💻 computer science

GeomHair: Reconstruction of Hair Strands from Colorless 3D Scans

The paper presents GeomHair, a novel method that reconstructs detailed hair strands directly from colorless 3D scans by combining multi-modal orientation extraction with a diffusion prior, while also releasing the Strands400 dataset to support data-driven generative models for digital avatar synthesis.

Original authors: Rachmadio Noval Lazuardi, Artem Sevastopolsky, Egor Zakharov, Matthias Niessner, Vanessa Sklyarova

Published 2026-05-14

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Rachmadio Noval Lazuardi, Artem Sevastopolsky, Egor Zakharov, Matthias Niessner, Vanessa Sklyarova

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a 3D scan of a person's head, but it's like a statue made of smooth, gray clay. It has the shape of the head and the bumps of the hair, but it's completely colorless and lacks the fine, individual strands that make hair look real. For a long time, computer scientists have struggled to turn this smooth "clay" into thousands of tiny, realistic hair strands, especially without color information to help guide them.

GeomHair is a new method that solves this puzzle. Think of it as a digital sculptor that can look at a smooth, gray 3D head and say, "I know exactly where every single hair strand should go, even though I can't see any color."

Here is how it works, broken down into simple steps:

1. The "Detective" Phase: Finding Clues in the Shape

Usually, to figure out which way hair is flowing, computers look at color and light (like how a highlight on a shiny hair strand tells you its direction). But since this scan has no color, GeomHair uses two other "detective tools" to find clues in the shape itself:

The Ridge Finder (3D Clues): Imagine running your finger over a mountain range. You feel the sharp ridges and the deep valleys. The computer does the same thing on the 3D scan. It looks for the "crests" (the highest ridges) and "ravines" (the deepest dips) on the surface of the hair. These sharp curves act like a roadmap, telling the computer the general direction the hair is flowing.
The Shadow Reader (2D Clues): The computer then takes a picture of the gray 3D scan, lighting it up from different angles to create shadows. It uses a smart AI (called a neural detector) to look at these shadows and edges, much like how you can tell the direction of grass in a field just by looking at the shadows it casts. This helps catch the finer details that the 3D ridges might miss.

By combining the "ridge map" and the "shadow map," the computer gets a very clear picture of how the hair is supposed to flow.

2. The "Artist" Phase: Drawing the Strands

Once the computer knows the direction, it needs to actually draw the hair. But drawing millions of strands from scratch is hard; they might look messy or unnatural.

To fix this, GeomHair uses a Diffusion Prior. Think of this as a "style guide" or a "memory bank" of thousands of real hairstyles the computer has studied before.

The computer asks a smart AI (a Vision-Language model) to look at the 3D scan and describe the hair in words (e.g., "wavy," "short," "messy bun").
It then uses these words to "prompt" the style guide. The guide says, "Okay, based on this description and the shape we see, here is how real hair usually behaves."
The computer then "denoises" the hair, slowly refining a messy cloud of strands into a clean, realistic hairstyle that fits the head perfectly.

3. The Result: A New Library of Hair

The authors didn't just build the tool; they used it to create Strands400.

Imagine a library where, instead of books, you have 400 different 3D heads, each with thousands of perfectly reconstructed hair strands.
Before this, most hair data was either made by hand (which is slow and expensive) or came from expensive, multi-camera setups that required perfect lighting and color.
Strands400 is the largest collection of its kind, built entirely from "colorless" 3D scans. It proves that you don't need color to get high-quality hair; you just need the right shape and the right AI tools.

Why This Matters (According to the Paper)

The paper highlights three main wins:

It works on "Gray" Data: It can turn simple, colorless 3D scans (from handheld scanners or video games) into realistic hair.
It helps Artists: If a game designer or movie artist makes a hair "mesh" (a blocky shape) and wants to turn it into realistic strands for simulation, this tool can do it automatically.
It teaches AI: Because they created the Strands400 dataset, other AI models can now learn from real-world hair data to generate new hairstyles from text or images, without needing to be trained on fake, computer-generated hair.

In short, GeomHair is like a magic translator that turns a smooth, featureless 3D head into a detailed, realistic hairstyle by reading the "topography" of the hair and consulting a library of how real hair behaves.

Technical Summary: GeomHair

Problem Statement

Reconstructing individual hair strands from 3D data is a critical challenge for high-fidelity digital avatars, animation, and AR/VR applications. While physically based rendering and motion simulation require explicit strand-level representations, existing methods face significant limitations:

Dependency on Color: Most state-of-the-art reconstruction techniques rely heavily on high-quality, multi-view color (RGB) information to infer strand orientation. They often fail or produce noisy results when applied to colorless 3D geometry, such as scans from structured-light sensors or designer mesh assets.
Data Scarcity: There is a lack of large-scale, diverse datasets of real-world hair strands. Existing datasets rely on costly handcrafted parametric models or extensive augmentation, limiting the training of data-driven generative models.
Workflow Gaps: 3D artists often model hair as meshes (using grooming guides) but lack automated tools to convert these meshes into dense, strand-level representations required for simulation.

