Large Displacement Motion Transfer with Unsupervised Anytime Interpolation

ICML 2025 Aims to transfer pose from a driving video to an object in a source image, animating the source object. Hot research topic with applications in film animation, game production, and face exchange. Goal: Animate a still image by transferring pose from a driving video to generate a video with the same pose. Key challenges: Accurately transfer motion patterns. Maintain identity consistency. Struggle with large displacement motions. What is Motion Transfer? Problem: Current unsupervised methods struggle with large displacement motions. Existing Methods (Limitations): Supervised methods: Rely on prior knowledge (landmarks, 3D models), often fail with out-of-training data. Unsupervised methods: FOMM, MRAA: Use local linear affine transformations, struggle with non-linear object motion. TPSMM: Uses Thin-plate splines, but keypoint detection is often inaccurate. CPABMM: Uses continuous piecewise affine transformation, limited in finer motions, causing artifacts. CoP: Based on chain-of-pose, difficult to obtain pose chains with different identity information or large displacement. Problem Statement & Existing Methods Core Idea: Decompose large displacement motion into many small displacement motions by inserting intermediate images. Keypoint-based anytime interpolation: Estimates keypoint information of interpolated images based on source and driving image keypoints.