Table 4: Effect of joint cross attention on depth condition learning. Joint Cross Attention is our proposed method, which trains without tracking condition.

Depth Condition	Background Consistency ↑	Subject Consistency ↑	Dynamic ↑	Imaging Quality ↑	CLIP-I ↑	DINO-I ↑
✗	0.934	0.942	0.480	0.675	0.698	0.563
Controlnet	0.918	0.880	0.333	0.558	0.720	0.393
Joint Cross Attention (ours)	0.935	0.945	0.480	0.682	0.720	0.575