FurE: Efficient Instance-Specific 3D Fur Reconstruction without Animal-Fur Datasets

What it claims

FurE reconstructs explicit, editable animal fur strands from calibrated multiview images. On 36 views with an NVIDIA A5000, strand training takes 52 minutes, against 10.5 hours for NeuralFur (Sklyarova et al., 2026). Preprocessing takes 1 hour, against 10 hours for NeuralFur. The authors report comparable or better rendering quality. They describe a real-world bison sequence as, to their knowledge, the first instance-specific, strand-based fur reconstruction from noisy real multiview images. NeuralFur struggles on that sequence. The bison data is released, and the project page is https://toshi2k2.github.io/fure.

How it works

The pipeline has two stages and two inputs to defurring that meet before strands are placed.

flowchart LR
  A[Multiview images + masks] --> B[NeuS2 outer mesh]
  A --> C[Gaussian Frosting shell widths]
  B --> D[3D part labels]
  C --> E[Inward vertex displacement]
  D --> E
  E --> F[Defurred mesh]
  C --> G[Per-part strand length]
  D --> G
  H[Optional VLM length] -.-> G
  F --> I[Root sampling + TBN frame]
  G --> I
  I --> J[Latent codes to PCA decoder]
  J --> K[Cylindrical Gaussians, photometric loss]

• Outer surface. NeuS2 reconstructs the outer coat surface. Body parts (face, ears, mane, belly, legs, paws, tail and others) come from ALIGN-Parts (Paul et al., 2026), with manual verification. This replaces SMAL fitting.
• Defurring. Gaussian Frosting layer widths serve as relative fur-thickness cues. They are not measurements of skin depth. The method solves for smooth inward vertex displacements. Weakly supported regions lean on part-level estimates, and smoothing is weaker across part boundaries. Non-fur regions stay fixed, and offsets that cause face flips or self-intersections are reduced.
• Strand length. Length is set separately from defurring. V1 uses the 75th percentile of each part's shell widths (in cm) times a fixed part-specific multiplier. Optional V2 blends 60% of that value with 40% of a VLM estimate, clamped to 35–135% of the VLM estimate.
• Strand codec. Each root gets a latent code predicted from its positional encoding, part label, normalized Frosting thickness and target length. A PCA decoder initialized from PERM's human-hair basis maps the code to a canonical strand. The strand is then scaled to the target length and oriented by the root's tangent-bitangent-normal frame. The encoder is optimized and the decoder fine-tuned per instance. No animal-fur training set is used. An optional UV texture generator with guide and style maps can produce the codes.
• Rendering. Each strand segment becomes an elongated cylindrical Gaussian with spherical-harmonic appearance. The loss combines RGB, silhouette, image-space orientation, an envelope-attraction term, an SDF anti-penetration term and a fur-mask term.

Evidence

• Dataset. Five synthetic fur styles from Artemis (Luo et al., 2022). Training uses 36 uniformly sampled frames per sequence, and novel views are evaluated on the rest.
• Geometry against ground truth. On a synthetic tiger with artist-made ground-truth strands, strand precision, recall and F-score are reported against GaussianHaircut and NeuralFur.
• Unsupervised geometry. Four Artemis scenes are scored on strand length mean and standard deviation, local and global curvature variance, and local direction variance.
• Rendering. PSNR, LPIPS and SSIM follow the GaussianHair protocol. The authors say FurE "largely outperforms" NeuralFur and Gaussian Haircut.

The fetched text does not include the numeric values of Tables 2–4. Only the runtime table is quantified.

Limits

• Inputs. The method needs calibrated multiview images and foreground masks.
• Part labels. Labels were manually verified, so the pipeline is not fully automatic as described.
• Strand length. Length relies on fixed part-specific multipliers. The authors state that shell thickness is relative evidence, not a strand-length measurement.
• Real-world evaluation. Real-world validation is one bison sequence, and no quantitative metrics are given for it.
• Code and licence. The fetched text states no code release or licence. It mentions only the bison data release.
• Truncation. The fetched text cuts off in the qualitative results. Appendix details on calibration and on the NeuralFur failure case are not visible.