Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
DGH: Dynamic Gaussian Hair
Authors: Junying Wang, Yuanlu Xu, Edith Tretschk, Ziyan Wang, Anastasia Ianina, Aljaz Bozic, Ulrich Neumann, Tony Tung
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 4 Experiments Evaluation. We evaluate our method on deformation and appearance, comparing it with baselines across 5 hair subjects using our synthetic hair dataset. Each groom s training dataset consists of 90 motion sequences, while testing is performed on the remaining 10 sequences. For dynamic appearance, we assess unseen hair motions (100 frames per subject) and novel views, capturing 100 views via horizontal camera rotation. See the Appendices for more dataset and evaluation details. Metrics. We evaluate the quality of rendered images via reconstruction fidelity (PSNR [54]), local structural similarity (SSIM [54]), and perceptual similarity (LPIPS [52]) between the synthesized images and the ground truth. We evaluate per-frame motion via the L2 error, Chamfer distance [55], between our dense hair point cloud and the ground truth. We evaluate the temporal consistency of the hair motion via the L2 error between the flow vectors in the predicted and ground-truth point cloud sequences, and we show the flow error of different settings in Fig. 7 Baselines. Our framework learns dynamic hair motion and time-varying appearance using Gaussians in a differentiable manner. We compare against 3DGS [34] and Gaussian Haircut [9], retraining both on our synthetic dataset for each static hairstyle. Since dynamic hair modeling is underexplored, we integrate our dynamics model into each baseline, optimize their canonical hair appearance, and reanimate hair using our estimated motion for fair comparison. We further benchmark dynamics against two references: rigid-transformed canonical hair (lower bound) and physics-based XPBD [53] results (upper bound). Full baseline-training and implementation details are provided in the Appendices. 4.2 Ablation Study |
| Researcher Affiliation | Collaboration | 1University of Southern California 2Meta Reality Labs Research |
| Pseudocode | Yes | Algorithm 1 Coarse-to-Fine Dynamic Hair Modeling Input: Canonical hair P can hair, proxy mesh (head and shoulders), head poses {Ht 2, Ht 1, Ht} Output: Dynamic hair P t hair at time t |
| Open Source Code | No | Our project page: https://junyingw.github.io/paper/dgh We will release our synthetic dynamic-hair dataset to accelerate research on dynamic hair modeling. Justification: The main contribution of this work lies in the proposed method rather than the dataset or experiments, however, we plan to release our dataset in the future to support further research in dynamic hair modeling. |
| Open Datasets | No | Due to the lack of real captures of dynamic hair deformations with accurate strand tracking, we create a new synthetic dynamic hair dataset from scratch. Hairstyles with strands are modeled from industry experts, and animated with a physics-based simulation engine. Mutiple-view images are generated with a render farm. The dataset includes frame-by-frame hair geometry deformation for various hair styles, along with corresponding upper body geometry and head motions. The 3D models are rendered using fine-tuned hair shaders, resulting in photorealistic videos (see Appendices). We will release our synthetic dynamic-hair dataset to accelerate research on dynamic hair modeling. |
| Dataset Splits | Yes | Evaluation. We evaluate our method on deformation and appearance, comparing it with baselines across 5 hair subjects using our synthetic hair dataset. Each groom s training dataset consists of 90 motion sequences, while testing is performed on the remaining 10 sequences. Training dataset. We train each hairstyle independently to obtain a hairstyle-specific hair deformation model. For each hairstyle, we use 90 motion subjects, resulting in a geometry training dataset of 9K frames. Testing dataset. We evaluate our dynamic hair model across different hairstyles using 10 motion sequences, resulting in 1K frames. |
| Hardware Specification | Yes | We train our model on a single A100 GPU using the Adam optimizer and a learning rate of 1 10 4 for both stages. We train and test our model on a single A100 GPU, with training times of approximately 20 hours for the dynamic hair coarse stage, 20 hours for the fine stage, and 26 hours for the appearance model. For hair deformation runtime analysis, we report runtime with various strand numbers on RTX4090 GPU, compared with XPBD, as shown in Fig 16. |
| Software Dependencies | No | The paper mentions "Adam optimizer" and "Blender" but does not specify their version numbers or any other key software dependencies with specific versions. For example, it does not mention Python or specific deep learning frameworks (like PyTorch or TensorFlow) with their versions. |
| Experiment Setup | Yes | We train our model on a single A100 GPU using the Adam optimizer and a learning rate of 1 10 4 for both stages. In Stage I, each iteration samples 200k points from the hair point cloud. Here we provide the formal definitions for Eq. 1, including the point loss Lpoint = 1 N PN i=1 ˆpi p GT i 2 2 and the SDF penalty loss LSDF = 1 N PN i=1 max(0, SDF(ˆpi)), where ˆpi is the predicted 3D hair point, p GT i is the corresponding ground-truth hair point, and SDF(ˆpi) denotes the signed distance to the body surface. We penalize points inside the mesh via the Re LU (i.e., max(0, )) operation. For Eq. 1, we set λp=1.0 and λSDF=0.01; for Eq. 6, we set λrgb=1.0, λssim=0.1, and λlpips=0.1. |