We present Sapiens, a family of models for four fundamental human-centric vision tasks: 2D pose estimation, body-part segmentation, depth estimation, and surface normal prediction. Our models natively support 1K high-resolution inference and are extremely easy to adapt for individual tasks by simply fine-tuning models pretrained on over 300 million in-the-wild human images. We show that self-supervised pretraining on a curated dataset of human images significantly boosts performance for a diverse set of human-centric tasks, given the same computational budget. Model performance across tasks consistently improves as the number of parameters is scaled from 0.3 to 2 billion. Sapiens achieves significant improvements over prior state-of-the-art on Humans-5K (pose) by 7.6 mAP, Humans-2K (part-seg) by 17.1 mIoU, Hi4D (depth) by 22.4% relative RMSE, and THuman2 (normal) by 53.5% relative angular error.

Human-centric vision encompasses a wide range of tasks that focus on understanding human appearance, pose, and geometry from images. These tasks are critical for applications in augmented reality, fitness tracking, virtual try-on, human-computer interaction, and autonomous systems. Despite their importance, existing approaches typically train separate, task-specific models, each requiring specialized architectures and training procedures. This fragmented approach limits knowledge sharing across related tasks and makes it difficult to leverage the full potential of large-scale pretraining. In contrast, foundation models in NLP and general vision have demonstrated that a single, well-pretrained model can serve as a powerful starting point for diverse downstream tasks.

The Sapiens framework consists of two stages: large-scale self-supervised pretraining followed by task-specific fine-tuning. For pretraining, we curate a dataset of over 300 million human images from public sources, filtered to ensure high quality and diversity in terms of body poses, camera viewpoints, clothing, and environmental conditions. We use Masked Autoencoder (MAE) as our pretraining objective, which learns representations by randomly masking large portions (75%) of the input image and reconstructing the missing patches. The key distinction from general-purpose MAE pretraining is our focus on human-only images, which allows the model to develop specialized representations for human body structure.

We adopt Vision Transformer (ViT) as our backbone architecture, scaling from ViT-Small (0.3B parameters) to ViT-Huge (2B parameters). A critical design choice is our support for native 1024x1024 resolution during both pretraining and fine-tuning. This is essential for human-centric tasks because human body details such as finger positions, facial features, and clothing textures require high-resolution input to be accurately perceived. For each downstream task, we attach a simple task-specific head to the pretrained backbone and fine-tune end-to-end. The simplicity of adaptation — requiring only the addition of a lightweight head — underscores the quality of the learned representations.

We evaluate Sapiens on four tasks across multiple benchmarks. For 2D pose estimation on Humans-5K, Sapiens-2B achieves 78.2 mAP, improving over the previous state-of-the-art by 7.6 mAP. For body-part segmentation on Humans-2K, we achieve 68.3 mIoU, a 17.1 mIoU improvement. For depth estimation on Hi4D, we reduce the relative RMSE by 22.4%, and for surface normal prediction on THuman2, we reduce the relative angular error by 53.5%. Critically, we demonstrate consistent scaling behavior across all tasks: performance improves monotonically as we scale from 0.3B to 2B parameters, with no signs of saturation, suggesting that further scaling would yield additional improvements.

We have introduced Sapiens, a family of foundation models for human-centric vision that demonstrates the power of large-scale, domain-specific self-supervised pretraining. By curating 300M+ human images and pretraining ViT models at 1K resolution with up to 2B parameters, we achieve significant advances across four fundamental human vision tasks. Our results establish that the foundation model paradigm — pretrain once, fine-tune for many tasks — is highly effective for the human-centric vision domain. We release our pretrained models and code to accelerate research in this important area.