sapiens-pretrain-2b-bfloat16

Pretrain-Sapiens-2B-Bfloat16

Model Details

Sapiens is a family of vision transformers pretrained on 300 million human images at 1024 x 1024 image resolution. The pretrained models, when finetuned for human-centric vision tasks, generalize to in-the-wild conditions. Sapiens-2B natively support 1K high-resolution inference. The resulting models exhibit remarkable generalization to in-the-wild data, even when labeled data is scarce or entirely synthetic.

• Developed by: Meta
• Model type: Vision Transformer
• License: Creative Commons Attribution-NonCommercial 4.0
• Task: pretrain
• Format: bfloat16
• File: sapiens_2b_epoch_660_bfloat16.pt2

Model Card

• Image Size: 1024 x 1024
• Num Parameters: 2.163 B
• FLOPs: 8.709 TFLOPs
• Patch Size: 16 x 16
• Embedding Dimensions: 1920
• Num Layers: 48
• Num Heads: 32
• Feedforward Channels: 7680

More Resources

• Repository: https://github.com/facebookresearch/sapiens
• Paper: https://arxiv.org/abs/2408.12569
• Demo: Sapiens Gradio Spaces
• Project Page: https://about.meta.com/realitylabs/codecavatars/sapiens
• Additional Results: https://rawalkhirodkar.github.io/sapiens
• HuggingFace Collection: https://huggingface.co/collections/facebook/sapiens-66d22047daa6402d565cb2fc

Uses

Pretrained 2B model can be used for feature extraction, fine-tuning, or as a starting point for training new models.