transformers/docs/source/en/model_doc/sapiens2.md

This model was published in HF papers on 2026-04-23 and contributed to Hugging Face Transformers on 2026-06-03.

Sapiens2

Overview

The Sapiens2 model was proposed in Sapiens2 by Rawal Khirodkar, He Wen, Julieta Martinez, Yuan Dong, Zhaoen Su, Shunsuke Saito. Sapiens2 is a family of high-resolution vision transformers pretrained on ~1 billion curated human images, designed for human-centric computer vision tasks including pose estimation, body-part segmentation, surface normal estimation, and pointmap estimation.

You can find all the original Sapiens2 checkpoints under the Sapiens2 collection.

The abstract from the paper is the following:

We present Sapiens2, a family of high-resolution transformers for human-centric vision focused on generalization, versatility, and high-fidelity outputs. We pretrain on ~1 billion curated high-quality human images with improved task annotations and combine masked image reconstruction with self-distilled contrastive objectives to learn both low-level and semantic features. Our models scale from 0.4B to 5B parameters and train at native 1K resolution, with hierarchical 4K variants for extended spatial reasoning. Sapiens2 achieves substantial improvements over its predecessor: +4 mAP in pose estimation, +24.3 mIoU in body-part segmentation, and 45.6% error reduction in normal estimation, while extending to new tasks like pointmap and albedo estimation. Code is publicly available.

Tips:

• Sapiens2 uses Rotary Position Embeddings (RoPE) and supports arbitrary input resolutions. The default image processor resizes images to 1024×768 (height×width).
• The model uses Grouped Query Attention (GQA) for middle layers and full multi-head attention for the first and last 8 layers.
• Register tokens (8 by default) reduce high-norm artifacts in patch tokens, yielding cleaner attention maps and better performance on dense prediction tasks.

This model was contributed by guarin. The original code can be found here.

Usage examples

The example below shows how to obtain the CLS token (whole-image embedding) with [Sapiens2Model].

import torch
from transformers import AutoImageProcessor, AutoModel
from transformers.image_utils import load_image

image = load_image("http://images.cocodataset.org/val2017/000000004016.jpg")

image_processor = AutoImageProcessor.from_pretrained("facebook/sapiens2-pretrain-0.4b")
model = AutoModel.from_pretrained("facebook/sapiens2-pretrain-0.4b", device_map="auto")

inputs = image_processor(images=image, return_tensors="pt").to(model.device)
with torch.inference_mode():
    outputs = model(**inputs)

# outputs.pooler_output is the CLS token (whole-image embedding)
cls_token = outputs.pooler_output
print("CLS token shape:", cls_token.shape)  # [1, 1024]

[Sapiens2Backbone] exposes patch tokens already reshaped to spatial dimensions and CLS tokens directly on the output object.

import torch
from transformers import AutoBackbone, AutoImageProcessor
from transformers.image_utils import load_image

image = load_image("http://images.cocodataset.org/val2017/000000004016.jpg")

image_processor = AutoImageProcessor.from_pretrained("facebook/sapiens2-pretrain-0.4b")
model = AutoBackbone.from_pretrained("facebook/sapiens2-pretrain-0.4b", device_map="auto")

inputs = image_processor(images=image, return_tensors="pt").to(model.device)
with torch.inference_mode():
    outputs = model(**inputs, return_class_token=True)

# Patch tokens shaped (batch, height, width, channels)
patch_features = outputs.feature_maps[0]
cls_token = outputs.cls_tokens[0]
print("CLS token shape:", cls_token.shape)           # [1, 1024]
print("Patch features shape:", patch_features.shape) # [1, 64, 48, 1024]

The example below shows how to estimate surface normals with [Sapiens2ForNormalEstimation]. The output normals are raw (unnormalized); use post_process_normal_estimation to resize and L2-normalize them.

import torch
from transformers import AutoImageProcessor, AutoModelForNormalEstimation
from transformers.image_utils import load_image

image = load_image("http://images.cocodataset.org/val2017/000000004016.jpg")

image_processor = AutoImageProcessor.from_pretrained("facebook/sapiens2-normal-0.4b")
model = AutoModelForNormalEstimation.from_pretrained("facebook/sapiens2-normal-0.4b", device_map="auto")

inputs = image_processor(image, return_tensors="pt").to(model.device)
with torch.inference_mode():
    outputs = model(**inputs)

# outputs.normals shape: (batch_size, 3, height, width) — raw, unnormalized XYZ normals
print("Normals shape:", outputs.normals.shape)  # [1, 3, 1024, 768]

# Remove preprocessing padding, resize to original size, and L2-normalize to unit vectors in [-1, 1]
original_size = (image.height, image.width)
result = image_processor.post_process_normal_estimation(
    outputs, source_sizes=[original_size], target_sizes=[original_size]
)
normals = result[0]["normals"]
print("Normals shape:", normals.shape)   # [3, original_height, original_width]


