Qualitative comparison on dense control. 2D-based methods, MagicMotion and Go-with-the-Flow struggle to capture fine-grained details, while DiffusionAsShader also fails as its trajectory representation cannot handle newly emerging points. In contrast, our method outperforms them by closely following the motion in the source frames.
Qualitative comparison on spatial-sparse control. The subject in the left of input image is occluded by the subject in the right. 2D-based methods (MagicMotion, ToRA) fail in handling occlusion, U-Net-based method LeviTor introduces artifacts, while ours accurately captures occlusion with high visual fidelity.
Qualitative comparison on temporal-sparse control. SparseCtrl yields unsatisfactory results, while MagicMotion shows weak alignment and blurriness. Our method aligns with the anchor-frame motion and generates coherent in-between frames.
Qualitative comparison on unaligned control. DAS introduces artifacts (red blurriness around subject) from strict alignment, while Go-with-the-Flow produces implausible results. Our method flexibly follows input motion.