Methodology: GeomHair

The authors propose GeomHair, the first method to reconstruct hair strands directly from colorless 3D scan geometry. The approach operates in two primary stages: Orientation Extraction and Strands-based Reconstruction.

1. Orientation Extraction

To overcome the lack of color information, GeomHair combines two complementary orientation signals:

3D Geometric Cues (Crest Lines): The method extracts salient surface features directly from the mesh geometry. Using local cubic polynomial fitting at mesh vertices, it identifies "crest lines" where principal curvatures reach extrema. These lines capture the sharp bends and curves characteristic of hair, providing a robust 3D directional field independent of texture or lighting.
2D Neural Cues (TEED): The method renders the colorless scan from multiple viewpoints using a forward-facing point light to highlight surface details. It then applies TEED (Tiny and Efficient Edge Detection), a neural orientation detector, to these grayscale renderings. The resulting 2D edge maps are skeletonized, converted to graph representations, and lifted back into 3D space to form a directional field.
Fusion: These two signals are combined to create a comprehensive orientation field that guides strand growth, balancing the structural coherence of 3D geometry with the fine-grained detail of 2D edge detection.

2. Strands-based Reconstruction

The reconstruction process optimizes a latent geometry texture to generate 3D polylines representing hair strands. The optimization is guided by a multi-term loss function:

Orientation Loss ( $L_{orient}$ ): Ensures generated strands align with the fused 3D+2D orientation field.
Volume Loss ( $L_{volume}$ ): Utilizes an Unsigned Distance Function (UDF) approximator trained on the input hair mesh. This constrains strands to the surface volume, accommodating non-watertight meshes common in real-world scans.
Coverage Loss ( $L_{chamfer}$ ): A one-way Chamfer distance ensures uniform coverage of the visible scalp surface.
Diffusion Prior ( $L_{prior}$ ): To enhance realism and structural plausibility, the method incorporates a conditional diffusion prior (based on the HAAR model). This prior is conditioned on text embeddings generated by a Vision-Language Model (VQA) that analyzes the rendered shading of the input scan. The prior is applied via a two-phase noise schedule (Karras discretization) to refine the latent representation.

The final objective function combines these geometric constraints with the diffusion-based regularization to produce high-fidelity strand reconstructions.

Key Contributions

GeomHair Method: The first framework capable of reconstructing hair strands directly from colorless 3D scans, eliminating the dependency on RGB data.
Strands400 Dataset: The creation of the largest publicly available real-world strand dataset, comprising 400 subjects. This dataset is curated from high-quality scans (NPHM and new collections) and addresses the data bottleneck for training generative hair models.
Practical Workflow Integration: Demonstration of the method's utility in converting designer mesh assets (e.g., from Houdini or 3ds Max) and generated meshes (from text-to-mesh or image-to-mesh pipelines) into dense strand representations.
Superior Orientation Detection: Validation that combining 3D crest lines with neural 2D edge detection (TEED) outperforms traditional Gabor filters, particularly in scenarios lacking color or under challenging lighting conditions.

Results and Evaluation

Qualitative Performance: Ablation studies confirm that the combination of 3D and 2D orientation signals yields the most coherent and detailed strands. Removing either component degrades the result (3D alone produces overly smooth strands; 2D alone produces noisy, incoherent structures).
Comparison to SOTA: When compared to methods relying solely on RGB supervision (e.g., Gaussian Haircut) or generative baselines without per-scene optimization (HAAR), GeomHair achieves higher fidelity to the input geometry while maintaining realistic strand structure.
Dataset Validation: Perceptual evaluations using GPT-4V indicate that hairstyles generated from the Strands400 dataset are preferred over those from existing synthetic datasets (Perm) in 74.4% of cases.
Downstream Utility: Retraining the HAAR generative model on Strands400 results in more realistic generation of specific hairstyles (e.g., ponytails, wavy textures) compared to training on the original synthetic data.

Significance and Claims

The paper claims that GeomHair unlocks scalable dataset creation from existing 3D scan collections and designer assets, which were previously unusable for strand-level reconstruction due to the lack of color data. By enabling the extraction of strands from geometry alone, the method facilitates:

The creation of Strands400, a foundational resource for training data-driven generative models for tasks like image-to-strands and text-to-strands.
A practical pipeline for CG artists to convert mesh-based hair assets into simulation-ready strand representations.
The ability to generate hair strands from simple input modalities (text or images) via intermediate mesh generation, purely through geometric processing.

The authors acknowledge limitations, noting that the method struggles with very short hairstyles (insufficient curvature details) and extremely curly hair (where high-frequency details are lost in the scan, leading to noise in orientation estimation). However, the work establishes a new paradigm for geometry-first hair reconstruction.