Surface Normals

Per-pixel surface-normal estimation. Output channels are the (x, y, z) components of the normal vector in camera space.

Real sapiens_normal output (0.4b) — surface normals mapped to RGB, blended with input

Signature

sapiens_normal(
    input_path:  str,
    output_dir:  str,
    model_size:  str = "0.4b",
    device:      str = "cuda:0",
    save_pred:   bool = True,
) -> dict

Example

from strands_sapiens import sapiens_normal

sapiens_normal(
    input_path="person.jpg",
    output_dir="out/",
    model_size="0.4b",
)

Output:

• out/person.jpg - side-by-side input vs. remapped RGB normals
• out/person_normal.npy - 3×H×W float array, unit-normalized

Input

Surface normals (RGB-mapped)

Blended overlay

Visualization

The viz maps each normal (x, y, z) ∈ [-1, 1]³ to RGB via

rgb = (normal * 0.5 + 0.5) * 255

Red channel = x (horizontal), Green = y (vertical), Blue = z (depth / toward camera).

Flat walls facing the camera show up as ~(128, 128, 255).

Use-cases

• Relighting / portrait studio: combine with albedo → relight photos after the fact.
• Clothing wrinkle analysis: normals are sensitive to fine fabric detail.
• AR filter preprocessing: drive shading in real time without a depth sensor.

Consume the raw .npy

import numpy as np
n = np.load("out/person_normal.npy")  # (3, H, W)
# unit-normalize (already close but re-do for safety)
norm = np.linalg.norm(n, axis=0, keepdims=True)
n = n / np.clip(norm, 1e-6, None)

Related

• Albedo → · Pointmap →

Normals data flow