Supplementary Materials

On this page, we present a demo video and detailed comparative results showcasing the performance of our method against state-of-the-art controllable video generation approaches. We conduct the comparative experiments on data from two distinct domains: human and non-human, highlighting our model's generalization across diverse subject categories.

Demo Video

demo_video

Human Data Comparison

This section provides comparison with state-of-the-art methods on human data. Each row corresponds to a single test case with same input. Our model achieves excellent continuity in motion, consistent appearance, and stable background, while current SOTA methods show distortion in key areas like the hands and face, or experience catastrophic frame failure.

Case 1

Pose

Ours

UniAnimate

Animate-X

MagicPose

Case 2

Pose

Ours

UniAnimate

Animate-X

MagicPose

Case 3

Pose

Ours

UniAnimate

Animate-X

MagicPose

Case 4

Pose

Ours

UniAnimate

Animate-X

MagicPose

Non-Human Data Comparison

This section provides comparison with state-of-the-art methods on non-human data. Each row corresponds to a single test case. Since ATI, SG-I2V, and Tora are trajectory-guided video generation methods, we manually constructed the input control information for Tora and SG-I2V, and utilized ATI's self-proposed control information extraction mechanism for its input. Furthermore, given that the default generation frame counts of SG-I2V and Tora are less than the default frame count 81, used by our PoseAnything and ATI methods, we align the frame count of all videos to 81 frames by duplicating the last frame to ensure a intuitive comparison. The results indicate that our PoseAnything model demonstrates a significant advantage in precise object pose control, whereas competing methods struggle to achieve frame-level pose alignment and tend to generate hallucinations during large-range motion synthesis.

Case 1