### Example code for visualization

# Convert L2-normalized normals in [-1, 1] to RGB in [0, 255]
normals_rgb = ((normals + 1.0) / 2.0 * 255.0).clamp(0, 255).to(torch.uint8)

# Apply background removal using the segmentation model output.
# `segmentation` is the output of `post_process_semantic_segmentation` — a (H, W) tensor
# of per-pixel class IDs, where class 0 is background.
background_mask = segmentation == 0
normals_rgb[:, background_mask] = 0
print("Normals RGB shape:", normals_rgb.shape)   # [3, original_height, original_width]

The example below shows how to estimate per-pixel 3D coordinates with [Sapiens2ForPointmapEstimation]. Use post_process_pointmap_estimation to remove preprocessing padding, resize to the original image size, and apply the predicted focal-length scale.

import torch
from transformers import AutoImageProcessor, AutoModelForPointmapEstimation
from transformers.image_utils import load_image

image = load_image("http://images.cocodataset.org/val2017/000000004016.jpg")

image_processor = AutoImageProcessor.from_pretrained("facebook/sapiens2-pointmap-0.4b")
model = AutoModelForPointmapEstimation.from_pretrained("facebook/sapiens2-pointmap-0.4b", device_map="auto")

inputs = image_processor(image, return_tensors="pt").to(model.device)
with torch.inference_mode():
    outputs = model(**inputs)

# outputs.pointmaps shape: (batch_size, 3, height, width) — raw XYZ in canonical camera space
print("Pointmaps shape:", outputs.pointmaps.shape)  # [1, 3, 1024, 768]

# Remove preprocessing padding, resize to original size, and apply focal-length scale
original_size = (image.height, image.width)
result = image_processor.post_process_pointmap_estimation(
    outputs, source_sizes=[original_size], target_sizes=[original_size]
)
pointmap = result[0]["pointmap"]
print("Pointmap shape:", pointmap.shape)  # [3, original_height, original_width]


### Example code for visualization

# Visualize the pointmap as an RGB image using inverse-depth and the turbo colormap.
import matplotlib.pyplot as plt

# `segmentation` is the output of `post_process_semantic_segmentation` — a (H, W) tensor
# of per-pixel class IDs, where class 0 is background.
foreground_mask = segmentation != 0
depth = pointmap[2]  # Z channel: depth in camera space, shape (H, W)
pointmap_rgb = torch.zeros(3, *depth.shape, dtype=torch.uint8)
foreground_depth = depth[foreground_mask]
if foreground_depth.numel() > 0:
    depth_low, depth_high = torch.quantile(foreground_depth, torch.tensor([0.01, 0.99]))
    inverse_depth = 1.0 / foreground_depth.clamp(min=1e-6)
    inverse_depth_low = 1.0 / depth_high.clamp(min=1e-6)
    inverse_depth_high = 1.0 / depth_low.clamp(min=1e-6)
    inverse_depth_normalized = ((inverse_depth - inverse_depth_low) / (inverse_depth_high - inverse_depth_low + 1e-8)).clamp(0, 1)
    turbo = plt.get_cmap("turbo")
    foreground_colors = torch.from_numpy(turbo(inverse_depth_normalized.cpu().numpy())[..., :3] * 255).to(torch.uint8)  # (N, 3)
    pointmap_rgb[:, foreground_mask] = foreground_colors.T
print("Pointmap RGB shape:", pointmap_rgb.shape)  # [3, original_height, original_width]

The example below shows how to run pose estimation with [Sapiens2ForPoseEstimation]. The model predicts per-keypoint heatmaps; use post_process_pose_estimation to decode them back to image-space keypoint coordinates. It requires opencv-python (pip install opencv-python).

import torch
from transformers import AutoImageProcessor, AutoModelForPoseEstimation
from transformers.image_utils import load_image

image = load_image("http://images.cocodataset.org/val2017/000000004016.jpg")

image_processor = AutoImageProcessor.from_pretrained("facebook/sapiens2-pose-0.4b")
model = AutoModelForPoseEstimation.from_pretrained("facebook/sapiens2-pose-0.4b", device_map="auto")

# Provide bounding boxes in COCO format (x, y, width, height) for each person
boxes = [[[270.8, 0.6, 294.1, 379.5]]]
inputs = image_processor(image, boxes=boxes, return_tensors="pt").to(model.device)
with torch.inference_mode():
    outputs = model(**inputs)

# outputs.heatmaps shape: (num_persons, num_keypoints, heatmap_height, heatmap_width)
print("Heatmaps shape:", outputs.heatmaps.shape)  # [1, 308, 256, 192]

# Decode heatmaps to image-space keypoint coordinates
results = image_processor.post_process_pose_estimation(outputs, boxes=boxes)[0]
keypoints = results[0]["keypoints"]   # (num_keypoints, 2) — x/y in image coordinates
scores = results[0]["scores"]         # (num_keypoints,) — per-keypoint confidence
print("Keypoints shape:", keypoints.shape)

