How to Swap Faces with Python: Step‑by‑Step Guide Using dlib and OpenCV

Last year a deepfake video swapping Zhu Yin with Yang Mi went viral, demonstrating the power of AI‑driven face swapping.

In this article we show how to achieve a similar effect on static images using Python, dlib and OpenCV.

Overview

The process consists of four main steps: detecting facial landmarks, aligning the second image to the first, correcting color differences, and blending the faces.

1. Detect facial landmarks

We use dlib’s Python bindings, which implement the algorithm from “One Millisecond Face Alignment with an Ensemble of Regression Trees”. The get_landmarks() function returns a 68×2 array of (x, y) coordinates for each landmark.

2. Align the faces (Ordinary Procrustes Analysis)

With the two landmark matrices we solve for the optimal rotation R , scale s and translation T that minimize the distance between corresponding points.

3. Color correction

To reduce visible seams we adjust the color of the second image by dividing it by its Gaussian‑blurred version and multiplying by the blurred first image, applying a local RGB scaling factor.

4. Blend the faces

A mask is created for the second image, feathered at the edges, and combined with the first image using an affine warp. The final result is obtained by applying the mask to merge the two images.

Following these steps produces a convincing face swap on static pictures.