinjay Sarkar and Prakhar Kaushik and Soumava Paul and Alan Yuille},
  title     = {{FurE: Efficient Instance-Specific 3D Fur Reconstruction without Animal-Fur Datasets}},
  journal   = {arXiv preprint arXiv:2609.35770v1},
  year      = {2026},
  url       = {https://arxiv.org/abs/2609.35770v1}
}