Single-view 3D Face Reconstruction

Given a single image of a person, the task is to produce a 3D mesh representing the person's head. There are 2 tracks:

The challenge is conducted on 391 images from 20 different persons of various ethnicities, ages, and genders. The images have a resolution of 512x512 pixels and are already cropped to the face region. The meshes for each image can be provided with FLAME topology or with arbitrary topology, in which case additional 7 landmarks are required for alignment.

Please see the NeRSemble Benchmark Toolkit on how to obtain the data to participate in the benchmark and prepare a submission.

Evaluation and Metrics

We evaluate the geometric accuracy of submitted meshes by comparing them to the corresponding ground-truth point clouds that were obtained via Multi-view stereo. To eliminate the effect of arbitrary world spaces on the evaluation, we first rigidly align the submitted mesh to the ground-truth pointcloud with the Kabsch-Umeyama algorithm using 7 landmarks (NoW benchmark). This alignment serves as initialization for the subsequent Iterative Closest Points (ICP) registration which ensures that the submitted mesh and the ground-truth pointcloud match as good as possible. After registration, geometric evaluation can be performed in the pointcloud's metrical space. We measure both L1 and L2 uni-directional point-to-mesh Chamfer distance in millimeters (Chamfer L1 [mm] and Chamfer L2 [mm]) by finding the nearest neighbor on the mesh for each point in the ground-truth pointcloud. We also measure cosine similarity of the matching points' normal vectors (Normals Similarity).

Results

Methods Chamfer L1 [mm] Chamfer L2 [mm] Normals Similarity
RealDenseFace 1.610 1.086 0.8857
Linzhou Li; Tianjia Shao; Kun Zhou. RealDenseFace: Real-time Monocular 3D Face Reconstruction from Dense UV-space Priors.
Pixel3DMM 1.659 1.118 0.8847
Simon Giebenhain, Tobias Kirschstein, Martin Rünz, Lourdes Agapito, Matthias Nießner. Pixel3DMM: Versatile Screen-Space Priors for Single-Image 3D Face Reconstruction. ICLR 2026
PartFusionMI 1.700 1.145 0.8860
PartFusion 1.868 1.256 0.8828
FlowFace 1.973 1.330 0.8802
Felix Taubner, Prashant Raina, Mathieu Tuli, Eu Wern Teh, Chul Lee, Jinmiao Huang. 3D Face Tracking from 2D Video through Iterative Dense UV to Image Flow. CVPR 2024
SHeaP 2.083 1.407 0.8764
Liam Schoneveld, Zhe Chen, Davide Davoli, Jiapeng Tang, Saimon Terazawa, Ko Nishino, Matthias Nießner. SHeaP: Self-Supervised Head Geometry Predictor Learned via 2D Gaussians. ICCV 2025
Skullptor 2.260 1.531 inf
Noe Artru, Rukhshanda Hussain, Emeline Got, Alexandre Messier, David B. Lindell, Abdallah Dib. Skullptor: High Fidelity 3D Head Reconstruction in Seconds with Multi-View Normal Prediction. CVPR 2026
smirk 2.276 1.533 0.8697
George Retsinas, Panagiotis P. Filntisis, Radek Danecek, Victoria F. Abrevaya, Anastasios Roussos, Timo Bolkart, Petros Maragos. SMIRK: 3D Facial Expressions through Analysis-by-Neural-Synthesis. CVPR 2024
DECA 2.385 1.611 0.8710
Yao Feng, Haiwen Feng, Michael J. Black, and Timo Bolkart . DECA: Learning an Animatable Detailed 3D Face Model from In-The-Wild Images. SIGGRAPH 2021
TokenFace 2.627 1.779 0.8655
Tianke Zhang, Xuangeng Chu, Yunfei Liu, Lijian Lin, Zhendong Yang, Zhengzhuo Xu. Accurate 3D Face Reconstruction with Facial Component Tokens. ICCV 2023
EMOCA 2.636 1.777 0.8602
Radek Daněček, Michael J. Black, Timo Bolkart . EMOCA: Emotion Driven Monocular Face Capture and Animation. CVPR 2022

Please refer to the submission instructions before making a submission

Submit results