SynthHair: Leveraging MetaHumans for a High-Quality 4K Hair Matting Dataset

Recent progress in computer vision enables dense alpha matte predictions at resolutions of 4K and higher. While portrait matting has benefited from large-scale, high-resolution datasets, progress in hair matting remains limited due to the difficulty of annotating sub-pixel fine hair details. In this work, we analyze critical limitations in image resolution and annotation quality of such existing datasets and introduce our own synthetic hair matting dataset, SynthHair, which consists of 16k RGBA portrait images in $4096\times 4096$ image resolution, containing labels for hair matting, as well as hair categorization, growth direction, and depth estimation. We create this dataset using MetaHuman Creator and Unreal Engine to obtain highly detailed and realistic renders of human faces with complex hair structures. We adapt a generalized version of the isoperimetric inequality quotient (IPQ) that we call the soft isoperimetric inequality quotient (SIPQ), as a scale-invariant label complexity metric that enables comparison of both segmentation and matte labels. Using this metric, we demonstrate that hair matting complexity increases significantly with resolution, reaching an SIPQ of 325.41 at 4K, in contrast with portrait matting datasets at similar resolutions, which reach an SIPQ of  10. To validate our data, we train separate matting models on SynthHair and existing real-world datasets. Our results demonstrate strong domain generalization and show clear improvements in high-resolution images. A/B testing further confirms these findings, with over 80% of users preferring our model's results on high-resolution images.