Portrait Animation from a Single Photo

Every system has four parts: a source image that supplies appearance, a driving signal (a video of another face, or audio), a motion representation, and a generator that warps the source and fills in gaps. The First Order Motion Model (Siarohin et al., NeurIPS 2019) set the template. It learns keypoints without any labels, together with local affine transformations around each one. From the source and driving keypoints it predicts a dense motion field, which is a per-pixel flow, plus an occlusion map marking regions the source image cannot supply. It warps the source's feature maps with the flow and has a decoder paint in the occluded areas. It also uses relative motion transfer: it applies how the driving keypoints move relative to the driving video's first frame, not their absolute positions, so the source keeps its own face shape. SadTalker (CVPR 2023) animates from audio. It predicts 3D Morphable Model coefficients from speech, using ExpNet for expression and PoseVAE for head pose, then renders them through a keypoint-based face generator. LivePortrait, released by Kuaishou in 2024, uses implicit keypoints trained on a large dataset. It adds a stitching module that pastes the animated face back into the full image without misaligned shoulders, and retargeting modules for eye and lip openness. It is fast enough for near real-time use on a GPU. Diffusion-based methods such as Alibaba's EMO and the open Hallo project give more expressive results at much higher compute cost. One photo has no information about the sides of the head, the inside of the mouth or the teeth, so these regions are invented. Large head turns stretch the face, backgrounds and hair may warp along with it, and identity can drift. A common misconception is that these tools build a full 3D model of the person. Even 3DMM-based methods use only an approximate face model.