NOTE: This is a native timm (EVA) remap of facebook/sapiens2-pretrain-5b. Checkpoint keys have been converted to timm naming; the weights have not been fine-tuned. The original Sapiens2 License applies. The upstream model card is reproduced below with timm usage instructions.
Sapiens2-5B
Sapiens2 is a family of high-resolution vision transformers pretrained on 1 billion human images — designed for human-centric tasks such as pose estimation, body-part segmentation, surface normals, and pointmaps.
This repository contains the 5B parameter pretrained backbone. It produces dense per-patch features suitable for fine-tuning downstream task heads.
- 📄 Paper: arXiv:2604.21681
- 🌐 Project Page: rawalkhirodkar.github.io/sapiens2
- 💻 Code: github.com/facebookresearch/sapiens2
Model Details
- Developed by: Meta
- Model type: Vision Transformer
- License: Sapiens2 License
- Task: pretrain
- Format: safetensors
- File:
model.safetensors
Quick Start
Use a timm version that includes Sapiens2 support.
import torch
import timm
from PIL import Image
device = "cuda" if torch.cuda.is_available() else "cpu"
model = timm.create_model(
"hf-hub:timm/vit_5b_patch16_sapiens2.fb", pretrained=True, use_naflex=False,
).eval().to(device)
data_config = timm.data.resolve_model_data_config(model)
transform = timm.data.create_transform(**data_config, is_training=False)
image = Image.open("image.jpg").convert("RGB")
x = transform(image).unsqueeze(0).to(device)
with torch.inference_mode():
tokens = model.forward_features(x)
cls_features = tokens[:, 0]
patch_features = tokens[:, model.num_prefix_tokens:] # exclude CLS and register tokens
model(x) uses CLS-token pooling by default, matching the original Sapiens2 convention.
Pass global_pool="avg" to create_model for average pooling over patch tokens.
Model Card
| Field | Value |
|---|---|
| Architecture | Sapiens2 ViT (RoPE, GQA, SwiGLU, RMSNorm, QK-norm) |
| Parameters | 5.071 B |
| FLOPs | 15.722 T |
| Embedding dim | 2432 |
| Layers | 56 |
| Attention heads | 32 |
| Pretraining resolution | 1024 × 768 (H × W) |
| Patch size | 16 |
| Pretraining data | 1B human images |
Sapiens2 Family
| Model | Params | FLOPs | Embed dim | Layers | Heads |
|---|---|---|---|---|---|
| Sapiens2-0.1B | 0.114 B | 0.342 T | 768 | 12 | 12 |
| Sapiens2-0.4B | 0.398 B | 1.260 T | 1024 | 24 | 16 |
| Sapiens2-0.8B | 0.818 B | 2.592 T | 1280 | 32 | 16 |
| Sapiens2-1B | 1.462 B | 4.715 T | 1536 | 40 | 24 |
| Sapiens2-1B-4K | 1.607 B | — | 1536 | 40 | 24 |
| Sapiens2-5B (this) | 5.071 B | 15.722 T | 2432 | 56 | 32 |
See the Sapiens2 Collection for all variants and downstream task checkpoints (pose, segmentation, normals, pointmaps).
Intended Use
- Feature extraction for human-centric downstream tasks
- Initialization for fine-tuning task heads (pose, segmentation, normals, pointmap)
- Research on human-centric vision
License
Released under the Sapiens2 License.
Citation
@article{khirodkarsapiens2,
title={Sapiens2},
author={Khirodkar, Rawal and Wen, He and Martinez, Julieta and Dong, Yuan and Su, Zhaoen and Saito, Shunsuke},
journal={arXiv preprint arXiv:2604.21681},
year={2026}
}