FurE reconstructs editable 3D animal fur 10x faster without animal-fur datasets

Srinjay Sarkar · hf · 2026-09-30

FurE is a strand-based 3D animal fur reconstruction method that recovers per-strand, editable fur from multi-view images—without any animal-fur datasets. Highlights:

Optimizes a root-conditioned latent field decoded into strand geometry via a PCA-based decoder trained on human-hair strand data, sidestepping animal-data scarcity.
Reconstructs the defurred body using local fur-thickness cues from a surface-constrained Gaussian Frosting representation plus part-based priors.
Achieves 10x speedup in strand training over SOTA dense per-strand optimization while retaining strand fidelity and generalizing to synthetic and real-world sequences.

Original post →

FurE reconstructs editable 3D animal fur 10x faster without animal-fur datasets

More from Multimodal