Horizontal flip augmentation (test-time augmentation) improves keypoint accuracy by averaging predictions from the original and mirrored image. Pass flip_pairs — a tensor of [left_keypoint, right_keypoint] pairs — to the second forward pass. The model flips the heatmaps back to the original orientation before returning them, so you can average both outputs directly.

import torch
from transformers import AutoImageProcessor, AutoModelForPoseEstimation
from transformers.image_utils import load_image

image = load_image("http://images.cocodataset.org/val2017/000000004016.jpg")

image_processor = AutoImageProcessor.from_pretrained("facebook/sapiens2-pose-0.4b")
model = AutoModelForPoseEstimation.from_pretrained("facebook/sapiens2-pose-0.4b", device_map="auto")

boxes = [[[270.8, 0.6, 294.1, 379.5]]]
inputs = image_processor(image, boxes=boxes, return_tensors="pt").to(model.device)
pixel_values = inputs["pixel_values"]

flip_pairs = torch.tensor(model.config.flip_pairs, device=model.device)

with torch.inference_mode():
    outputs = model(pixel_values)
    outputs_flipped = model(pixel_values.flip(-1), flip_pairs=flip_pairs)

results = image_processor.post_process_pose_estimation(outputs, outputs_flipped=outputs_flipped, boxes=boxes)[0]
keypoints = results[0]["keypoints"]
scores = results[0]["scores"]

The example below shows how to perform body-part segmentation with [Sapiens2ForSemanticSegmentation].

import torch
from transformers import AutoImageProcessor, AutoModelForSemanticSegmentation
from transformers.image_utils import load_image

image = load_image("http://images.cocodataset.org/val2017/000000004016.jpg")

image_processor = AutoImageProcessor.from_pretrained("facebook/sapiens2-seg-0.4b")
model = AutoModelForSemanticSegmentation.from_pretrained("facebook/sapiens2-seg-0.4b", device_map="auto")

inputs = image_processor(image, return_tensors="pt").to(model.device)
with torch.inference_mode():
    outputs = model(**inputs)

# outputs.logits shape: (batch_size, num_labels, height, width)
print("Logits shape:", outputs.logits.shape)  # [1, 29, 1024, 768]

# Get per-pixel class predictions, optionally resized to the original image size
original_size = (image.height, image.width)
segmentation = image_processor.post_process_semantic_segmentation(
    outputs, target_sizes=[original_size]
)[0]
print("Segmentation map shape:", segmentation.shape)  # [original_height, original_width]

The example below shows how to run image matting with [Sapiens2ForImageMatting]. Outputs are sigmoid-activated and already in [0, 1]; use post_process_image_matting to resize and split into alphas, foregrounds, and an optional composite image. The composite image shows the foreground overlaid over the background with the formula: composite = foreground * (1 - alpha) * background.

import torch
from transformers import AutoImageProcessor, AutoModelForImageMatting
from transformers.image_utils import load_image

image = load_image("http://images.cocodataset.org/val2017/000000004016.jpg")

image_processor = AutoImageProcessor.from_pretrained("facebook/sapiens2-matting-1b")
model = AutoModelForImageMatting.from_pretrained("facebook/sapiens2-matting-1b", device_map="auto")

inputs = image_processor(image, return_tensors="pt").to(model.device)
with torch.inference_mode():
    outputs = model(**inputs)

# outputs.foregrounds: (1, 3, H, W), outputs.alphas: (1, 1, H, W) — both in [0, 1]
original_size = (image.height, image.width)

# Pass an optional background to composite the foreground over it.
# A (3, 1, 1) tensor broadcasts as a uniform color; PIL images and numpy arrays are also accepted.
background = torch.tensor([0, 177, 64], dtype=torch.uint8).view(3, 1, 1)  # chroma green in RGB
result = image_processor.post_process_image_matting(
    outputs, target_sizes=[original_size], backgrounds=background
)[0]
print("Alpha shape:", result["alpha"].shape)        # [1, original_height, original_width]
print("Foreground shape:", result["foreground"].shape)  # [3, original_height, original_width]
print("Composite shape:", result["composite"].shape)    # [3, original_height, original_width] — uint8 [0, 255]

Sapiens2Config

autodoc Sapiens2Config

Sapiens2HeadConfig

autodoc Sapiens2HeadConfig

Sapiens2ImageProcessor

autodoc Sapiens2ImageProcessor - preprocess - post_process_image_matting - post_process_normal_estimation - post_process_pointmap_estimation - post_process_pose_estimation - post_process_semantic_segmentation

Sapiens2Model

autodoc Sapiens2Model - forward

Sapiens2Backbone

autodoc Sapiens2Backbone - forward

Sapiens2ForImageMatting

autodoc Sapiens2ForImageMatting - forward

Sapiens2ForNormalEstimation

autodoc Sapiens2ForNormalEstimation - forward

Sapiens2ForPointmapEstimation

autodoc Sapiens2ForPointmapEstimation - forward

Sapiens2ForPoseEstimation

autodoc Sapiens2ForPoseEstimation - forward

Sapiens2ForSemanticSegmentation

autodoc Sapiens2ForSemanticSegmentation - forward